Você está na página 1de 220

A Market Approach to Forecasting: Background, Theory and Practice

Georgios G. Tziralis

Submitted in partial fulllment of the requirements for the degree of Doctor of Philosophy in the Sector of Industrial Management and Operational Research

SCHOOL OF MECHANICAL ENGINEERING NATIONAL TECHNICAL UNIVERSITY OF ATHENS


June 2013

c June 2013 Georgios G. Tziralis All Rights Reserved

ABSTRACT
A Market Approach to Forecasting: Background, Theory and Practice Georgios G. Tziralis
Advisor: Prof. Ilias Tatsiopoulos

My PhD Thesis focuses on the omnipresent problem of business forecasting. In its very core, the problem refers to estimating a variables future price, potentially being correlated to a set of variables, or ideally following a specic pattern. There exists a variety of statistical approaches, ranging from classical econometric techniques to recent data mining algorithms, attempting to extract such underlying relationships from past data and reproduce them to predict future values. However, in practice, these correlations are highly dynamic and patterns, if any, usually evolve in chaotic ways. Therefore even highly adjustable machine learning techniques tend to perform poorly, suffering from inherent systemic impotencies. Human experts, on the other hand, seem to be better suited to address this request. A single expert can process a big number of variables and quickly adjust predictions into varying circumstances. However, bias and shortcomings reduce the value of such estimates. Consequently, there lies a need and opportunity to develop mechanisms that can demonstrate a capacity to aggregate information from experts, in a dynamic, cost and time-effective manner that negates their inherent shortcomings. Financial markets arise as an emerging paradigm to handle such a request. Wellestablished market institutions typically focus on serving either investment, hedging or speculation purposes, while their fundamental function of information aggregation remains latent in all but newly, and yet of limited exposure, prediction markets institutions. I provide a market topology scheme and locate existing market institutions in terms of usage of these ubiquitous functions. And I am studying prediction markets in this Thesis.

I start with a survey. I provide an extensive literature review of the eld and a detailed classication scheme of the existing literature, both appearing for the rst time in the bibliography. I then move on to identify inappropriate attributes ensued from market mechanisms typically used to serve a prediction market institution. I build on the latest advances of market scoring rules and dynamic pari-mutuel markets to propose an analytical framework for a coherent price function of a dynamic pari-mutuel market. I construct and validate a function satisfying this framework, nally contributing what may be perceived as a coherent hybrid between dynamic pari-mutuel market and market scoring rules. I also provide an attempt to evaluate the hybrids performance, by shaping a dynamic market model and simulating various agent behaviors and strategies to study the properties of convergence, acquisition of recently arrived information and equilibrium, among others. I then proceed from models and simulations to practice. To examine the behavior of the proposed market mechanism and its forecasting capacity, I build and present a fully-scaled web platform, AskMarkets. Taking advantage of it, I run and analyze a number of experiments in a variety of academic, professional and social contexts to empirically validate the applicability of theoretical ndings. I also provide a practical deployment framework for prediction markets implementation, leveraging on the experience accumulated due to the extent of this study. I also address some advanced concepts and the open question of prediction markets efciency. I approach it by studying the convolution of market prices and relevant news streams. I empirically validate the approach using the most varied data set of real-money prediction markets contracts ever examined. I nally propose a sophisticated analytical technique for prediction markets to serve for event detection and eliciting market inefciencies. Overall, I believe that this Thesis provides a unique contribution to the eld of prediction markets, forecasting, decision sciences and operations research at large, studying the mechanism at full scale, ranging from an extensive literature review to a coherent market mechanism and from market modeling to simulations and extended case studies. The Thesis was completed in 2008, yet its defence took place in June 2013.

I would like to thank Professor Georgios Kosmetatos, who taught me to love research; my supervisor Professor Ilias Tatsiopoulos, who gave me the chance to work on what I love; Athanasios Tolis, who made me stay in the place I love; fellow coworkers Athanasios Rentizelas and Konstantinos Kirytopoulos, who instructed me to pursue excellence and results; also, Panos Ipeirotis for his support in Chapter 6 of this Thesis. Beyond my life in an NTUA lab, I would also like to extend this thank you note to a number of people who happily bore my peculiar behaviour, throughout the years I spent writing this Thesis.

Stouc gonec mou, gia th strix touc ensw afierwnmoun sta akadhmak.

Contents
Table of Contents List of Figures List of Tables 1 Introduction 1.1 1.2 1.3 2 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v viii x 1 1 2 4 7 7 7 10 12 12 13 13 14 18 18 19 19

Background 2.1 2.2 2.3 2.4 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Markets, in general . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fundamentals of Prediction Markets . . . . . . . . . . . . . . . . . . . . . . . Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 2.4.2 2.4.3 2.4.4 2.5 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Research Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Early works and evolution . . . . . . . . . . . . . . . . . . . . . . . .

Classication Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.0.1 2.5.0.2 2.5.0.3 Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . Theoretical work . . . . . . . . . . . . . . . . . . . . . . . . . Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . i

2.5.0.4 2.5.1 2.6 2.7

Law and Policy . . . . . . . . . . . . . . . . . . . . . . . . .

20 20 23 26 26 27 29 29 30 30 31 32 33 35 35 36 39 43 44 47 50 50 55 61 61 67 75 75 76

Classication Results . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Conclusions and Research Implications . . . . . . . . . . . . . . . . . . . . . Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7.1 2.7.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Theoretical Properties of Prediction Markets 3.1 3.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Market Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 3.2.2 3.2.3 3.2.4 3.3 Description & Characteristics . . . . . . . . . . . . . . . . . . . . . . . Continuous Double Auction . . . . . . . . . . . . . . . . . . . . . . . Market-maker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Pari-mutuel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Prediction Markets Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 3.3.2 3.3.3 Disadvantages for use in prediction markets . . . . . . . . . . . . . . Market Scoring Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . Dynamic Pari-Mutuel Market . . . . . . . . . . . . . . . . . . . . . . .

3.4

Properties for a Coherent Price Function . . . . . . . . . . . . . . . . . . . . 3.4.1 3.4.2 Simple case: n = 2 outcomes . . . . . . . . . . . . . . . . . . . . . . . Extension for n = k outcomes . . . . . . . . . . . . . . . . . . . . . . .

3.5

A Coherent Price Function I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.1 3.5.2 Simple case: n = 2 outcomes . . . . . . . . . . . . . . . . . . . . . . . Extension for n = k outcomes . . . . . . . . . . . . . . . . . . . . . . .

3.6

A Coherent Price Function II . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.1 3.6.2 Simple case: n = 2 outcomes . . . . . . . . . . . . . . . . . . . . . . . Extension for k outcomes . . . . . . . . . . . . . . . . . . . . . . . . .

3.7

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.1 3.7.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

Theoretical Evaluation of Prediction Markets 4.1 4.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Background & Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 4.2.2 4.3 Theoretical work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Empirical studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77 77 79 79 82 84 84 85 86 87 88 88 90 91 93 97 97

Design of Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 4.3.2 4.3.3 4.3.4 Model of markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Information structure . . . . . . . . . . . . . . . . . . . . . . . . . . . Market mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Agent strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4.4

Convergence Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 4.4.2 4.4.3 4.4.4 Price Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Convergence Speed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Best Possible Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . Convergence to the Best Prediction or Not . . . . . . . . . . . . . . .

4.5

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 4.5.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 103

Empirical Analysis of Prediction Markets 5.1 5.2

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Review of Empirical Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 5.2.1 5.2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.3 5.4

The askmarkets platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Experiment A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 5.4.1 5.4.2 5.4.3 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

5.5

Experiment B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 5.5.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 iii

5.5.2 5.5.3 5.6

Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

Experiment C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 5.6.1 5.6.2 5.6.3 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

5.7

A Deployment Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 5.7.1 5.7.2 5.7.3 5.7.4 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 Stocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

5.8

Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 5.8.1 5.8.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 129

Advanced Prediction Markets Concepts: Event Detection 6.1 6.2 6.3

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Background & Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 6.3.1 6.3.2 Text mining & stock market prediction . . . . . . . . . . . . . . . . . 132 Volatility analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

6.4

Time Series Analysis of Prediction Markets . . . . . . . . . . . . . . . . . . . 135 6.4.1 6.4.2 Machine Learning Techniques . . . . . . . . . . . . . . . . . . . . . . 135 Garch analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 6.4.2.1 6.4.2.2 6.4.2.3 Pre-estimation analysis . . . . . . . . . . . . . . . . . . . . . 137 Parameter estimation . . . . . . . . . . . . . . . . . . . . . . 144 Post-estimation analysis . . . . . . . . . . . . . . . . . . . . 145

6.5

Volatility Modeling for Detecting Information Flows . . . . . . . . . . . . . 149 6.5.1 6.5.2 Volatility Modeling and Prediction . . . . . . . . . . . . . . . . . . . . 150 Analyzing Variance Movements . . . . . . . . . . . . . . . . . . . . . 150 iv

6.6

Experimental Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

6.7 6.8

Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 6.8.1 6.8.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 157

Conclusions 7.1 7.2

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 189 191

Bibliography A Revision

A.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 A.2 Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 A.3 Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 A.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 Index 199

vi

List of Figures
2.1 2.2 2.3 2.4 4.1 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 6.1 6.2 6.3 6.4 6.5 6.6 Number of articles per term used to describe the concept of PM. . . . . . . Publication trend. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Classication of topics in prediction markets literature. . . . . . . . . . . . . Classication results of PM literature. . . . . . . . . . . . . . . . . . . . . . . 15 17 18 21

Breakdown Structure of Prediction Market Modeling. source: Chen et al. [2006b] 84 Homepage, screenshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Market page, screenshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Market creation page, screenshot . . . . . . . . . . . . . . . . . . . . . . . . . 110 Example closed market page with slow convergence, screenshot . . . . . . 112 Number of transactions per active user . . . . . . . . . . . . . . . . . . . . . 115 Market A, stock prices evolution, graph . . . . . . . . . . . . . . . . . . . . . 116 Market B, nal results, screenshot . . . . . . . . . . . . . . . . . . . . . . . . 117 Example market, stock prices evolution, graph . . . . . . . . . . . . . . . . . 120 SVM predictions vs actual results, Clinton. . . . . . . . . . . . . . . . . . . . 138 Autocorrelation of prices series, Clinton, 2004Q4. . . . . . . . . . . . . . . . 139 Autocorrelation of differences series, Clinton, 2004Q4. . . . . . . . . . . . . 140 Autocorrelation of returns series, Clinton, 2004Q4. . . . . . . . . . . . . . . . 141 Autocorrelation of squared returns series, Clinton, 2004Q4. . . . . . . . . . 142 Innovations, conditional standard deviations and returns series, Clinton, 2004Q4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

6.7

Standardized innovations of returns series, Clinton, 2004Q4. . . . . . . . . . 147 vii

6.8

Autocorrelation of the squared standardized innovations of returns series, Clinton, 2004Q4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

viii

List of Tables
1.1 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.1 3.2 3.3 4.1 5.1 5.2 6.1 6.2 6.3 6.4 Research approaches and questions . . . . . . . . . . . . . . . . . . . . . . . Type of Publications & Number of Articles . . . . . . . . . . . . . . . . . . . Number of Description Articles . . . . . . . . . . . . . . . . . . . . . . . . . . Number of Theoretical Work Articles . . . . . . . . . . . . . . . . . . . . . . Number of Application Articles . . . . . . . . . . . . . . . . . . . . . . . . . . Number of Law and Policy Articles . . . . . . . . . . . . . . . . . . . . . . . Classication of Reviewed Literature, (a) Description . . . . . . . . . . . . . Classication of Reviewed Literature, (b) Theoretical Work . . . . . . . . . . Classication of Reviewed Literature, (c) Applications . . . . . . . . . . . . Classication of Reviewed Literature, (d) Law and Policy . . . . . . . . . . 5 14 22 22 22 23 24 24 25 26 36 38 43 98

Properties of Market Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . Examples of Proper Scoring Rules . . . . . . . . . . . . . . . . . . . . . . . . Properties of Adequate Market Mechanisms . . . . . . . . . . . . . . . . . . Review of the market model . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Experiment A, survey & market results . . . . . . . . . . . . . . . . . . . . . 113 Experiment C, market results . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 SVMs Input Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 SVMs Output Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 SVM errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Ljung-Box-Pierce test on returns series, Clinton 2004Q4 . . . . . . . . . . . . 144 ix

6.5 6.6 6.7

Ljung-Box-Pierce test on square of returns series, Clinton 2004Q4 . . . . . . 144 GARCH model estimated parameters, Clinton 2004Q4 . . . . . . . . . . . . 145 Ljung-Box-Pierce test on square of standardized innovations series, Clinton 2004Q4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

6.8 6.9

Example of hits in news archive found, typical contract . . . . . . . . . . . . 153 Metrics of average results for all contracts . . . . . . . . . . . . . . . . . . . . 153

Chapter 1

Introduction
1.1 Scope
What men have seen they know; but what shall come hereafter, no man before the event can see. Sophocles Ajax, Chorus, lines 1417-1419, ca 450-440 BC [Jebb, 1896]. Since the oracles of Pythia and before, people have always wanted to forecast the future. Across geographies and ages, the uncertainty that future brings settles the need for its efcient forecasting. Responding to this life fact, modern age researchers and practitioners attempted to replace rituals with reason. Yet, this transition remains an ongoing process and the Thesis on hand aspires to further contribute towards it. Every decision requires a number of assumptions about the future. Predictions, in other words, are the ip side of decisions. In this context, the importance of forecasting, namely any statement about the future, arises as remarkable in decision sciences. Management, decision science and operations research serve as natural elds for the study and the advancement of forecasting problems. Forecasting is essentially a task of aggregation and processing. Ideally, one needs to collect all potentially related information, then lter only the relevant to the target variable parts and process the remaining data to arrive at conclusions. No matter the fascinating progress that has been recently made in handling such problems, in elds ranging from machine learning, to econometrics to social sciences, this expectably remains a cyclopean 1

CHAPTER 1. INTRODUCTION

task. One may ponder though about the existence of alternative mechanisms, that organically perform the above core tasks of information aggregation and processing in dynamic environments. Human intelligence seems to excel at most parts of these, yet it remains hindered by endemic problems, such as biases. That said, a number of ideas that could take advantage of collective intelligence and at the same time cancel out such limitations remain underexplored, provoking for further research. The scope of this Thesis is to study in depth such an approach, with the vision to help making it the default choice for a number of forecasting applications.

1.2 Motivation
The core approaches to the generic problem of forecasting are essentially three [Armstrong, 2001]. The rst set refers to causal methods. These focus on discovering correlations between the target variable and other potentially related ones [Pindyck and Rubinfeld, 1981]. Then, they apply these causal relationships to predict future prices of the target variable, assuming that such relationships do not change signicantly and provided that the prices of the correlated variables are given [Rodriguez Poo, 2003]. Causal methods have in many cases demonstrated a capacity to capture underlying relationships and provide sufcient results. However, it is quite common for such relationships to be difcult to discover, also for them to evolve over time, thus for causal methods to eventually lead to mediocre results [Fildes, 1985, Armstrong, 1978]. The second set of forecasting methods refers to time-series based ones [Makridakis et al., 1998, Brockwell and Davis, 1996]. Such approaches are established on the fundamental hypothesis that past data, usually limited to the target variable itself, contain enough information for the prediction of future prices. In other words, time-series approaches attempt to discover underlying patterns across past data and extrapolate them in the future, assuming that such patterns will remain valid [Lutkepohl, 1993]. As expected, time-series approaches perform generally well in relatively static problems. However, they typically prove to be incapable of sufciently capturing underlying patterns in dynamic environments, as a driving by the rear-view mirror approach suggests.

1.2. MOTIV ATION

The third set is comprised of judgmental methods [Wright and Ayton, 1987, Onkal-Atay et al., 2008]. In such approaches experts opinions and intuitive judgments are incorporated, via surveys or other similar tools, to arrive at estimates about the future. Typically capturing the view of a single expert, such methods serve as the cheapest and most common technique of forecasting in practice. They expectably come with some caveats though [Lawrence et al., 2006a, Sniezek, 1990]. It remains hard for an expert to formulate her knowledge and experience in a quantitive manner. It is also difcult for a forecaster to identify experts, then elicit and weight their various opinions [Ashton and Ashton, 1985, Batchelor and Dua, 1995]. Moreover, judgmental methods require substantial time to complete, resulting in a poor t for cases where time is of the essence. Articial intelligence (AI) approaches have more recently emerged, attempting to provide a more decent solution to some of the above issues. A wide variety of algorithms is put into use in both experiments and practice, mimicking among others the underlying principles of causal and time-series methods, but also judgmental ones (e.g. neural networks [Hill et al., 1994, Zhang et al., 1998]). Performance against some of the issues has improved, yet others problems have been raised, while the research eld remains denitely hot. Beyond any doubt though, the holy grail of a universal forecasting algorithm remains for the foreseeable future an elusive promise. Traditional forecasting approaches together with AI ones seem to have converged upon a ceiling that is far from easy to transcend, no matter the computational complexity of the algorithms being used. In this context, breakthroughs are rather to be expected in new research directions, challenging widely accepted assumptions while bringing together knowledge from discrete domains. Towards such a target, research on judgmental techniques seems shallow and less on par with the latest blossom of AI. A revisit of core judgmental practices to take full advantage of virtues while mitigating drawbacks might provide for contribution opportunities. More specically, their capacity to discover, aggregate and process relevant information, lying at the core of the forecasting problem, might be founded on a different basis, given an alternate approach to incorporate the knowledge of the many.

CHAPTER 1. INTRODUCTION

1.3 Structure
This Thesis will focus on the study of markets as a forecasting tool. The so-called prediction markets will be investigated from four diverse perspectives. Literature review, theoretical properties, theoretical evaluation and empirical analysis will each contribute towards understanding and utilization of markets as an essential tool for forecasting purposes. Table 1.1 provides a closer look at the research approach to be followed, as well as to the general and specic questions that will be attempted to address. Literature review will attempt to study all relevant to the topic publications that the author was able to locate by the time of writing. Theoretical properties will focus on the algorithmic foundation of markets as a prediction tool, contributing a generic mathematical framework and specic functions for its coherent operation. Theoretical evaluation will attempt to provide evidence on the theoretical soundness of the proposed mechanism. Empirical analysis will seek to expand such results into a wide set of experiments within and beyond the laboratory, along with giving shape to a framework for the practical implementation of markets for prediction purposes. Finally, advanced topics will investigate the applicability of prediction markets for event detection where continuous information ow is available.

1.3. STRUCTURE

Research Approaches & Questions Research Approaches General Research Questions Why & how do prediction markets work? Literature Review What is the literature on prediction markets? What are their fundamentals of operation? What is their evolution in recent years? What is the existing volume of publications? Which are the topics covered by literature? What is a proper Theoretical Properties prediction market mechanism? What market mechanisms are used in general? What prediction market mechanisms are used? Which are the properties for a coherent (proper) price function of a prediction market? Which are coherent price functions? Does a proper Theoretical Evaluation prediction market work in theory? Will a proper prediction market converge to a consensus equilibrium? If yes, how fast is the convergence process? What is the best possible equilibrium? Will a proper prediction market always converge to it? Does a proper Empirical Analysis prediction market work in practice? Is there a software platform supporting a proper prediction market mechanism? Is a proper prediction market able to converge to equilibrium in practice? What is a practical framework for prediction markets deployment? Are prediction markets Advanced Topics useful for event detection? What is the state-of-the-art in text mining for market prediction and volatility analysis? What is a model of volatility for prediction markets? How can such a model serve to detect events? Does such a model work in practice? Specic Research Questions

Table 1.1 Research approaches and questions

CHAPTER 1. INTRODUCTION

Chapter 2

Background
2.1 Overview
Following Chapter 1, the current Chapter provides the context within which this Thesis will be positioned. Markets and their generic functions are studied to begin with in Section 2.2. Then, in Section 2.3 focus is put on those market functions that stand as the reason detre of prediction markets. The denition and core properties of this forecasting mechanism is then followed by an extended literature review [Tziralis and Tatsiopoulos, 2007] that follows in Section 2.4. As part of this review, a classication of literature is also presented in Section 2.5, which might be of value for beginners or experts, also researchers and practitioners of the eld. Finally, Section 2.7 summarizes the outcomes of the Chapter while providing a number of suggestions for future work on top of it.

2.2 Markets, in general


[Aristotle, 1972] (ca 334-323 BC) suggested that when diverse groups all come together [...] they may surpass-collectively and as a body, although not individually-the quality of the few best. [...] When there are many who contribute to the process of deliberation, each can bring his share of goodness and moral prudence... some appreciate one part, some another, and all together appreciate all. 7

CHAPTER 2. BACKGROUND

Here, then, is a clear suggestion that many minds, deliberating together, may improve on the quality of the few best Sunstein [2008]. Provided the mass of the people is not too slave-like, each individual will indeed be a worse judge than the experts, but collectively they will be better, or at any rate no worse. ([Aristotle, 1972], ca 334-323 BC) Fast forward twenty three centuries, Hayek [1945] takes this further: Fundamentally, in a system in which the knowledge of the relevant facts is dispersed among many people, prices can act to coordinate the separate actions of different people. [...] We must look at the price system as such a mechanism for communicating information if we want to understand its real function. [...] if it were the result of deliberate human design, and if the people guided by the price changes understood that their decisions have signicance far beyond their immediate aim, this mechanism would have been acclaimed as one of the greatest triumphs of the human mind. Within this context, markets may be generally perceived as the best available mechanism for gathering and aggregating dispersed information from private, self-interested economic agents [Kagel and Roth, 1995, Hahn and Tetlock, 2005b]. Indeed, yet arguably, the most descriptive theory of market operations as of today is the efcient market hypothesis [Fama, 1970], suggesting that an efcient market continuously reects the sum of all available information about future events into security prices. On the same direction, the theory of rational expectations acknowledges the ability of markets to convey information through the prices and volumes of traded assets [Muth, 1961, Lucas, 1972, Grossman, 1981]. In the more recent decades, serial research attempts on experimental economics suggested that markets might be created specically to collect, aggregate and publish information, potentially requiring special market designs and institutions to leverage on such capacity [Plott and Sunder, 1982, Forsythe et al., 1982, Plott and Sunder, 1988, OBrien and Srivastava, 1991, Plott, 2001], also returning a Nobel prize in Economics to the pioneer of the eld, Vernon Smith [E-Museum, 2003, Altman, 2004].

2.2. MARKETS, IN GENERAL

The above essentially laid the foundations of what matured to serve the concept of markets for information and forecasting, also the core of the thesis on hand; however, before a closer look to the specic market function takes place, a rough depiction for markets in general follows hereby. Financial markets are institutions that incorporate by their very nature and facilitate, in one way or another, all of the four fundamental functions of investment, hedging, speculation and information aggregation. Several institutions have been deployed during the years to efciently support these functions, mostly by focusing on just one of them. To name a few, stock markets are created and operating with the primary purpose of capital allocation, futures markets scope in hedging risks and betting markets serve as a physical destination for wagering. However, each one of these institutions facilitates also and in parallel the entirety of the core market functions. For example, stock markets serve as a hospitable environment for speculation as well, while hedging could be accomplished by holding a wide portfolio of the traded assets and stock prices are unbiased estimators of rm fundamentals [Mandelbrot, 1966]. In this context, there exists signicant evidence supporting the general presence of the fourth native function of markets operation, namely their aggregative and predictive nature. This presence is strong even in institutions that were designed with different objectives in mind, in both investing [Admati and Peiderer, 1987, Chen et al., 2007, Grossman and Stiglitz, 1980b, Hellwig, 1980, Holmstrom and Tirole, 1993, Lo, 1997], hedging [Jackwerth and Rubinstein, 1996b, Krueger and Kuttner, 1996, Roll, 1984] and wagering cases [Boulier and Stekler, 2003, Debnath et al., 2003, Figlewski, 1979, Gandar et al., 1998, Schmidt and Werwatz, 2002, Thaler and Ziemba, 1988b, Winkler, 1971]. But, while there exist well-established and widely used market institutions to fully exploit each of the rst three essential features of markets operation, the fourth fundamental function of information aggregation was not a raison detre of a market till the comparatively recent emergence and development of prediction markets. A computational perspective to the information (or belief) aggregation problem is also applicable. Such an approach describes a number of agents (or experts) holding different and non-independent sets of information about an uncertain variable, with the target of designing a function to extract and summarize agents information, providing a

10

CHAPTER 2. BACKGROUND

collective estimate which ideally equals the omniscient forecast that has direct access to all the information available to all informants (a sensible consensus). The problem is a classic one in statistics and decision science, often studied under the concept of opinion pools and expert weights. Pennock and Wellman [1997] attempts to summarize a number of related references, the gist of which is well put by French [1985], suggesting that no proposed solution avoids a certain arbitrariness, also casting doubts for the very likelihood of a universally acceptable solution. In this context, the idea of using markets for the purpose of belief aggregation arose as early as forty years ago by Eisenberg and Gale [1959]. The researcher experimented with a typical pari-mutuel scheme, aggregating agents bets for a number of events, nally resulting in a consensus probability equal to the proportion of the total bet on each event, as Pennock and Wellman [1997] mentions. While Eisenberg and Gale [1959] demonstrated that such a mechanism returned a unique set of equilibrium probabilities as desirable, and Norvig [1967] gured out a dynamic process for reaching this equilibrium through iterated bids, Genest and Zidek [1986] observes that this approach has never enjoyed much popularity, due to the arbitrariness of the results in the case of low number of agents and volume. That said, however, and while the vast majority of work already performed on the information aggregation problem has nothing to do with the concept of a market, the inherent limitations of other solutions and references such as the above highlight the opportunities for a market solution and suggest its further exploration and exploitation, also from a computational perspective next to an economics and operations research one.

2.3 Fundamentals of Prediction Markets


Prediction markets emerged fairly recently as a promising forecasting mechanism, capable of efciently handling the dynamic aggregation of dispersed information among various agents. The interest that this mechanism attracts seems to be growing, both in terms of business applications and academic work [Tziralis and Tatsiopoulos, 2007], while it is by now accepted as a widely used forecasting mechanism [Armstrong, 2006]. This section provides a denition of prediction markets, next to a summary of their fundamental

2.3. FUNDAMENTALS OF PREDICTION MARKETS

11

properties that may serve as a short introduction to the concept. There is no universally accepted terminology and denition of prediction markets. After performing an extensive research on the existing literature, the following denition, based on the one given by Berg and Rietz [2003], is adopted and suggested for broader acceptance and use. Prediction markets are dened as markets designed and run for the primary purpose of mining and aggregating information scattered among traders; then transforming such information into market prices serving as predictions about specic future events. [Tziralis and Tatsiopoulos, 2007] Due to their unique objectives and nature, serving simultaneously as a forecasting tool and market mechanism, prediction markets feature a number of distinct characteristics, the most important of which are summarized hereby. First, prediction markets serve as an efcient way of arriving into consensus about the possibility of a future event. As Hanson [1992] rst observed, the mechanism may be the only available one to enable creating a consensus about future events in a way simultaneously open, egalitarian, honest, self-consistent and operationally cheap. Second, as also mentioned before, prediction markets serve as a well-working framework for belief aggregation. Pennock and Wellman [1997] suggest that markets incentivize self-motivated agents to gather all cost-effective relevant information and truthfully reveal their private beliefs. By setting up nancial securities that their return is related to future events, belief aggregation is performed in a natural manner and market prices may be interpreted as an aggregate probability of the participants beliefs. Third, and similar to the previous characteristic, prediction markets serve as an efcient mechanism with regards to information incorporation for future events. Market participants get informed by observing prices, thus information get spread, processed and incorporated by everyone for prices to uctuate as if everyone had access to all information [Pennock et al., 2002]. As a result, it is suggested that prediction markets may be used to efciently incorporate all related information to a future event, with empirical results owing in to conrm the appeal of such an approach [Chen et al., 2006b]. Fourth, in contrast to other market institutions, prediction markets accuracy can be assessed. Stock or future markets, for example, measure what will happen over an innite

12

CHAPTER 2. BACKGROUND

period, but this is not the case with prediction markets. Being related to a specic future event, prediction markets simultaneously suggest their probability of being mistaken but can also be judged based on that, when the predicted event gets materialized, or not [Einbinder, 2006]. Fifth, and the nal core characteristic is that, no matter how prediction markets are modeled or not, they effectively and largely stand as an empirical science. Human agents participation might be enough to testify that the study of prediction markets can be attributed to strictly be a social science, and this has to be taken into account in every approach that is not such. In this context, much of the existing literature, as it will be documented in the next paragraph, focuses on laboratory and eld experiments, testing prediction markets behaviors in various settings. Such attempts are also provided across this dissertation, however, and having said that, as happens with other species of markets, economic, computational and operations research theories and practices still apply and they will serve as the core perspective and set of tools for this thesis.

2.4 Literature Review


2.4.1 Context

This review surveys and examines the relevant existing literature and its trends, while it is also designed to provide a unique starting point for the further study of PM literature. Prediction markets provides an exciting eld for research, partly due to its novelty and growth. This section attempts a comprehensive review and classication of the literature on prediction markets research, starting from its introduction and the rst applications of the PM concept in the early nineties [Hanson, 1990b, 1992, Kuon, 1991, Forsythe et al., 1992] up until the writings of this section. The scheme used represents the authors view of the focus and direction of prediction market research and reveals a rapid growth in the number of published articles. The current state and direction of research topics should be of interest to many and it is hoped that it will serve both academics and practitioners as a point of reference for the study of the subject. The literature review is organized as follows. First, the research methodology is

2.4. LITERATURE REVIEW

13

described, followed by a commentary on the current diverse terminology existing in the prediction market eld, along with the evolution and growth of the literature and research itself. Subsequently, a classication method is introduced and its results are analyzed. The review nally concludes by presenting research implications and an extended list of prediction market references.

2.4.2

Research Methodology

This survey is the outcome of an attempt to collect and list an extended volume of prediction market related academic work. Before the publication of its results in Tziralis and Tatsiopoulos [2007], no relevant literature review was identied. Hitherto, there existed, for example, no publication outlet dedicated exclusively to prediction research. Therefore, the inclusion of various potential sources of academic knowledge dissemination was essential. As a result, all journal articles, conference proceedings papers, books or book chapters, masters theses, doctoral dissertations or other unpublished academic working papers and reports that are referring to the concept of prediction markets were collected, studied and are cited herein. The search was conducted mostly through the internet, as well as electronic libraries and academic databases. The literature review nally resulted in identifying 155 articles, which are classied by type of publication as shown in Table 2.1. By its very nature, this review could therefore be characterized as extended but by no means as exhaustive. Nevertheless, it may serve as a comprehensive basis for understanding prediction market research and its state of art.

2.4.3

Terminology

Prediction markets is not a unique and globally adopted descriptor of the concept and mechanism that was dened previously. On the contrary, the terminology used to address this concept is rather wide. The literature search was based on the following ve more usual and relevant descriptors: prediction markets, information markets, decision markets, electronic markets and virtual markets. Moreover, the references of each article found were further examined as to identify relevant citations that use perhaps

14

CHAPTER 2. BACKGROUND

Type of Publications Journal articles Books & book chapters Conference proceedings Masters theses & doctoral dissertations Working papers, reports & unpublished work Total

Number of Articles 58 (37%) 22 (14%) 15 (10%) 7 (4%) 53 (35%) 155 (100%)

Table 2.1 Type of Publications & Number of Articles another descriptor. The full text of each article was then reviewed to eliminate those articles that were not actually related to prediction markets. The nal selection of 155 articles was then classied on the basis of the prevalent descriptor used in each article to describe the concept of prediction markets. The distribution is depicted in Figure 2.1. Other descriptors which were identied during the research, include political stock markets [Berlemann and Schmidt, 2001, Bohm and Sonnegard, 1999, Brueggelambert, 2004, Forsythe et al., 1994, 1992, Hansen et al., 2004, Hauser and Huber, 2005, Jacobsen et al., 2000, Murauer, 1997, Ortner et al., 1995], election stock markets [Antweiler and Ross, 1998, Brueggelambert and Crueger, 2002, Forsythe et al., 1995, 1998, Kou and Sobel, 2004, Kuon, 1991], articial markets [Pennock et al., 2000, 2001a,b] and idea futures [Hanson, 1990b, 1992, Passmore et al., 2005b]. It becomes clear that the terminology used to describe the same concept is very diverse and extensive. This fact could lead to the division of the prediction markets community and its research products at a very early stage of its development and makes the agreement on globally accepted and standardized terminology all the more important, the author argues.

2.4.4

Early works and evolution

The series of articles that Hanson published between 1990 and 1992 [Hanson, 1990a,b, 1991, 1992] are the very rst introductory texts on the topic of prediction markets. The

2.4. LITERATURE REVIEW

15

Figure 2.1 Number of articles per term used to describe the concept of PM.

16

CHAPTER 2. BACKGROUND

earliest application of a prediction market mechanism, the Iowa Electronic Markets, was initiated in 1988 and was originally designed to study market dynamics while predicting the outcome of US presidential elections. Forsythe, Nelson, Neumann, and Wright [1992] wrote the rst academic article on the Iowa Electronic Markets in 1992. The early works of the nineties focused mainly on political stock markets applications. Aside from the papers on the most popular prediction market, the Iowa Electronic Markets [Berg et al., 1996, 1997, Forsythe et al., 1994, 1992, 1999], other election markets were described and analysed, like the one founded as early as 1990 in Germany [Beckmann and Werding, 1996, Kuon, 1991], as well as others in Canada [Antweiler and Ross, 1998, Forsythe et al., 1995, 1998], Austria [Murauer, 1997, Ortner et al., 1995] and Sweden [Bohm and Sonnegard, 1999]. Ortner also made important contributions to the eld with his doctoral dissertation in 1996 [Ortner, 1996] and the description of prediction markets rst application as a business tool by Siemens Austria in 1997 [Ortner, 1997, 1998]. Prediction markets literature up until 1998 is limited to mainly those above-mentioned articles. In the following years however, this survey witnesses a signicant increase in the volume of publications. The publication trend, as depicted in Figure 2.2, could be roughly described as being of exponential growth: the number of relevant articles in 2002 corresponded to 14, increased to 22 during 2004, while in the rst 8 months of 2006 there were already 34 published articles. Among these most recent articles it is of paramount importance to mention the pioneering work of Pennock on a dynamic pari-mutuel market framework [Pennock, 2004] and of Hanson on combinatorial market design [Hanson, 2003a]. Other equally signicant contributions, both in terms of citations and implications, were made by Spann and Skiera [2003a], Wolfers and Zitzewitz [2004a] and Berg and Rietz [2003] among others. This substantial increase of literature makes the need for further classication of articles in terms of their nature of prediction markets research indispensable. This need is addressed in the following section.

2.4. LITERATURE REVIEW

17

Figure 2.2 Publication trend.

18

CHAPTER 2. BACKGROUND

2.5 Classication Method


The classication or taxonomy framework, presented in Figure 2.3, was based on the above literature review and the nature of prediction markets research. The papers were classied into four broad categories: (a) description, (b) theoretical work, (c) applications and (d) law and policy; and each category is further divided into subcategories. It has to be stated, however, that this framework is designed to be rather practical than strictly documentary, serving as a navigation tool for researchers. Each paper was assigned to the category that describes most accurately the core of its prediction markets relevant contents alone. The categories breakdown is described hereafter.

Figure 2.3 Classication of topics in prediction markets literature.

2.5.0.1

Description

This category covers descriptive literature on prediction markets research, including introductory texts, general description, open questions, etc.

2.5. CLASSIFICATION METHOD

19

1. Introduction: This subcategory contains mostly short and rudimental texts on the basics of prediction markets, which are often a subsidiary topic of the publication. 2. General description: The subcategory covers lengthy and detailed articles that usually address the analysis of a variety of aspects on prediction markets. 3. Open problems: This consists of works that highlight issues which have yet to be addressed by the literature in a fully satisfactory way. 4. Other descriptive issues: This subcategory contains papers that discuss other descriptive issues on prediction markets, such as taxonomy, its potential use in education and other elds. 2.5.0.2 Theoretical work

The literature in this category includes papers of theoretical nature and is divided into the following three areas: 1. Market modeling and design: This contains various texts dealing with aspects on PM modeling, framework design and analysis. 2. Information aggregation convergence and equilibrium: The subcategory consists of papers discussing the convergence and equilibrium properties of the information aggregation process that is hosted by prediction markets. 3. Other theoretical issues: This includes works on other theoretical issues that could not be assigned to the previous two subcategories, such as the interpretation of prediction markets prices. 2.5.0.3 Applications

This broad category includes the totality of papers describing or analyzing applications of the prediction markets concept, either of experimental or practical nature. 1. Experiments: This is comprised of various experimental applications of the PM concept, held in academic or some other environment.

20

CHAPTER 2. BACKGROUND

2. Iowa Electronic Markets: The subcategory contains all the papers that focus on the description and analysis of results of the Iowa Electronic Markets. 3. Other political markets: This covers all the literature referring to political stock markets applications, with the exception of the Iowa Electronic Markets. The references include political stock markets in Germany, Canada, Austria, Sweden, Netherlands, Australia and Taiwan. 4. Markets on sport events: This subcategory comprises of articles of prediction markets applications in various sport events. Comparisons of real-money and play-money markets are also included in this subcategory. 5. Other applications: The subcategory contains the rest of applications that could not be assigned to any of the previous ones and includes among others business and entertainment web games applications. 2.5.0.4 Law and Policy

This last category consists of law and policy literature on PM research. 1. Legality and regulation: This subcategory is comprised of papers referring to aspects on the legality of PM and provides directions for their regulation. 2. Public policy and decision making: The works of this subcategory address the potential of prediction markets in improving policy analysis and public decision making. 3. The Policy Analysis Market: This covers all the literature describing the Policy Analysis Market, a prediction markets application that was designed to support policy analysis on sensitive political issues, such us international affairs and terrorism. 4. Other law and policy issues: This subcategory covers other law and policy aspects on prediction markets.

2.5.1

Classication Results

The 155 papers found were classied according to the above mentioned model. The distribution of articles by topics is shown in Figure 2.4. The majority of published research

2.5. CLASSIFICATION METHOD

21

concerns prediction markets applications (72 articles, 47%), whereas 36 articles (23%) were found to be mainly of descriptive nature and 27 (17%) of theoretical nature.

Figure 2.4 Classication results of PM literature.

Table 2.2 lists the number of description articles. 40% (13 articles) were general descriptions to the concept of prediction markets, while 30% (10 articles) were of introductory nature. Table 2.3 shows the number of articles categorised as theoretical works. The majority of them (16 articles, 59%) refers to market modelling issues, followed by 33% (9 articles) denoting to the study of convergence and equilibrium properties. Table 2.4 lists the number of articles of each PM application subcategory. 21 articles (29%) were written on other prediction markets than the Iowa political markets, 16 on Iowa Electronic Markets (22%), 15 on other applications (21%) and 13 on various experiments (18%). Table 2.5 shows the number of articles in law and policy related topics. Public policy and decision making was the dominant subcategory, as 55% (11 articles) were published

22

CHAPTER 2. BACKGROUND

Description Introduction General description Open problems Other descriptive issues Total

Number of Articles 13 (36%) 13 (36%) 5 (14%) 5 (14%) 36 (100%)

Table 2.2 Number of Description Articles

Theoretical work Market modelling Information aggregation convergence & equilibrium Other theoretical issues Total

Number of Articles 16 (59%) 9 (33%) 2 (7%) 27 (100%)

Table 2.3 Number of Theoretical Work Articles

Applications Experiments Iowa Electronic Markets Other political markets Markets on sport events Other applications Total

Number of Articles 13 (18%) 16 (22%) 21 (29%) 7 (10%) 15 (21%) 72 (100%)

Table 2.4 Number of Application Articles

2.6. CONCLUSIONS AND RESEARCH IMPLICATIONS

23

Law & Policy Legality & regulation Public policy & decision making The Policy Analysis Market Other law & policy issues Total

Number of Articles 4 (20%) 11 (55%) 4 (20%) 1 (5%) 20 (100%)

Table 2.5 Number of Law and Policy Articles on this topic. Tables 2.6, 2.7, 2.8 & 2.9 present a summary of all reviewed articles and assigns each of them to their respective subcategory. This is a helpful resource for anyone looking for prediction markets articles in a specic area.

2.6 Conclusions and Research Implications


As the nature of research on prediction markets is difcult to be limited to specic disciplines and the origin and growth of the literature is rather recent, the relevant material is scattered across various sources of academic writings. As a result, the research for this literature review was not focused exclusively on journal articles, but also extended to conference proceedings papers, books, book chapters, masters theses, doctoral dissertations and other unpublished academic working papers and reports. This literature survey was undertaken in order to identify as many as possible prediction markets related academic articles from various sources of prediction markets research. It resulted in the identication of 155 prediction markets articles published between 1990 and 2007. Although this review cannot claim to be exhaustive, it does provide reasonable insights into the state of the prediction markets research. The author feels that the results presented in this review have several important implications. (a) Based on the data presented and the existing trends, it is expected that prediction markets research and applications will signicantly increase in the future.

24

CHAPTER 2. BACKGROUND

(a) Description Introduction

Reference Boyle and Videbeck [2005], Cherry and Rogers [2006], Dahan and Hauser [2002], Hahn and Tetlock [2005a], Hahn and Tetlock [2006a], Kambil and van Heck [2002], Passmore et al. [2005a], Spann and Skiera [2003b], Sunstein [2006a]

General description

Abramovicz [2006b], Ankenbrand and Rudzinski [2005a], MacKieMason and Wellman [2006], Schrieber [2004], Soukhoroukova and Spann [2006], Spann and Skiera [2003a], Spann and Skiera [2004], Surowiecki [2004], Tziralis and Tatsiopoulos [2006], Wolfers and Zitzewitz [2004a], Wolfers and Zitzewitz [2006a], Williams [2005], Yang [2005]

Open problems

Bubb [2005], Hanson [2006a], Sunstein [2004], Sunstein [2006b], Wolfers and Zitzewitz [2006b]

Other descriptive issues

Ankenbrand and Rudzinski [2005b], Manne [2005], Passmore et al. [2005b], Pennock and Wellman [2001], Simkins [1999]

Table 2.6 Classication of Reviewed Literature, (a) Description

(b) Theoretical Work Market modelling and design

Reference Bergfjord [2006], Chan [2001], Chen [2005], Chen et al. [2006b], Fang et al. [2005], Fortnow et al. [2004], Hanson [2002b], Hanson [2003a], Kazumori [2004], McAdams and Malone [2005], Pennock [1999], Pennock [2004], Pennock et al. [2002], Pennock and Wellman [2001], Tetlock and Hahn [2006], Tetlock et al. [2005]

Information aggregation convergence & equilibrium Other theoretical issues

Berg et al. [2003], Feigenbaum et al. [2005], Gjerstad [2005], Hanson [2002a], Hanson and Oprea [2004], Koessler et al. [2005], Noeth et al. [1999], Ottaviani and Sorensen [2005], Pennock and Wellman [1997] Manski [2006], Wolfers and Zitzewitz [2006c]

Table 2.7 Classication of Reviewed Literature, (b) Theoretical Work

2.6. CONCLUSIONS AND RESEARCH IMPLICATIONS

25

(c) Applications Experiments

Reference Barner et al. [2004], Berlemann and Forrest [2002], Chan et al. [2002], Chan et al. [2001], Chan et al. [1999], Chen et al. [2001], Chen et al. [2003], Chen et al. [2004], Hanson et al. [2006], Jaisingh et al. [2002], Ledyard et al. [2005], Plott and Chen [2002], Rietz [2005]

Iowa Electronic Markets

Berg et al. [2000], Berg et al. [1996], Berg et al. [1997], Berg and Rietz [2002], Berg and Rietz [2003], Berg and Rietz [2006], Bondarenko and Bossaerts [2000], Erikson and Wlezien [2006], Forsythe et al. [1994], Forsythe et al. [1992], Forsythe et al. [1999], Fowler [2006], Kou and Sobel [2004], Oliven and Rietz [2004], Pagon [2005], Rickershauser [2006]

Other political markets

Antweiler and Ross [1998], Beckmann and Werding [1996], Berlemann and Schmidt [2001], Bohm and Sonnegard [1999], Brueggelambert [2004], Brueggelambert and Crueger [2002], Filzmaier et al. [2003], Forsythe et al. [1995], Forsythe et al. [1998], Hansen et al. [2004], Hauser, Hauser and Huber [2005], Jacobsen et al. [2000], Kuon [1991], Leigh and Wolfers [2006], Murauer [1997], Ortner [1996], Ortner et al. [1995], Rhode and Strumpf [2006], Wang et al. [2005], Wolfers and Leigh [2002]

Markets on sports events

Bean [2005], Chen et al. [2005], Debnath et al. [2003], Rosenbloom and Notz [2006], Schmidt and Werwatz [2002], Servan-Schreiber et al. [2004], Smith et al. [2005]

Other applications

Gruca [2000], Gruca et al. [2001], Gruca et al. [2003], Gruca et al. [2005], Gurkaynak and Wolfers [2006], Mangold et al. [2005], Ortner [1997], Ortner [1998], Pennock et al. [2000], Pennock and Wellman [2001]. Skiera and Spann [2004], Snowberg et al. [2005], Snowberg et al. [2006], Soukhoroukova and Spann [2005], Tetlock [2004]

Table 2.8 Classication of Reviewed Literature, (c) Applications

26

CHAPTER 2. BACKGROUND

(d) Law & Policy Legality and regulation Public policy and decision making

Reference Abramovicz [1999], Bell [2002], Bell [2006], Hahn and Tetlock [2006b] Abramowicz [2004], Abramovicz [2006a], Einbinder [2006], Hahn and Tetlock [2004], Hahn and Tetlock [2005b], Hahn and Tetlock [2005c], Hanson [2003b], Hanson [2006c], Ledyard [2006], Wolfers and Zitzewitz [2004b]

The Policy Analysis Market Other law and policy issues

Hanson [2005], Hanson [2006b], Meirowitz and Tucker [2004], Polk et al. [2003] Abramovicz [2004]

Table 2.9 Classication of Reviewed Literature, (d) Law and Policy (b) There is a strong need to standardize the terminology used to refer to the prediction markets concept. (c) The formation and dissemination of a fully appropriate prediction markets mechanism, such as the dynamic pari-mutuel presented by Pennock [2004], could lead to the expansion of prediction markets research and applications.

2.7 Discussion
2.7.1 Summary

This Chapter provided a close look to the concept of prediction markets, serving as the context on top of which this thesis will be deployed. Starting with a general description of the mechanism and a short summary of the developments that led to it, it proceeded to outline the mechanisms core characteristics and its differentiation to other existing market tools or forecasting approaches. After that, an extensive literature review on prediction markets was performed and deployed at full detail, followed by a classication of the literature that may serve as a stand-alone resource for anyone willing to study a subset of the topic. In the authors knowledge there does not exist a similar to Tziralis and Tatsiopoulos [2007] work on prediction markets literature published so far, so this might

2.7. DISCUSSION

27

stand as a contribution of value.

2.7.2

Future work

Introductory material may seem to be the most limited one in terms of suggesting directions for future work. However, this is not the case for the thesis at hand. Prediction markets stand as a relatively recent topic with a wide array of introductory material already existing, as this section documented. That said, it is clear to the author that the concepts description itself has yet to become widely understood and, thus, utilized to bring its potential into reality. This might be considered to serve as a mostly commercial, rather than purely academic, target, but it is in the belief of the author that prediction markets becoming a mainstream tool and approach for forecasting will result into a further growth of the related research as well. Popularizing the prediction market concept largely remains an unmet challenge for both the research community and prediction market industry alike, while the author hopes this section to serve as a trigger for more people interested to work towards that direction. Next to that, other research orientations include the expansion of literature review with future publications, also creating a different taxonomy for alternate usage.

28

CHAPTER 2. BACKGROUND

Chapter 3

Theoretical Properties of Prediction Markets


3.1 Overview
Chapter 1 set the generic forecasting problem to address; Chapter 2 shaped out the market framework towards its solution. The current Chapter proceeds to the algorithmic perspective of the problem on hand and the mechanism that orchestrates the operation of a prediction market, standing as the theoretical core of the Thesis. The mechanisms in use by well-established market institutions, focusing on all but the market function of information aggregation, appear not to be suitable for that latter one. These mechanisms typically include Continuous Double Auction, Market Maker and Pari-mutuel Market, each one of them appearing to have at least one signicant drawback for the use under focus. The need for a more appropriate mechanism was addressed by the recent introduction of Market Scoring Rules by [Hanson, 2003a] and Dynamic Pari-mutuel Market by [Pennock, 2004]. The current Chapter moves forward to introduce some additions and shape a coherent framework of properties for a prediction market mechanism, further addressed by a couple of market functions which are also deployed. This contribution intends to provide an alternative to the core market problem, which is thoroughly studied under various perspectives in the Chapters to follow. The Chapter starts with Section 3.2, discussing the typical market mechanisms in use 29

30

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

by well-established market institutions, next to their drawbacks for use in a prediction market setting. Section 3.3 proceeds to deploy the state-of-the-art mechanisms of Market Scoring Rules and Dynamic Pari-mutuel Market, while highlighting some directions for further improvement. Section 3.4 builds on these directions to contribute a framework of properties for a coherent price function in a DPM setting. Then, Sections 3.5 & 3.6 introduce such functions, proving their partial or fully compliance with the proposed framework. Finally, Section 3.7 sums up and discusses the Chapters contributions, while providing directions for future deployment.

3.2 Market Mechanisms


3.2.1 Description & Characteristics

During the years, a wide variety of nancial and wagering mechanisms have been developed and matured to support various operations of well-dened market institutions, as already shaped out in Section 2.2. Money allocation, speculative trading (i.e. wagering) and hedging (i.e. insuring) against exposure to uncertain events, next to information aggregation, all are operations facilitated in one way or another within an extended set of institutions, which have matured over time. However, the variance of market institutions occurs not only due to the differentiation of their scope of operations (what is being traded, and why), but also because of the various nancial and wagering mechanisms that have been developed to efciently serve each one of these operations (how it is traded). This section describes and studies the operational properties of these mechanisms, providing the needed input for assessing their suitability of usage in the case of prediction markets. To authors knowledge and according to Pennock [2004]s sufcient review, the main and most used market mechanisms are: (a) CDA: Most modern nancial markets use a continuous double auction mechanism to store and match orders and facilitate trading [Smith et al., 2003]. The dominant mechanism used in nancial circles is the continuous double auction (CDA). (b) Market-maker: The primary mechanism used for sports wagering is a bookie or

3.2. MARKET MECHANISMS

31

bookmaker, who essentially acts exactly as a market maker. A CDA with market maker (CDAwMM) is also common in nancial circles. (c) Pari-mutuel: Horse racing and other wagering games traditionally employ the pari-mutuel mechanism. The next paragraphs describe in more detail the operational properties of each one of these market mechanisms, followed by a study on their suitability for information aggregation purposes.

3.2.2

Continuous Double Auction

The continuous double auction (CDA) stands as a very popular and widely used market mechanism. [Smith et al., 2003] note that it is used by most of the modern nancial markets to store and facilitate trading; examples include stock markets, option and other derivatives markets, insurance markets and market games [Jackwerth and Rubinstein, 1996b, Roll, 1984, Forsythe et al., 1992, 1999, Hanson, 1990a, Chen et al., 2001, Pennock et al., 2001a,b]. Moreover the CDA has been documented to effectively perform the tasks of information aggregation and prediction of future events [Plott and Sunder, 1988, Copeland and Friedman, 1987, 1991, 1992, Forsythe and Lundholm, 1990, Friedman, 1993, Sunder, 1992, 1995, Plott, 2000]. The building block of a CDA is the simple call market auction. In such an auction, bids arrive asynchronously and are collected over time, then processed together in large batches. If m the number of sellers, the clearing price is typically dened as equal or in between of the mth and m + 1th lowest prices. A CDA is a continuous version of the call market, where a transaction is immediately executed at the time that a trade is acceptable by any two bidders (usually at the bid price of the least recent bidder). In more detail, if at any time a trader is willing to buy one unit of an asset at a bid price pbid , while another party is willing to sell one unit of the asset at an ask price p ask , then any case that pbid > p ask results into a transaction (typically at some price between pbid and p ask ). In other words, buyers can only buy as many shares as sellers are willing to sell, and for any transaction to occur, there must be a counterpart on the other side willing to accept the trade. In this context, the auctioneer takes on no

32

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

risk as her only function is to match willing bidders, and this serves as another signicant advantage of the mechanism. However, in the case that the highest bid price is less than the lowest ask, nothing occurs. As a result, it is expected for prices in a CDA-powered market either to remain illiquid or rapidly change as new information arrives and traders reassess the value of an asset. Illiquidity is rather typical especially when only few traders participate, turning to be a signicant drawback of CDA. Moreover, the spread between the highest bid price and the lowest ask price may vary to arbitrarily large values, or one or both queues may be completely empty, further discouraging trading. This chicken and egg scenario of scarce traders do not participating at all because the expected time for a trading partner to arrive is indenite is also known as the thin market problem. This is particularly highlighted in combinatorial markets, where the number of available stocks is extensive, further squeezing the likelihood of a match between traders in a specic stock. The above problems result in reduced incentives for traders to participate, and limited informativeness of the market itself.

3.2.3

Market-maker

To address CDAs problems, the mechanism is often used with a market maker (CDAwMM) to induce liquidity. The Market Maker (MM) is an agent who is nearly always ready to trade, willing to accept a much larger volume of buy and sell orders than typical traders. MM may be a person or automated algorithm, posting various bid and ask prices for other traders to trade with. In other words, a single trader is enough for a transaction to take place under the existence of a MM, who efciently cancels out the thin market problem and other CDAs drawbacks, resulting in a continuous and uninterrupted ow of information. The typical bookmaker, the default mechanism for wagering, also functions like a MM in a CDA. The main difference lies in the direction of the transactions available; in the bookmaker case the market institution sets the odds only for other players to buy and not to sell, in a take it or leave it fashion. The odds remain xed at the time of a bet, initially dened according to expert opinion and later updated in response to the relative

3.2. MARKET MECHANISMS

33

level of wagering on the various outcomes Pennock [2004]. Both MM and the bookmaker have also a strong record of success in the information agreggation task [Jackwerth and Rubinstein, 1996b, Krueger and Kuttner, 1996, Roll, 1984]. It is expected though for MMs performance to come at a price; and this is exposure to risk of losing considerable amounts of money for the MM. As the mechanism does not only provide the matching functionality, there is an operational cost or benet associated with the performance of the trader that essentially a MM is, depending on the evolution of prices in stocks that the MM comes to own. This risk is not necessarily a negative property, providing a proper way of injecting subsidies into the market Pennock and Sami [2007]. In this way, informed traders rationally expect to receive a prot out of their transactions. However, it is of specic importance for the operational risk of a MM to be bounded, next to the normal requirements of providing proper incentives for truthful information submission by traders, and a computationally tractable way of determining the new prices after any transaction. To achieve this, a fully systematic way for the adjustment of bid and ask prices after every trade needs to be set up. The result is known as an automated market maker.

3.2.4

Pari-mutuel

The third market mechanism available is the pari-mutuel market (PMM) one. Pennock [2004] describes its usage as common at horse races [Ali, 1977, Rosett, 1965, Snyder, 1978, Thaler and Ziemba, 1988b, Weitzman, 1965] and other similar wagering games. In such a market, people place wagers on which of two or more mutually exclusive and exhaustive outcomes will occur at some time in the future. When the true outcome becomes known, players who wagered on the correct outcome split the total amount of money invested in proportion to the amount they have wagered, while those invested in all other outcomes lose their money (in practice, the total amount redistributed is diminished, for the market institution to receive a fee). In essence, wagers compete against each other to receive the largest share possible out of all money invested. In mathematical terms, let n be the mutually exclusive and exhaustive outcomes (e.g., n horses, exactly one of which will win), while M1 , M2 , . . . , Mn euro in total are bet on

34

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

each outcome. If outcome i occurs, then everyone who bet on an outcome j = i loses their wager, while everyone who bet on outcome i receives an equal amount of n j = 1 M j / Mi euro for every euro they wagered (in practice, the total amount redistributed is lowered due to institutions fees). In other words, the cost of purchasing an equal share of the winnings remain constant (for example, 1 euro) [Pennock and Sami, 2007], and it is not conditional to the time the wager was placed, or the amount of money invested in the various outcomes, but just depends on the nal amounts wagered on all outcomes when the market closes, next to the identity of the correct outcome. Pari-mutuel markets have a number of different operational characteristics when compared to CDA or market-maker. Unlike a CDA, in a pari-mutuel market anyone can place a wager of any amount at any time, without the need for a matching offer from another bettor or a market maker, providing in a sense innite liquidity for buying. Moreover, unlike a CDAwMM, this liquidity comes at no risk for the market institution, since money is only redistributed from losing wagers to winning wagers. On the other hand, pari-mutuel markets are not suitable for situations where information arrives over time, since there is no incentive for buying until either all information is revealed, or the market is about to close [Pennock and Sami, 2007]. By not allowing bettors to prot before the market closes through predicting future demand (aka buying low and selling high, as occurs in a CDA), prices in a standard pari-mutuel mechanism cannot be considered an accurate reection of current information, and this results in a questionable applicability in a large number of settings [Mangold et al., 2005]. Moreover, the pay-off rate of a particular bet uctuates arbitrarily according to the amount of money wagered on the particular outcomes at each time (while remaining positive if the correct outcome is chosen), and bettors need to wait to get a better sense of the nal pay-out rates before wagering [Pennock, 2004]. That said, pari-mutuel markets ability to disseminate information, at least for a innitesimal time window, is remarkable [Plott et al., 1997], and the paragraphs to follow move on to develop a mechanism leveraging on their benets, while addressing the drawbacks deployed above.

3.3. PREDICTION MARKETS MECHANISMS

35

3.3 Prediction Markets Mechanisms


3.3.1 Disadvantages for use in prediction markets

The previous section described in detail the operational characteristics of the most widespread market mechanisms. All three of them were also documented to perform well in the information aggregation task, a fact suggesting their suitability for facilitating prediction markets. At the same time, though, specic disadvantages prevent each one of these mechanisms from being an ideal t. In the next paragraphs these drawbacks are summarized and their implications are discussed, while what follows is a description of some more advanced mechanisms recently deployed to efciently support the operation of prediction markets. A CDA matches willing traders, so poses no risk for the market institution, while typically provides a motive for traders to participate as soon as new information arrives and traders reassess the value of an asset. At the same time, however, CDA is vulnerable to the thin market problem, resulting in illiquidity in the case that only few traders participate, a scenario rather usual for a prediction market. The expectation of no or low liquidity may also diminish the willingness of a trader to participate immediately when she receives new information, further reducing the suitability of the mechanism. MM and the similar bookmaker mechanism address the above problems, by operating as an agent willing to trade in both buy and sell orders (in the general case), or only in buy orders (in the case of a bookmaker). As a result, the market always remain liquid, and information capture is dynamically performed according to traders demand. However, this comes with a potentially signicant risk for the market institution running the MM, as inducing liquidity may result to large monetary losses, which are expectably anything but desired in a real money market setting. Pari-mutuel markets, standing as the third option available, succeed in both canceling out any risk for the market institution, and continuously providing liquidity to the marketplace. By limiting the amount of money redistributed to the amount of money invested, the market institution can only receive prots (equal to the transaction costs that are potentially introduced). On the other hand, liquidity does not necessarily imply current information capture, and this is the inconvenient case with a pari-mutuel market. The lack

36

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

No risk for the mar- Innite liquidity ket institution CDA Market maker Pari-mutuel yes no yes no yes yes

Current information capture yes yes no

Table 3.1 Properties of Market Mechanisms of any proper incentive for the trader to submit her bet as soon as new information arrives results into the inappropriateness of simple pari-mutuel markets on serving a prediction market institution as its empowering mechanism. Table 3.1 provides a synopsis of the above and a concrete view of the positive and negative features of each market mechanism; the lack of a fully appropriate market mechanism for the usage under focus becomes apparent. This lack was recently addressed by the introduction of Market Scoring Rules and Dynamic Pari-Mutuel Market mechanisms, the detailed description of which follows hereafter.

3.3.2

Market Scoring Rules

Description The building block of Market Scoring Rules (MSR) is, not surprisingly, Scoring Rules. Scoring rules serve as an alternative mechanism to prediction markets for tackling the information aggregation problem; in fact, their concept and origin is much older, as the original Savage [1971] paper manifests. Hanson [2003a] introduced an innovative interpretation of scoring rules, for them to empower the function of a market mechanism. More specically, Hanson showed how any proper scoring rule can be interpreted as an automated market maker, the latter being conceived as a sequential shared version of the core scoring rule [Pennock and Sami, 2007]. In what follows, a more detailed description of scoring rules is provided, followed by an elegant narrative of an automated MM, to nally enable a proper deployment and understanding of MSR. Finally, the extended pros and cons of MSR for the under study prediction problem are reported and discussed.

3.3. PREDICTION MARKETS MECHANISMS

37

Scoring rules were built with a twofold aim in mind [Schervish, 1989, Clemen, 2002]: In an ex-ante sense, they are designed to give agents proper incentives for acquiring information and making careful and honest assessments of it. At the same time and in an ex-post sense, scoring rules serve as devices for eliciting and evaluating subjective probabilities. Typically discussed across the literature under the latter perspective, strictly proper scoring rules can be used for assessing forecasts, by assigning a numerical score based on the forecast and the actual outcome of the event or value occurred [Schervish, 1989]. Denitions More specically, a scoring rule is strictly proper if the expected score is maximized by the forecaster for an observation drawn from the distribution F, when she issues the probabilistic forecast F rather than any G = F. Moreover, a strictly proper rule enables attractive loss and utility functions that can be tailored to the problem at hand [Clemen, 2002]. Providing a proper mathematical documentation to the above by following the set-up of Savage [1971], Hanson [2002b, 2003a], consider an expected-utility maximizing agent, with subjective beliefs pi on a complete set of n mutually exclusive and exhaustive events i (i pi = 1). The agent is expected to receive a cash amount x, dened according to a proper scoring rule s = {si }i satisfying xi = si (r ), where xi the cash payment of event i, ri the probability assigned to i by the agent, and r = {ri }i her full report. s is a proper scoring rule, if a risk neutral agent with event-independent utility maximizes her expected monetary returns by reporting her true expectations, namely ri = pi . More properly, p=

pi si (r) given ri = 1
i i

Examples of proper scoring rules are presented in Table 3.2 that follows. Hanson [2002b] also reports that power law rules are proper scoring rules for a 1, while both the quadratic and logarithmic rules are special cases of the power law rules, for a of 2 and 1 respectively. Scoring rules perform well against the thin market problem by design, however, a formula on extracting a single consensus estimate out of probability reports submitted by a number of experts is not straightforward, if not problematic. In the ideal case, though, the available information by all traders would eventually converge to identical estimates,

38

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Quadratic Spherical Logarithmic Power Law

si = ai + bri b j r2 j /2
1/2 si = ai + bri /( j r2 j)

si = ai + b log(ri ) si = ai + ba
r i a 2 ( i ) d i 0

b j ra j

Table 3.2 Examples of Proper Scoring Rules provided that rational agents are repeatedly made aware of each others estimates, given common priors [Geanakoplos and Polemarchakis, 1982, Hanson, 1998, 2002b]. MSR came to provide an elegant solution to this problem. Starting with an initial report and prices set by the market maker, each new trader t that arrives agrees to compensate the previous trader t 1 according to her previously submitted probability estimate, while receiving the scoring rule payment associated with the probability estimate that she submits by herself [Pennock and Sami, 2007]. In a sense and under this prism, the scoring rule actually enables a stepwise and sequential process, while supporting the liquidity and current information capture prerequisites described before, and in parallel limiting the risk undertaken by the (automated) market maker. At the same time, a MSR always provide a single consensus probability distribution over the entire state space, each part of which remains available for the next trader to change [Hanson, 2002b, 2003a]. To provide a more proper description to the above, following the interpretation of [Pennock and Sami, 2007], let j be the relevant security paying off for the event i, following the notation already deployed. If q j is the total quantity of security j held by all traders combined, then the market maker essentially implements a cost function C (q) tracking the total amount of money traders have spent as a function of the total number of shares held of each security. In this context, each transaction of any trader purchasing (or selling) shares of security j costs C (qnew ) C (qold ) = C (q1 , . . . , q j + , . . . , qi ) C (q) monetary units, with the instantaneous price of each security j given by C /q j (the cost per share for purchasing an innitesimal quantity), and the full cost of purchasing any nite quantity computed by

C (z)dz.

3.3. PREDICTION MARKETS MECHANISMS

39

For a given scoring rule, the computation of the cost function becomes tractable. Hanson [2003a] proposed the logarithmic one, for which the cost and price functions respectively become: C (q) = b ln exp(q j /b)
j

exp q j /b C = q j k exp (qk /b) The parameter b serves to both shape the sensitivity of the market, as well as bound the risk of the market maker. More specically, the larger the b is, the smaller the change a given amount of investment induces, and vice versa. It can also be showed that the market maker loss is bounded at b ln|n|, where n the number of available outcomes [Hanson, 2002b]. Finally, when the true outcome becomes known, traders holding the winning security receive one monetary unit for each share they own. Discussion After all, the MSR functions as a simple scoring rule when dealing with

a single trader, while transforming into an automated market maker in the case of a number of traders participating. At the same time, the mechanism succeeds into carrying over all advantages of the simple scoring rule into an automated market maker setting, thus sustaining the core advantages of the latter, while signicantly limiting its core disadvantage. In other words, the MSR, and especially the logarithmic one, exhibits liquidity in both small and large quantities of buy and sell orders, plus it provides sufcient incentives for current information capture. On top of that, the operational risk for the market maker is now controllable, in practice being equal or less to the maximum possible payment to the nal trader minus the payment received from the rst trader. These benets make the MSR really appealing, and the mechanism is indeed utilized in a number of real world settings [Pennock and Sami, 2007].

3.3.3

Dynamic Pari-Mutuel Market

Description Dynamic Pari-mutuel (DPM), introduced by Pennock [2004], is designed as a hybrid between the classic pari-mutuel and CDA mechanisms, trying to avoid the disadvantages of both, while sustaining their benets. DPM in essence functions like a

40

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

pari-mutuel one (or, from a traders perspective, similar to a market maker) by always enabling purchases of each outcome, while resembling the CDA in that prices increase with demand [Mangold et al., 2005, Pennock and Sami, 2007]. In more detail, the market operation starts with the market institution submitting a pre-determined subsidy, in the form of an initial amount, for each stock to begin with an equal, non-zero investment on it (in practice, this amount can be arbitrarily small and is overbalanced from transaction fees and other commissions). Traders are then able to purchase, at any time, any number of shares they want to, similarly to a pari-mutuel market. Selling, however, is not equally exible in a DPM, as it is only available to traders already owning an amount of shares, for them to sell it directly back to the market maker, in a fashion similar to CDA. Alternatively, the inability of short selling (selling shares a trader does not own) can be circumvented -though admittedly in a less straightforward fashion- by buying shares of all other stocks, therefore hedging a traders risk by owning a stake of all complementary outcomes. A critical difference in the DPM case is that prices are not stable but vary over time, being automatically upgraded by the market maker according to a predened function, based on the current stage of wagering. More specically, buying shares of an outcome results into its stock price increasing, while the price of other outcomes moves to the opposite direction. The price function may vary according to the properties needed and its complexity is not required to be communicated to traders, as it is only its outcomes who matter each time. Stock prices get updated for every new share traded, and the total cost of a transaction occurs as the integral of the price function, from zero to the number of shares purchased or sold each time. When the market closes, all money invested in the various outcomes get redistributed to those traders owning shares of the winning stock, proportionately only to the number of shares they hold, in a pure pari-mutuel way. Given that the amount a share of a winning stock nally returns remains steady and non-related to its buying price, the mechanism provides improved returns to those traders buying early (when stocks prices are low), therefore bringing them to submit their information in a real time fashion. In this way, DPM succeeds in enabling current information capture, thus invalidating the core disadvantage of the simple pari-mutuel mechanism, while retaining all of its benets.

3.3. PREDICTION MARKETS MECHANISMS

41

Denitions To provide a solid mathematical basis to the above, following the set-up and notation of Pennock [2004], lets consider, for the sake of simplicity, the binary case of two mutually exclusive and exhaustive outcomes, A and B. Let, also, p A and p B be the instantaneous price per share at time t of outcomes A and B respectively, while PA and PB the payoffs per share, if the winning outcome is revealed and the market closes at t. Moreover, let M A the total amount wagered on outcome A, MB the total amount wagered on B and T = M A + MB the amount wagered on both outcomes and the market in general till time t; also NA the total number of shares purchased of A and NB the total number of shares purchased of B. Given that, when the market closes, all money invested in the various outcomes get redistributed to the winning outcome, the payoff per share equals, respectively, PA = M A + MB T M A + MB T = , PB = = NA NA NB NB

In other words, if A occurs, traders who own shares of A at time t lose their initial money wagered, but receive PA per share owned, while shareholders of B simply lose all money wagered; the opposite applies if B is the winning outcome. A risk-neutral trader is now expected to buy a number of shares of a stock in the case that her expected returns are bigger that the purchase cost. The cost of buying n shares of A (or B) occurs out of the continuous interval of the price function from 0 to n,
n 0

p( x )dx,

while the price function may vary according to the desirable properties of the market institution. The traders optimization problem, which is more perplexed in the DPM case, compared to other more widespread mechanisms like the CDA, is to select n so as to maximize the quantity Pr( A) n PA
n 0

p A ( x )dx

where Pr( A) is the probability of outcome A occurring, and Pr( A) n PA is the expected earning of n shares of A. In the following, a couple of price functions, proposed by Pennock [2004], Pennock and Sami [2007] are described, while their operational pros and cons are discussed, for a more convenient solution to be proposed. The core consideration of Pennock [2004]s function is the price of one stock being proportional to the amount of money wagered on it. That, in the general case, would result in the price function p A = aM A , which is straightforward but results in non-bounded,

42

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

ascending prices, a non-desirable property in most cases. A simpler variation is for the ratio of prices being equal to the ratio of money wagered (money-ratio function), namely p A / p B = M A / MB . Out of this condition, Pennock [2004] computes the instantaneous price per share as equal to pA =

( M A )2 + M A M B M A + MB = MB M A NA + MB NB NA + M NB A

while, given the price function, the number of shares n that can be purchased for m monetary units equals n= T ( M A + m) N ( T + m) m( NA NB ) + B ln T MB M A ( T + m)

and, nally, the price as a function of m becomes p A (m) = where denom = ( M A + m) MB NA + ( MB m) MB NB + T ( M A + m) NB ln T ( M A + m) M A ( T + m) dm ( M A + m) MB T = , dn denom

Another function proposed by Pennock and Sami [2007] is the share-ratio one, where the ratio of instantaneous prices is proportional to the number of shares purchased for each outcome. In the two-outcomes case, this is interpreted as N pA = k A pA = k pB NB For k = 1, the share-ratio function simplies to pA N = A pA = pB NB NA
2 + N2 NA B

NA
2 + N2 NA B

and the core fundamental equation p A (m) = dm/dn results in


n

m ( n ) = m (0) + while n(m) now equals

NA + s
2 ( NA + s)2 + NB

ds =

2 ( NA + n)2 + NB

2 + N2 NA B

n(m) =

2m

2 + N 2 + N 2 + m2 N NA A B A

3.4. PROPERTIES FOR A COHERENT PRICE FUNCTION

43

Risk for the mar- Innite ket institution Market Scoring Rules Dynamic Pari-Mutuel limited no liquidity yes yes

Direction of in- Current informanite liquidity buy/sell buy tion capture yes yes

Table 3.3 Properties of Adequate Market Mechanisms Discussion DPM succeeds in all core properties of removing risk for the market institu-

tion, providing innite liquidity, and capturing current information. Moreover, it provides participants with the option to sell any amount of shares they own at any time, just like traditional stock markets, in contrast to a standard pari-mutuel mechanism. At the same time though, it features a couple of practical drawbacks. First, the expected payoff per share is a function of the price at the time of the transaction and the nal payoff at the time of closing. The latter is not predened, as the case is in MSR for example, often resulting in less intuitive market results. The complexity of this function also stands for a greater difculty in a traders optimization strategy for expected returns. Second, short-selling, namely selling of shares that one does not already own, is not directly possible. There is a workaround if one buys a portfolio of all other existing stocks instead of short-selling shares of the one she considers overpriced, yet again the complexity of this process usually proves to be unpleasantly expensive. No matter its drawbacks, DPM, along with MSR, stands as an adequate mechanism for prediction markets usage. Table 3.3 summarizes the properties of both DPM and MSR. After all, the substantial difference between them lies in MSR paying a xed monetary unit per share versus DPM paying per share an equal portion of the total amount outstanding in the market at the time of closing [Pennock and Sami, 2007].

3.4 Properties for a Coherent Price Function


Focusing further on DPM, and no matter the generic adequacy of the mechanism, the price function each time used arises as of signicant importance. [Pennock, 2004] suggests a couple of price functions, the money-ratio and the share-ratio ones, namely p A / p B =

44

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

M A / MB & p A / p B = kNA / NB respectively in the simple case of two outcomes. At the same time he claries though that its not clear which price function is better or if a best one exists, inclining towards further research. In response to this need, this Section will focus on shaping a framework of properties for what it will be called a coherent price function for a DPM prediction market mechanism. Functions satisfying this set of rules will provide for clarity, brevity, intuitiveness and elegance, aiming to become the standard set of properties for future such quests.

3.4.1

Simple case: n = 2 outcomes

Notation The following notation in the case of n = 2 outcomes will be consistently used in this and the following Sections of this Chapter. outcomes: A, B (mutually exclusive and exhaustive) price: p A , p B (instantaneous, per share) payoff: PA , PB (per share) money: M A , MB (total amount wagered) shares: NA , NB (total number purchased) At time t At time t known are the number of shares and money invested per outcome,

namely NAt , NBt & M At , MBt . Based on these, the expected payoff (the payoff that would be distributed if the market was to be closed at time t) per share is given by the formula: PAt = M At + MBt Tt M At + MBt Tt = , PBt = = NA t NA t NBt NBt

Given the known variables, price is a function of p At = f ( M At , MBt , NAt , NBt ) =? At time t + dt Case 1: buy/sell n shares of A by paying m money (where m unknown) In this case known parameters are MBt +dt = MBt , NAt +dt = NAt + n, NBt +dt = NBt , while unknown parameters that will be computed below are M At +dt , PAt +dt , PBt +dt , p At +dt , p Bt +dt .

3.4. PROPERTIES FOR A COHERENT PRICE FUNCTION

45

By denition p = Then M At +dt = M At + m PAt +dt = PBt +dt

dm m ( n ) = m (0) + dn

n 0

p (s) ds m (n) : known

M At +dt + MBt +dt M + m + MBt T Tt + m = At = t+dt = NAt +dt NA t + n NAt +dt NA t + n M At +dt + MBt +dt M At + m + MBt Tt + m Tt+dt = = = = NBt +dt NBt + n NBt +dt NBt + n

Thus p At +dt , p Bt +dt = f ( M At , MBt , NAt , NBt , m, n) = known Similar formulas apply when buying (selling) n shares of B. Case 2: buy/sell m money on n shares of A (where n unknown) In this case known parameters are M At +dt = M At + m, MBt +dt = MBt , NBt +dt = NBt , while unknown parameters that will be computed below are NAt +dt , PAt +dt , PBt +dt , p At +dt , p Bt +dt . n is found by solving the fundamental equation of p= for n, thus n = f ( M At +dt , MBt +dt , NAt , NBt +dt , m) = f ( M At , MBt , NAt , NBt , m) : known Then (same as case 1) NAt +dt = NAt + n PAt +dt = PBt +dt M At +dt + MBt +dt M + m + MBt T Tt + m = At = t+dt = NAt +dt NA t + n NAt +dt NA t + n M + MBt +dt M + m + MBt T Tt + m = At +dt = At = t+dt = NBt +dt NBt + n NBt +dt NBt + n dm m ( n ) = m (0) + dn
n

p (s) ds
0

Thus p At +dt , p Bt +dt = f ( M At , MBt , NAt , NBt , m, n) = known Similar formulas apply when buying (selling) m money on n shares of B.

46

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Properties

The properties that a coherent price function shall satisfy are proposed to be

the following: Property 3.4.1.1. Reexiveness: p At ( NAt , NBt ) = 1 p At ( NBt , NAt ) Property 3.4.1.2. Summation to 1: p At + p Bt = 1 t Corollary 1. p At ( NAt , NAt ) = 1/2 NAt 0 Property 3.4.1.3. Differentiability: p At ( NAt , NBt ) differentiable NAt , NBt 0 (or smooth) Property 3.4.1.4. Injection: p At ( NAt , NBt ) injective NAt , NBt 0 (+& invert injective), namely p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ) NA1 = NA2 p At ( NA0 , NB1 ) = p At ( NA0 , NB2 ) NB1 = NB2 Property 3.4.1.5. Monotonicity (& invert monotonicity): p At ( NAt , NBt ) is monotonically (& invert monotonically) increasing function at NAt , namely NA1 < NA2 p At ( NA1 , NB0 ) < p At ( NA2 , NB0 ) and also monotonically (& invert monotonically) decreasing at NBt , namely NB1 < NB2 p At ( NA0 , NB1 ) > p At ( NA0 , NB2 ) Property 3.4.1.6. Convergence to 1: NA t p At ( NAt , NBt ) 1 NBt Property 3.4.1.7. Convergence to 0: NA t 0 p At ( NAt , NBt ) 0 NBt

3.4. PROPERTIES FOR A COHERENT PRICE FUNCTION

47

3.4.2

Extension for n = k outcomes

Notation The following notation in the case of n = 2 outcomes will be consistently used in this and the following Sections of this Chapter. outcomes: O1 , O2 , . . . , Oi , . . . , Ok (mutually exclusive and exhaustive) price: p1 , p2 , . . . , pi , . . . , pk (instantaneous, per share) payoff: P1 , P2 , . . . , Pi , . . . , Pk (per share) money: M1 , M2 , . . . , Mi , . . . , Mk (total amount wagered) shares: N1 , N2 , . . . , Ni , . . . , Nk (total number purchased) At time t In the following, i I , where { I } = {i : i = 1, . . . , k } for some positive integer

k. At time t known are the number of shares and money invested per outcome, namely Mit , Nit , where Mit , Nit > 0, Based on these, the expected payoff (the payoff that would be distributed if the market was to be closed at time t) per share is given by the formula: Pit = 1 Nit Mit + M jt
j =i

Tt Nit

Given the known variables, price is a function of pit = f ( Mit , Nit ) =? At time t + dt Case 1: buy/sell n shares of Oi by paying m money (m: unknown) In this case known parameters are M jt +dt = M jt , Nit +dt = Nit + n, Njt +dt = Njt , while unknown parameters that will be computed below are Mit +dt , Pit +dt , Pjt +dt , pit +dt , p jt +dt , where i [1, k ] N, j [1, k ] N, j = i. dm By denition p = m ( n ) = m (0) + dn
n 0

p (s) ds m (n) : known

48

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Then Mit +dt = Mit + m Pit +dt = Pjt +dt = 1 Nit +dt

Mit +dt + M jt
j =i

= =

1 Nit + n 1 Njt

Mit + m + M jt
j =i

Tt+dt Nit +dt

1 Njt +dt

Mit +dt + M jt
j =i

Mit + m + M jt
j =i

Tt+dt Njt +dt

Thus p At +dt , p Bt +dt = f ( M At , MBt , NAt , NBt , m, n) = known j [1, k ] N Case 2: buy/sell m money on n shares of Oi (n: unknown) In this case known param-

eters are Mit +dt = Mit + m, M jt +dt = M jt , Njt +dt = Njt , while unknown parameters that will be computed below are Nit +dt , Pit +dt , Pjt +dt , pit +dt , p jt +dt , where i [1, k] N, j

[1, k] N, j = i.
n is found by solving the fundamental equation of p= for n, thus n = f Mit +dt , M jt +dt , Nit , Njt +dt , m = f Mit , M jt , Nit , Njt , m : known Then (same as case 1) Nit +dt = Nit + n Pit +dt = Pjt +dt Mit +dt + M jt +dt Mit + m + M jt T Tt + m = = t+dt = Nit +dt Nit + n Nit +dt Nit + n Mit +dt + M jt +dt Mit + m + M jt T Tt + m = = = t+dt = Njt +dt Njt + n Njt +dt Njt + n dm m ( n ) = m (0) + dn
n

p (s) ds
0

Thus pit +dt , p jt +dt = f Mit , M jt , Nit , Njt , m, n = known j [1, k ] N

3.4. PROPERTIES FOR A COHERENT PRICE FUNCTION

49

Properties

The properties that a coherent price function shall satisfy in the case of n = k

outcomes are proposed to be the following: Property 3.4.2.1. Reexiveness: This property does not naturally extend for n > 2 outcomes. Property 3.4.2.2. Summation to 1:

i =1

pi

= 1 t

Corollary 2. if N1t , . . . , Nkt = c, then pit ( N1t , . . . , Nkt ) = 1 , c 0 k

Property 3.4.2.3. Differentiability: pit ( N1t , . . . , Nkt ) differentiable i { I }, where { I } = {i : i = 1, . . . , k } for k positive integer (or smooth) Property 3.4.2.4. Injection: pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Ni1 = Ni2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit Nl0 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Nj1 = Nj2

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer


Property 3.4.2.5. Monotonicity (& invert monotonicity): Ni1 < Ni2 pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Nj1 < Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 < pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer


Property 3.4.2.6. Convergence to 1: if Nit j [1, k] N, j = i, then pit 1 Njt

Property 3.4.2.7. Convergence to 0: if Nit 0 j [1, k] N, j = i, then pit 0 Njt

It is here to be noted that the functions proposed by Pennock [2004] do not satisfy a number of the above properties, thus this framework of coherency.

50

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

3.5 A Coherent Price Function I


3.5.1 Simple case: n = 2 outcomes
The rst coherent price function introduced is, for the simple case of n = 2

At time t outcomes

p At = At time t + dt

N 1 1 exp At 2 NBt

+ exp

NBt NA t

(3.1)

Case 1: buy/sell n shares of A by paying m money (m: unknown) As deployed in Section 3.4.1, known parameters are MBt +dt = MBt , NAt +dt = NAt + n, NBt +dt = NBt , while unknown parameters that will be computed below are M At +dt , PAt +dt , PBt +dt , p At +dt , p Bt +dt . By denition p= dm m ( n ) = m (0) + dn
n 0 n 0

p (s) ds 1 exp NA t + s NBt

m=
This integral equals to m= N +n 1 n + NBt exp At 2 NBt

+ exp

NBt NA t + s

ds

+ ( NAt + n) exp

NBt NA t + n

+ NBt Ei

NBt NA t + n

where Ei denotes the exponential integral function, which is dened as Ei ( x ) = The integral can be easily computed numerically. Then M At +dt = M At + m PAt +dt = PBt +dt M At +dt + MBt +dt M + m + MBt T Tt + m = At = t+dt = NAt +dt NA t + n NAt +dt NA t + n M + MBt +dt M + m + MBt T Tt + m = t+dt = = At +dt = At NBt +dt NBt + n NBt +dt NBt + n

exp (t) dt. t

3.5. A COHERENT PRICE FUNCTION I

51

Thus p At +dt = 1 1 exp 2 1 = 2 1 = 1 exp 2 1 = 2

NAt +dt NBt +dt

+ exp

NBt +dt NAt +dt

p Bt +dt

NA t + n NBt + exp NBt NA t + n NA t NBt + exp = NBt NA t NBt N +n 1 exp + exp At NA t + n NBt 1 exp

Similar formulas apply when buying (selling) n shares of B. Case 2: buy/sell m money on n shares of A (n: unknown) As deployed in Section 3.4.1, known parameters are M At +dt = M At + m, MBt +dt = MBt , NBt +dt = NBt , while unknown that will be computed below are NAt +dt , PAt +dt , PBt +dt , p At +dt , p Bt +dt . n is found by numerically solving the equation
n

m=
0

1 exp

NA t + s NBt

+ exp

NBt NA t + s

ds

for n (no analytical solution exists). Then (same as case 1) NAt +dt = NAt + n PAt +dt = PBt +dt M At +dt + MBt +dt M + m + MBt Tt + m T = At = t+dt = NAt +dt NA t + n NAt +dt NA t + n M + MBt +dt M + m + MBt T Tt + m = At +dt = At = t+dt = NBt +dt NBt + n NBt +dt NBt + n

52

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Thus p At +dt = 1 1 exp 2 1 = 2 1 = 1 exp 2 1 = 2

NAt +dt NBt +dt

+ exp

NBt +dt NAt +dt

p Bt +dt

NA t + n NBt + exp NBt NA t + n NA t NBt + exp = NBt NA t NBt N +n 1 exp + exp At NA t + n NBt 1 exp

Similar formulas apply when buying (selling) m money on n shares of B. Properties Property 3.5.1.1. Reexiveness: p At ( NAt , NBt ) = 1 p At ( NBt , NAt ) Proof. p At ( NAt , NBt ) + p At ( NBt , NAt ) = N 1 1 exp At 2 NBt

+ exp

NBt NA t

1 N 1 exp Bt 2 NA t

+ exp

NA t NBt

=1

Property 3.5.1.2. Summation to 1: p At + p Bt = 1t Proof. p At ( NAt , NBt ) + p At ( NBt , NAt ) = 1 N 1 exp At 2 NBt

+ exp

NBt NA t

1 N 1 exp Bt 2 NA t

+ exp

NA t NBt

=1

Corollary: p At ( NAt , NAt ) = 1/2 NAt 0 (applies for p At = p Bt ) Property 3.5.1.3. Differentiability: p At ( NAt , NBt ) differentiable NAt , NBt 0 (or smooth) Proof. p At ( NAt , NBt ) 1 = NA t 2 NBt N exp Bt 2 NA t NA t

1 N exp At NBt NBt

3.5. A COHERENT PRICE FUNCTION I

53

which is a continuous function. p At ( NAt , NBt ) 1 = NBt 2 N NA t exp At 2 NBt NBt

1 N exp Bt NA t NA t

which is also a continuous function. Therefore, p At ( NAt , NBt ) differentiable NAt , NBt > 0. Property 3.5.1.4. Injection: p At ( NAt , NBt ) injective (& invert injective) NAt , NBt 0, therefore p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ) NA1 = NA2 p At ( NA0 , NB1 ) = p At ( NA0 , NB2 ) NB1 = NB2 Proof. Direct. p At ( NAt , NBt ) is differentiable in NAt (property 3.4.1.3). Also, 1 p At ( NAt , NBt ) = NA t 2 NBt N exp Bt 2 NA t NA t

N 1 exp At NBt NBt

> 0 NAt , NBt > 0

(derivative positive in all R + ). Thus, p At ( NAt , NBt ) injective (one-to-one) to NAt , namely p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ) NA1 = NA2 . p At ( NAt , NBt ) is differentiable in NBt (property 3.4.1.3). Also, p At ( NAt , NBt ) 1 = NBt 2 NA t N exp At 2 NBt NBt

1 N exp Bt NA t NA t

< 0 NAt , NBt > 0

(derivative negative in all R + ). Thus, p At ( NAt , NBt ) injective (one-to-one) to NBt , namely p At ( NA0 , NB1 ) = p At ( NA0 , NB2 ) NB1 = NB2 . Invert.

NA1 = NA2

exp

NA1 NB0

= exp

NA2 NB0

& exp

NB0 NA1

= exp

NB0 NA2

p At ( NA1 , NB0 ) = p At ( NA2 , NB0 )

and similarly NA1 = NA2 p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ).

54

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Property 3.5.1.5. Monotonicity (& invert monotonicity):p At ( NAt , NBt ) is monotonically increasing function at NAt , namely NA1 < NA2 p At ( NA1 , NB0 ) < p At ( NA2 , NB0 ) and also monotonically decreasing at NBt , namely NB1 < NB2 p At ( NA0 , NB1 ) > p At ( NA0 , NB2 ) Proof. Direct. p At ( NAt , NBt ) is differentiable in NAt (property 3.4.1.3). Also, p At ( NAt , NBt ) 1 = NA t 2 NBt N exp Bt 2 N NA At t

1 N exp At NBt NBt

> 0 NAt , NBt > 0

(derivative positive in all R + ). Thus, p At ( NAt , NBt ) is monotonically increasing to NAt , namely p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ) NA1 = NA2 . p At ( NAt , NBt ) is differentiable in NBt (property 3.4.1.3). Also, 1 p At ( NAt , NBt ) = NBt 2 NA t N exp At 2 NBt NBt

1 N exp Bt NA t NA t

< 0 NAt , NBt > 0

(derivative negative in all R + ). Thus, p At ( NAt , NBt ) is monotonically decreasing to NBt , namely p At ( NA0 , NB1 ) = p At ( NA0 , NB2 ) NB1 = NB2 . Invert. Proof was not possible. Property 3.5.1.6. Convergence to 1: NA t p At ( NAt , NBt ) 1 NBt Proof. NA t NBt exp NA t NBt

0 & exp

NBt NA t

p At ( NAt , NBt ) 1

Property 3.5.1.7. Convergence to 0:

3.5. A COHERENT PRICE FUNCTION I

55

NA t 0 p At ( NAt , NBt ) 0 NBt Proof. NA t 0 NBt exp NA t NBt

1 & exp

NBt NA t

p At ( NAt , NBt ) 0

3.5.2

Extension for n = k outcomes


The extension of the rst coherent price function introduced for n = k outcomes

At time t

is suggested to be the following. pit = 2 k2 N k k exp it 2 2 ( k 1) N jt j =i

Nj k exp t 2 ( k 1) j =i Nit

(3.2)

At time t + dt Case 1: buy/sell n shares of Oi by paying m money (m: unknown) As deployed in Section 3.4.2, in this case known parameters are M jt +dt = M jt , Nit +dt = Nit + n, Njt +dt = Njt , while unknown parameters that will be computed below are Mit +dt , Pit +dt , Pjt +dt , pit +dt , p jt +dt , where i [1, k ] N, j [1, k ] N, j = i. By denition p= dm m ( n ) = m (0) + dn 2 k2
n 0 n 0

p (s) ds

m=

Ni + s k k exp t 2 2 ( k 1) j =i Njt

Njt k exp 2 ( k 1) j =i Nit + s

ds

No analytical solution of the integral is known, to the author at least. However, a numerical solution can be easily computed.

56

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Then Mit +dt = Mit + m Pit +dt = Pjt +dt = 1 Nit +dt

Mit +dt + M jt
j =i

= =

1 Nit + n 1 Njt

Mit + m + M jt
j =i

Tt+dt Nit +dt

1 Njt +dt

Mit +dt + M jt
j =i

Mit + m + M jt
j =i

Tt+dt Njt +dt

where j [1, k ] N, j = i

Thus pit +dt = 2 k2 N k k exp it +dt 2 2 ( k 1) j =i Njt +dt 2 k2

Nj +dt k exp t 2 ( k 1) j =i Nit +dt

pit +dt = pit +dt =


p jt +dt =

N +n k k exp it 2 2 ( k 1) j =i Njt

Njt k exp 2 ( k 1) j =i Nit + n Njt Nit + n Nmt +dt Njt +dt

Ni + n 1 1 exp t k k ( k 1) Njt j =i exp Nit +dt Njt +dt

exp +

k 2 k 2 2 2 k 2 k 2 ( k 1) exp Njt +dt Nit +dt

j = i ,m

exp

k 2 k 2 2 ( k 1)

j = i ,m

exp

Nmt +dt Nkt +dt

p jt +dt =

1 1 k k ( k 1) Nit + n Njt

exp

exp

Njt Nit + n

j = i ,m

exp

Nmt Njt

exp

Njt Nmt

where m [1, k ] N, m = i, j

Case 2: buy/sell m money on n shares of Oi (n: unknown)

As deployed in Section 3.4.2,

in this case known parameters are Mit +dt = Mit + m, M jt +dt = M jt , Njt +dt = Njt , while unknown parameters that will be computed below are Nit +dt , Pit +dt , Pjt +dt , pit +dt , p jt +dt , where i [1, k ] N, j [1, k ] N, j = i. n is found by numerically solving the following equation for n:

3.5. A COHERENT PRICE FUNCTION I

57

m=

2 k2

n 0

k k 2 2 ( k 1)

exp
j =i

Nit + s Njt

Njt k exp 2 ( k 1) j =i Nit + s

ds

Then, same as case 1 Nit +dt = Nit + n Pit +dt = Pjt +dt = 1 Nit +dt

Mit +dt + M jt
j =i

= =

1 Nit + n 1 Njt

Mit + m + M jt
j =i

Tt+dt Nit +dt

1 Njt +dt

Mit +dt + M jt
j =i

Mit + m + M jt
j =i

Tt+dt Njt +dt

where j [1, k ] N, j = i

Thus pit +dt = p jt +dt = N +n 1 1 exp it k k ( k 1) j =i Njt 1 1 k k ( k 1) Nit + n Njt

exp

Njt Nit + n

exp

exp

Njt Nit + n

j = i ,m

exp

Nmt Njt

exp

Njt Nmt

where m [1, k ] N, m = i, j

Properties Property 3.5.2.1. Reexiveness: This property does not apply for n > 2. Property 3.5.2.2. Summation to 1:

i =1

pi

= 1 t

58

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Proof.

i =1

pit =
k

i =1 k

2 k k2 2 1

i =1 k

N k 2 exp it 2 k 2 ( k 1) j =i Njt

exp
j =i

Njt Nit

i =1

k k (k 1) exp
i =1 i =1 j = i

Nit Njt

exp
i =1 j = i

Njt Nit

= 1 t

Corollary: if Nic = c i { I }, where { I } = {i : i = 1, . . . , k } for k positive integer, then pit (c, . . . , c) = 1/k c 0 (applies for pit = ct i ). Property 3.5.2.3. Differentiability: pit ( N1t , . . . , Nkt ) differentiable i { I }, where { I } = {i : i = 1, . . . , k } for k positive integer (smooth) Proof. pit ( N1t , . . . , Nkt ) 1 = Nit k ( k 1)

j =i

Njt Ni2 t

exp

Njt Nit

+
j =i

N 1 exp it Njt Njt

which is a continuous function. Thus, pit ( N1t , . . . , Nkt ) differentiable for i { I }, where

{ I } = {i : i = 1, . . . , k} for k positive integer.

pit ( N1t , . . . , Nkt ) 1 = Njt k ( k 1)

j =i

Nit N exp it 2 N Njt jt

+
j =i

Nj 1 exp t Nit Nit

which is a continuous function. Thus, pit ( N1t , . . . , Nkt ) differentiable for j { I }, where

{ I } = { j : j = 1, . . . , k, j = i } for k positive integer.


Property 3.5.2.4. Injection: pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Ni1 = Ni2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit Nl0 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Nj1 = Nj2

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer


Proof. Direct. pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) differentiable in Nit (property 3.4.2.3). Also, pit ( N10 , . . . , Nk0 ) > 0 N10 , . . . , Nk0 > 0 Nit

3.5. A COHERENT PRICE FUNCTION I

59

(derivative positive in all R + ). Thus, pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) is injective (one-to-one) in Nit , namely pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Ni1 = Ni2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 differentiable in Njt (property 3.4.2.3). Also, pit ( N10 , . . . , Nk0 ) < 0 N10 , . . . , Nk0 > 0 Njt (derivative negative in all R + ). Thus, pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) is injective (one-to-one) in Njt , namely pit Nl0 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit Nl0 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Nj1 = Nj2 Invert. Ni1 = Ni2 exp Ni1 Nj0

= exp
Ni2 Nj0

Ni2 Nj0 &

& exp

Nj0 Ni1

= exp

Nj0 Ni2 Nj0 Ni2

exp
j =i

Ni1 Nj0

= exp
j =i

exp
j =i

Nj0 Ni1

= exp
j =i

pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 )

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer.


Similarly: Nj1 = Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0

Property 3.5.2.5. Monotonicity (& invert monotonicity): Ni1 < Ni2 pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Nj1 < Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 < pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer

60

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Proof. Direct. pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) is differentiable in Nit (property 3.4.2.3). Also, pit ( N1t , . . . , Nkt ) > 0 N1t , . . . , Nkt > 0 Nit (derivative positive in all R + ). Thus, pit ( N1t , . . . , Nk t ) monotonically increasing to Nit , namely Ni1 < Ni2 pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Similarly, pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 Also, pit ( N1t , . . . , Nkt ) < 0 N1t , . . . , Nkt > 0 Njt (derivative negative in all R + ). Thus, pit ( N1t , . . . , Nk t ) monotonically decreasing to Njt , namely Nj1 < Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 > pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Invert. Proof was not possible. Property 3.5.2.6. Convergence to 1: if Proof. Nit Njt exp Nit Njt Nit j [1, k] N, j = i, then pit 1 Njt differentiable in Njt (property 3.4.2.3).

0 & exp

Njt Nit

exp
j =i

Nit Njt

0&

exp
j =i

Njt Nit

k1

k k 2 k 0+ ( k 1) k2 2 2 ( k 1) 2 ( k 1) 2 pit ( N1t , . . . , Nkt ) = 1 k

pit ( N1t , . . . , Nkt )

Thus, the proposed function does not satisfy this property for n > 2.

3.6. A COHERENT PRICE FUNCTION II

61

Property 3.5.2.7. Convergence to 0: if Proof. Nit 0 Njt exp Nit Njt Nit 0 j [1, k] N, j = i, then pit 0 Njt

1 & exp

Njt Nit Njt Nit

0 0

exp
j =i

Nit Njt

k1&
2 k2

exp
j =i

pit ( N1t , . . . , Nkt )

k k k ( k 1) + 0 2 2 ( k 1) 2 ( k 1)

pit ( N1t , . . . , Nkt ) 0

Finally, this rst coherent price function proposed satises the suggested set of properties on its entirety but property 3.4.2.6 for n > 2.

3.6 A Coherent Price Function II


3.6.1 Simple case: n = 2 outcomes
The second coherent price function proposed is the conditional logit model, as

At time t

this was deployed in McFadden [2000]s Nobel Prize Lecture. The function resembles Hanson [2003a]s logarithmic scoring rule. Thus, in a sense, this price function shall be considered to serve as a unication of the core two prediction market mechanisms, DPM & MSR, and arguably stands as the core contribution of this Thesis. The formula in the simple n = 2 outcomes case is introduced below. exp p At = exp
NA t b NA t b

+ exp

NBt b

(3.3)

where, by denition, NAt , NBt > 0 t and b is a parameter that shapes markets sensitivity.

62

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

At time t + dt Case 1: buy/sell n shares of A by paying m money (m: unknown) As deployed in Section 3.4.1, known parameters are MBt +dt = MBt , NAt +dt = NAt + n, NBt +dt = NBt , while unknown parameters that will be computed below are M At +dt , PAt +dt , PBt +dt , p At +dt , p Bt +dt . By denition p= dm m ( n ) = m (0) + dn
n 0 n 0

p (s) ds ds
NBt b NBt b

m=

exp exp
NA t + s b

NA t + s b

+ exp

NBt b

m = b log

exp exp

NA t + n b NA t b

+ exp + exp

Then M At +dt = M At + m PAt +dt = PBt +dt M At +dt + MBt +dt M + m + MBt Tt + m T = At = t+dt = NAt +dt NA t + n NAt +dt NA t + n M + MBt +dt M + m + MBt Tt + m T = At = t+dt = = At +dt NBt +dt NBt + n NBt +dt NBt

Thus exp p At +dt = exp


NAt +dt b NAt +dt b

exp
NBt +dt b

=
exp

NA t + n b

+ exp
NBt +dt b

NA t + n b

+ exp
NBt b

NBt b

exp p Bt +dt = exp


NAt +dt b

exp
NBt +dt b

=
exp

+ exp

NA t + n b

+ exp

NBt b

Similar when buying (selling) n shares of B.

3.6. A COHERENT PRICE FUNCTION II

63

Case 2: buy/sell m money on n shares of A (n: unknown)

As stated in 3.4.1, in this case

known parameters are M At +dt = M At + m, MBt +dt = MBt , NBt +dt = NBt , while unknown parameters that will be computed below are NAt +dt , PAt +dt , PBt +dt , p At +dt , p Bt +dt . n is found by solving the fundamental equation of m = b log for n NA t + m b NBt + m b NBt b exp exp
NA t + n b NA t b

+ exp + exp

NBt b NBt b

n = b log exp Then (same as case 1) NAt +dt = NAt + n PAt +dt = PBt +dt

+ exp

exp

NA t

M At +dt + MBt +dt M + m + MBt Tt + m T = At = t+dt = NAt +dt NA t + n NAt +dt NA t + n M + MBt +dt M + m + MBt Tt + m T = At = t+dt = = At +dt NBt +dt NBt + n NBt +dt NBt

Thus exp p At +dt = exp


NAt +dt b NAt +dt b

exp
NBt +dt b

=
exp

NA t + n b

+ exp
NBt +dt b

NA t + n b

+ exp
NBt b

NBt b

exp p Bt +dt = exp


NAt +dt b

exp
NBt +dt b

=
exp

+ exp

NA t + n b

+ exp

NBt b

Similar formulas apply when buying (selling) m money on n shares of B. Properties Property 3.6.1.1. Reexiveness: p At ( NAt , NBt ) = 1 p At ( NBt , NAt )

64

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Proof. p At ( NAt , NBt ) + p At ( NBt , NAt ) = exp

=
exp

NA t b

exp
NBt b

NA t b

+
exp

NA t b

+ exp

NA t b

+ exp

NBt b

=1

Property 3.6.1.2. Summation to 1: p At + p Bt = 1t Proof. p At ( NAt , NBt ) + p At ( NBt , NAt ) = exp

=
exp

NA t b

exp
NBt b

NA t b

+
exp

NA t b

+ exp

NA t b

+ exp

NBt b

=1

Corollary: p At ( NAt , NAt ) = 1/2 NAt 0 (applies for p At = p Bt ) Property 3.6.1.3. Differentiability: p At ( NAt , NBt ) differentiable NAt , NBt 0 (smooth) Proof. p At ( NAt , NBt ) 1 = NA t b exp which is a continuous function. p At ( NAt , NBt ) 1 = NBt b exp exp
NA t b NAt + NBt b

exp
NA t b

NA t b

+ exp

NBt b

1 b exp

exp
NA t b

NA t b

2
NBt b

+ exp

+ exp

NBt b

which is also a continuous function. Thus, p At ( NAt , NBt ) differentiable NAt , NBt > 0. Property 3.6.1.4. Injection: p At ( NAt , NBt ) (& invert injective) injective NAt , NBt 0 thus p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ) NA1 = NA2 p At ( NA0 , NB1 ) = p At ( NA0 , NB2 ) NB1 = NB2

3.6. A COHERENT PRICE FUNCTION II

65

Proof. Direct. p At ( NAt , NBt ) differentiable in NAt (property 3.4.1.3). Also, p At ( NAt , NBt ) 1 = NA t b exp 1 = b exp exp
NA t b NA t b

exp
NA t b

NA t b NBt b

exp

+ exp

1 b exp
NA t b

exp
NA t b

NA t b

2
NBt b

+ exp

NBt b

+ exp

NBt b

exp

NA t b

+ exp

> 0 NAt , NBt > 0

(derivative positive in all R + ). Thus, p At ( NAt , NBt ) is injective (one-to-one) to NAt , namely p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ) NA1 = NA2 . p At ( NAt , NBt ) differentiable in NBt (property 3.4.1.3). Also, 1 p At ( NAt , NBt ) = NBt b exp exp
NA t b NAt + NBt b

+ exp

NBt b

< 0 NAt , NBt > 0

(derivative negative in all R + ). Therefore, p At ( NAt , NBt ) is injective (one-to-one) to NBt , namely p At ( NA0 , NB1 ) = p At ( NA0 , NB2 ) NB1 = NB2 . Invert.

NA1 = NA2

exp

NA1 b exp
NA1 b

= exp
NA1 b

NA2 b

& exp exp

NA1 b
NA2 b

+ exp

NB0 b

= exp

NA2 b

+ exp

NB0 b

exp

+ exp

NB0 b

=
exp

NA2 b

+ exp

NB0 b

p At ( NA1 , NB0 ) = p At ( NA2 , NB0 )

and similarly NA1 = NA2 p At ( NA1 , NB0 ) = p At ( NA2 , NB0 ). Property 3.6.1.5. Monotonicity (& invert monotonicity): p At ( NAt , NBt ) is monotonically increasing function at NAt , namely NA1 < NA2 p At ( NA1 , NB0 ) < p At ( NA2 , NB0 )

66

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

and also monotonically decreasing at NBt , namely NB1 < NB2 p At ( NA0 , NB1 ) > p At ( NA0 , NB2 ) Proof. Direct. p At ( NAt , NBt ) differentiable in NAt (property 3.4.1.3). Also, NA t NA t exp exp b b 1 p At ( NAt , NBt ) = 1 NA t NA t b exp NAt + exp NBt exp b + exp b b

NBt b

> 0 NAt , NBt > 0

(derivative positive in all R + ). Thus, p At ( NAt , NBt ) is monotonically increasing to NAt , namely NA1 < NA2 p At ( NA1 , NB0 ) < p At ( NA2 , NB0 ). p At ( NAt , NBt ) is differentiable in NBt (property 3.4.1.3). Also, 1 p At ( NAt , NBt ) = NBt b exp exp
NA t b NAt + NBt b

+ exp

NBt b

< 0 NAt , NBt > 0

(derivative negative in all R + ). Thus, p At ( NAt , NBt ) is monotonically decreasing to NBt , namely NB1 < NB2 p At ( NA0 , NB1 ) > p At ( NA0 , NB2 ). Invert. p At ( NA1 , NB0 ) < p At ( NA2 , NB0 )

exp

exp
NA1 b

NA1 b

exp
NB0 b

<
exp

NA2 b

+ exp exp < exp

NA2 b

+ exp < exp

NB0 b

exp
NA1 b

exp exp

NA1 b NA1 b

NA2 b NA2 b

+ exp

NB0 b

NA2 b

+ exp

NB0 b

NA1 < NA2

namely p At ( NA1 , NB0 ) < p At ( NA2 , NB0 ) NA1 < NA2 Also, p At ( NA0 , NB1 ) < p At ( NA0 , NB2 )

exp

exp
NA0 b

NA0 b

exp
NB1 b

<
exp

NA0 b

+ exp < exp

NA0 b

+ exp

NB2 b

exp

NB2 b

NB1 b

NB2 < NB1

3.6. A COHERENT PRICE FUNCTION II

67

namely p At ( NA0 , NB1 ) < p At ( NA0 , NB2 ) NB1 > NB2 Property 3.6.1.6. Convergence to 1: NA t p At ( NAt , NBt ) 1 NBt Proof. NA t NA t NBt NA t + exp b NBt exp NBt b NA t b NBt exp b exp NBt b

p At ( NAt , NBt ) 1

Property 3.6.1.7. Convergence to 0: NA t 0 p At ( NAt , NBt ) 0 NBt Proof. NA t NA t 0 NA t NBt exp NBt b NBt NA t + exp exp exp b b exp NBt b NBt b exp & exp

NA t b NBt b

p At ( NAt , NBt ) 0

3.6.2

Extension for k outcomes


The extension of the second coherent price function introduced for n = k

At time t

outcomes is suggested to be the following. exp pit ( N1t , . . . , Nit , . . . , Nkt ) =


Nit b Njt b

k j=1 exp

(3.4)

68

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

At time t + dt Case 1: buy/sell n shares of Oi by paying m money (m: unknown) As deployed in Section 3.4.2, in this case known parameters are M jt +dt = M jt , Nit +dt = Nit + n, Njt +dt = Njt , while unknown parameters that will be computed below are Mit +dt , Pit +dt , Pjt +dt , pit +dt , p jt +dt , where i [1, k ] N, j [1, k ] N, j = i. By denition p= dm m ( n ) = m (0) + dn
n 0 n

p (s) ds ds
Njt b Njt b

m=

exp exp
Nit +s b

0 Nit +s b Njt b

+ exp j =i
Nit +n b Nit b

m = b log

exp exp

+ j =i exp + j =i exp

Then Mit +dt = Mit + m Pit +dt = Pjt +dt = 1 Nit +dt 1 Njt +dt Mit +dt + M jt
j =i

= =

1 Nit + n 1 Njt

Mit + m + M jt
j =i

Tt + m Tt+dt = Nit +dt Nit + n

Mit +dt + M jt
j =i

Mit + m + M jt
j =i

Tt + m Tt+dt = Njt +dt Njt

where j [1, k ] N, j = i

Thus exp pit +dt ( N1t +dt , . . . , Nit +dt , . . . , Nkt +dt ) = exp
Nit +n b Nit +dt b Njt +dt b

exp

Nit +dt b

+ j =i exp

=
exp

Nit +n b

+ j =i exp

Njt b

3.6. A COHERENT PRICE FUNCTION II

69

p jt +dt N1t +dt , . . . , Njt +dt , . . . , Nit +dt , . . . , Nkt +dt = exp

=
exp

Njt +dt b

Njt +dt b

+ exp

Nit +dt b Njt b

+ m =i, j exp

Nmt +dt b

exp

=
exp

Njt b

+ exp

Nit +n b

+ m =i, j exp

Nmt b

where m [1, k ] N, m = i, j

Case 2: buy/sell m money on n shares of Oi (n: unknown)

As deployed in Section 3.4.2,

in this case known parameters are Mit +dt = Mit + m, M jt +dt = M jt , Njt +dt = Njt , while unknown parameters that will be computed below are Nit +dt , Pit +dt , Pjt +dt , pit +dt , p jt +dt , where i [1, k ] N, j [1, k ] N, j = i. n is found by solving m = b log for n: n = b log exp Then, same as case 1 Nit +dt = Nit + n Pit +dt = Pjt +dt = 1 Nit +dt Nit + m b exp exp
Nit +n b Nit b

+ j =i exp + j =i exp

Njt b Njt b

+ exp
j =i

Njt + m b

exp
j =i

Njt b

Nit

Mit +dt + M jt
j =i

= =

1 Nit + n 1 Njt

Mit + m + M jt
j =i

Tt+dt Tt + m = Nit +dt Nit + n

1 Njt +dt

Mit +dt + M jt
j =i

Mit + m + M jt
j =i

Tt+dt Tt + m = Njt +dt Njt

where j [1, k ] N, j = i

70

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

Thus exp pit +dt ( N1t +dt , . . . , Nit +dt , . . . , Nkt +dt ) = exp
Nit +n b Nit +dt b Njt +dt b

exp

Nit +dt b

+ j =i exp

=
exp

Nit +n b

+ j =i exp

Njt b

p jt +dt N1t +dt , . . . , Njt +dt , . . . , Nit +dt , . . . , Nkt +dt = exp

=
exp

Njt +dt b

Njt +dt b

+ exp

Nit +dt b Njt b

+ m =i, j exp

Nmt +dt b

exp

=
exp

Njt b

+ exp

Nit +n b

+ m =i, j exp

Nmt b

where m [1, k ] N, m = i, j

Properties Property 3.6.2.1. Reexiveness: This property does not apply for n > 2. Property 3.6.2.2. Summation to 1:

i =1

pi

= 1 t

Proof.

i =1

pi

i =1

exp

Nit b Njt b

k j=1 exp

= =

ik=1 exp k j=1 exp

Nit b Njt b

= 1 t

Corollary: if Nic = c i { I }, where { I } = {i : i = 1, . . . , k } for k positive integer, then pit ( c t , . . . , c t ) =


1 k

c 0 (applies for pit = ct i )

Property 3.6.2.3. Differentiability: pit ( N1t , . . . , Nkt ) differentiable i { I }, where { I } = {i : i = 1, . . . , k } for k positive integer (smooth)

3.6. A COHERENT PRICE FUNCTION II

71

Proof. exp bit pit ( N1t , . . . , Nkt ) 1 = Nit b b exp Njt j =1 b


N

2 Nit 1 exp b Njt b b j=1 exp b

which is a continuous function. Thus, pit ( N1t , . . . , Nkt ) differentiable for i { I }, where

{ I } = {i : i = 1, . . . , k} for k positive integer.


pit ( N1t , . . . , Nkt ) 1 = Njt b exp
Nit + Njt b Njt b 2

b j=1 exp

which is a continuous function. Therefore, pit ( N1t , . . . , Nkt ) differentiable for j { I }, where { I } = { j : j = 1, . . . , k, j = i } for k positive integer. Property 3.6.2.4. Injection: pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Ni1 = Ni2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit Nl0 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Nj1 = Nj2

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer


Proof. Direct. pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) is differentiable in Nit (property 3.4.2.3). Also, pit ( N1t , . . . , Nkt ) 1 1 = = Njt b Nit b b exp Njt b exp j =1 j =1 b b Nit Nit exp b exp b 1 1 > 0 N1t , . . . , Nk > 0 = t Njt b b b exp Njt exp
j =1 b j =1 b

exp

Nit b

exp

Nit b

(derivative positive in all R + ). Thus, pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) injective (one-to-one) in Nit , namely pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Ni1 = Ni2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 differentiable in Njt (property 3.4.2.3). Also, pit ( N1t , . . . , Nkt ) 1 = Njt b exp
Nit + Njt b Njt b 2

b j=1 exp

< 0 N1t , . . . , Nkt > 0

72

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

(derivative negative in all R + ). Thus, pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) injective (one-to-one) in Njt , namely pit Nl0 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit Nl0 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Nj1 = Nj2 Invert. Ni1 = Ni2

exp

Ni1 b

= exp
Ni1 b

Ni2 b

& exp

Ni1 b

+ exp
j =i

Nj0 b

= exp

Ni2 b

+ exp
j =i

Nj0 b

exp

exp
Ni1 Nb

exp
Nj0 b

+ j =i exp

=
exp

Ni2 b Nj0 b

Ni2 Nb

+ j =i exp

pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) = pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 )

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer.


Similarly: Nj1 = Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 = pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0

Property 3.6.2.5. Monotonicity (& invert monotonicity): Ni1 < Ni2 pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Nj1 < Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 < pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0

i, j { I }, where { I } = {i, j : i, j = 1, . . . , k, j = i } for k positive integer


Proof. Direct. pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) differentiable in Nit (property 3.4.2.3). Also, pit ( N1t , . . . , Nkt ) > 0 N1t , . . . , Nkt > 0 Nit (derivative positive in all R + ). Thus, pit ( N1t , . . . , Nk t ) monotonically increasing to Nit , namely Ni1 < Ni2 pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 )

3.6. A COHERENT PRICE FUNCTION II

73

Similarly, pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 Also,

differentiable in Njt (property 3.4.2.3).

pit ( N1t , . . . , Nkt ) < 0 N1t , . . . , Nkt > 0 Njt (derivative negative in all R + ). Thus, pit ( N1t , . . . , Nk t ) is monotonically decreasing to Njt , namely Nj1 < Nj2 pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 > pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Invert.

pit ( N10 , . . . , Ni10 , Ni1 , Ni+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 )

exp

exp
Ni1 Nb

Ni1 b Ni0 b

exp

+ j =i exp exp
Ni2 b Ni2 b

=
exp Ni1 b

Ni2 b Ni0 b

Ni2 Nb

+ j =i exp
j =i

exp
exp Ni1 b

Ni1 b

+ exp
Ni2 b

exp
Ni0 b

Ni0 b

<

exp
Ni1 b

+ exp
Ni2 b

exp
j =i

exp

< exp

Ni1 < Ni2

namely pit ( N10 , . . . , Ni10 , Ni1 , N+10 , . . . , Nk0 ) < pit ( N10 , . . . , Ni10 , Ni2 , Ni+10 , . . . , Nk0 ) Ni1 < Ni2 . Also, pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 < pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0

exp

exp
Nj1 Nb

Ni0 b Nk0 b

exp

+ k =i exp
exp Nj1 b

=
exp

Ni0 b Nk0 b

Nj2 Nb

+ k =i exp
Nj2 b

Nj2

<

exp

<

Nj1

74

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

namely pit N10 , . . . , Nj10 , Nj1 , Nj+10 , . . . , Nk0 < pit N10 , . . . , Nj10 , Nj2 , Nj+10 , . . . , Nk0 Nj2 < Nj1 . Property 3.6.2.6. Convergence to 1: if Proof. Nit j [1, k] N, j = i, then pit 1 Njt

j [1, k] N, j = i :
Thus: exp Nit b

Nit Nit Njt

Njt exp

Nit b

exp

Njt b

exp
j =i

Njt b

exp

Nit b

+ exp
j =i

Njt b

exp

Nit b

pit ( N1t , . . . , Nit , . . . , Nkt ) 1

Property 3.6.2.7. Convergence to 0: if Proof. Nit 0 j [1, k] N, j = i, then pit 0 Njt

j [1, k] N, j = i :
Thus: exp Nit b

Nit 0 Nit Njt

Njt exp

Nit b

exp

Njt b

exp
j =i

Njt b

exp

Nit b

exp

Nit b

+ exp
j =i

Njt b

exp

exp
Nit b

Nit b Njt b

+ j =i exp

0 pit ( N1t , . . . , Nit , . . . , Nkt ) 1

3.7. DISCUSSION

75

Finally, this second coherent price function proposed, namely the conditional logit model, satises the suggested set of properties on its entirety, thus it is recommended for wide usage.

3.7 Discussion
3.7.1 Summary

Serving at the core of this Thesis, this Chapter started with studying the mechanisms of well-established market institutions and the limits of their applicability under a prediction market setting. It then proceeded to review in detail Market Scoring Rules and Dynamic Pari-Mutuel, namely the couple of mechanisms that have been deployed with prediction markets in mind and now stand as the standard option when it comes to developing a market institution committed to forecasting at rst. By guring out their drawbacks, the Chapter next set forth towards a proposal for enhancements that address such shortcomings and attempt to advance the industry standard for a prediction market mechanism. Focusing on DPM, the essential part of markets price function was highlighted, also the limited sufciency of the available across the existing literature ones. In response to this, a framework of properties was developed for what was introduced to be a coherent price function, providing for a number of highly practical as well as aesthetic criteria that a DPM price function should meet to efciently deliver against a prediction markets purposes. On top of this framework, a couple of price functions were introduced and their compatibility to each one of the coherency properties above was studied in detail. Providing evidence on these functions adequacy compared to the previously available ones, it is hoped for them to stand as a contribution of value to the eld for researchers and practitioners alike. At the same time, the conditional logit model that was introduced to be the price function of choice, turns out to highly resemble the most popular MSR in use, eventually providing for a bridge or unication of the two standard prediction market mechanisms.

76

CHAPTER 3. THEORETICAL PROPERTIES OF PREDICTION MARKETS

3.7.2

Future work

It proved admittedly difcult to come up with such functions, yet -no matter the elegance of the results- the quest may not be over. The proposed coherence framework should not be cast in stone either. What this Chapter after all might play a part for is the introduction of some discipline when it comes to the study of a DPMs price function and the demonstration of such a disciplines benets. Furthermore, in another contribution that might be of value, the Chapter might serve as an indication on basic research still providing for a fertile ground in domains like a markets operating mechanism. There still remain a lot to be explored in better understanding core prediction market properties and return with responsive market design to limit each mechanisms drawbacks, while taking full advantage of its potential.

Chapter 4

Theoretical Evaluation of Prediction Markets


4.1 Overview
Chapter 3 essentially shaped a functioning prediction market mechanism, next to the theoretical framework of its properties. This contribution however highlights at the same time the need for a concrete model of market operation, taking into consideration all other primary components of operating a market institution, given the market mechanism. The design of such a model and the study of its fundamental properties consist the core of the hereby Chapter. Economic models typically fall within what might be called a fundamental equation of economics, according to Plott [1979]. The equation is comprised of four classes of parameters, including institutions, a commodity space, preferences and a feasible set of outcomes over that commodity space [Plott, 2001]. Given the proposed market mechanism as institution, under question for the model to deploy is the overriding principle of equilibrium, which completes the model and is to be studied among others in the following. The task of model design may also be described as invert, thus inherently relevant, to the one of algorithmic mechanism design. The related theory, which gained the 2007 Nobel Prize in Economics [Cihak, 2008], stands as a unique subeld of economics and game theory, 77

78

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

in having an engineering perspective [Nisan, 2007]. Just like operational researchers are interested in designing algorithms to implement good system-wide solutions to problems that involve multiple competitive subsystems, mechanism design considers how to arrive at solid system-wide solutions to problems involving multiple self-interested agents, each one with private information about their preferences. Therefore by in a sense inverting the process and scope of mechanism design theory, relevant techniques will be utilized to put focus on the study of the environmental parameters and agent preferences inuencing the operating characteristics of the given prediction market mechanism. After all, the results of mechanism design theory support Hayek [1945]s argument that markets efciently aggregate relevant private information [Nisan, 2007] and this essentially stands as the core question of this Chapter. The questions that will be ultimately addressed through the simplied market model to be deployed regard the existence and speed of markets convergence towards equilibrium, under conditions where information distributed among traders is partially revealed over time and incorporated into the market price. Moreover, focus is put on the coincidence of equilibriums price with the selection of an ideal trader who possesses all relevant information. In other words, the computational aspects that guarantee the convergence of a market-based decision support tool to the true rational expectations equilibrium are studied in detail. In that case of equilibrium, this translates to a meta-mechanism able to effectively aggregate the totality of available information throughout the number of contributing human agents. The Chapter proceeds as follows: Section 4.2 sets the generic problem, its background and approaches in use so far. Section 4.3 sketches the proposed model, in terms of its information structure, market mechanism and agent strategies, while Section 4.4 studies its properties, in terms of its convergence, speed to convergence and best possible prediction. Finally, Section 4.5 concludes and provides discussion on the set of outcomes, while it also suggests directions for future work.

4.2. BACKGROUND & RELATED WORK

79

4.2 Background & Related Work


Prediction markets initially occurred as a notional and experimental concept, according to what subsection 2.4.4 lays out. Theoretical examination naturally followed up and, as subsection 2.5.0.2 indicates, has lagged to their implementation and use, while yet remaining almost immature. That said, there exists a considerable pool of literature regarding topics inherently related to prediction markets modeling that provides a strong theoretical background for its study. What follows is a review of prior work on these topics, under the prism and fragmented extent of actually deploying such a model. Next to that, an analysis of existing empirical work in the area concludes the paragraph and nally sets the stage for the model deployment itself.

4.2.1

Theoretical work

Financial market institutions essentially bring buyers and sellers together to trade nancial products, namely securities, commodities, or other fungible items. By nature, such institutions incorporate and facilitate, in one way or another, the four functions of investment, hedging, speculation and information aggregation. Several institutions have been deployed during the years to efciently support these fundamental functions, mostly by focusing on just one of them. To name a few, stock markets are created and operating with the primary purpose of capital allocation, futures markets scope in hedging risks and betting markets serve as a physical destination for wagering. However, each one of these institutions facilitates also and in parallel the entirety of all core market functions. For example, stock markets serve as a hospitable environment for speculation as well, while hedging could be accomplished by holding a wide portfolio of the traded assets and stock prices are unbiased estimators of rm fundamentals [Mandelbrot, 1966]. In this context, there exists signicant evidence supporting the general presence of the fourth native function of markets operation, namely their aggregative and predictive nature. This presence is strong even in institutions that were designed with different objectives in mind, in both investing [Admati and Peiderer, 1987, Chen et al., 2007, Grossman and Stiglitz, 1980b, Hellwig, 1980, Holmstrom and Tirole, 1993, Lo, 1997], hedging [Jackwerth and Rubinstein, 1996b, Krueger and Kuttner, 1996, Roll, 1984] and

80

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

wagering cases [Boulier and Stekler, 2003, Debnath et al., 2003, Figlewski, 1979, Gandar et al., 1998, Schmidt and Werwatz, 2002, Thaler and Ziemba, 1988b, Winkler, 1971]. In other words, the information aggregation function is valid in every single market institution, a fact which gives further impetus to the study that follows. Feigenbaum et al. [2005] suggested that this function and process of information incorporation could be perceived, at its essence, as a distributed computation. In fact and practice, each trader begins with her own information, which then updates as she infers what information other traders may have, by observing market prices. The trader can then update her own beliefs based on the observations and, over time, market price and traders beliefs converge. In this ideal case, all private information becomes common knowledge, and thus any function of the private information can be directly evaluated by any agent or observer. This latter status of fully information disclosure in market prices, which results in no further trading even after observing the price itself, presents a cornerstone of the economics literature and consists the optimal status of each nancial market institution. Rational expectations equilibrium (REE) [Grossman, 1981, Lucas, 1972, McKelvey and Page, 1986], as it is called, and the theory of rational expectations are an extension to general equilibrium models that occurred from the need to understand markets where prices can lead to changes in the traders demand curves [Pennock and Sami, 2007]. This may be improbable for consumer products, in which each trader has a xed demand curve that describes the quantity demanded at each price. However, it is valid in the case of nancial securities markets, where securities have no direct consumption value, apart from an expectation of an increase in price for a protable future resell or cash out. In this context, the classical model of demand is invalidated and the essence of information aggregation is further highlighted, a process which ideally converges to a REE. The potential existence and stability of a REE has motivated a lot of research [Allen, 1981, 1982, Jordan and Radner, 1979, Radner, 1979], nally providing evidence that economies with asymmetric information do converge to a fully-revealing REE, under fairly general conditions on the value functions. Some criticism has also occurred and one has to note the ndings of Dubey et al. [1987b], who introduced two paradoxes that are implied from REE models. First, the simultaneous determination of equilibrium price and available information is nearly a problem of chicken-and-egg nature. If agents take into

4.2. BACKGROUND & RELATED WORK

81

account the equilibrium price in making decisions, then the way for these decisions to determine the prices remains unclear, as REE models do not formally consider the process of information ow into the market. Moreover, the nal communication of all information via market prices leaves few incentives for costly information gathering and this stands as another paradox of REE theory. The relevant no-trade theorems [Milgrom and Stokey, 1982], however, reveal that even tiny perturbations to the initial conditions invalidate them and, in practice, there exist several reasons that can lead an informed trader to expect prots from trading, such as the presence of irrational or towards hedging oriented traders. Nevertheless, these remarks need to be addressed when designing a market (or model) mechanism, namely taking the pricing formation process into consideration and providing sufcient incentives and conditions for information gathering [Pennock and Sami, 2007]. Another closely relevant notion to information aggregation is the one of common knowledge, namely information that all agents know, that all agents know that all agents know, and so on ad innitum [Fagin et al., 1996]. The process of people, initially disagreeing with each other, towards reaching an agreement could also be perceived as analogous to the process of information aggregation in nancial markets [Chen et al., 2006b]. Indeed, traders start with different information results in different opinions on the expected price of securities, but trading eventually leads to an agreement, expressed by the nal market price at equilibrium. Aumann [1976] was the rst to present the formal denition of common knowledge and proved that two people with common priors and posteriors to an event that are common knowledge, should also converge to common posteriors. Geanakoplos and Polemarchakis [1982] extended this outcome by lightening the second condition; common knowledge and equal posteriors occur when two people with same priors successively announce their posteriors to each other. McKelvey and Page [1986] generalized the previous results to n persons, only requiring the successive announcement of an aggregate statistic of individuals posteriors. When this statistic eventually becomes common knowledge, all posteriors of the n persons are equal. Finally, Nielsen et al. [1990] contributed the extension of the conditional probability (posterior) to the case of conditional expectation. Research on REE has also triggered the formulation of efcient market hypothesis, the strong form of which states that a securitys price fully reects the totality of relevant

82

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

information to the securitys value [Fama, 1970]. At its essence, the hypothesis validates the ubiquitous presence of the aggregative and predictive function of a markets operation. For a market to be efcient, information incorporation has to occur among traders who own a small amount of information about an event and that may serve as a theoretical foundation for the potential effectiveness of prediction markets. Finally, complete markets also stand as a concept of interest. The so-called, theoretically supreme, mechanism contains by denition enough linearly independent securities to span the entire state space of interest [Arrow, 1964, Varian, 1987]. In other words, the dimensionality (or independence) of the random variables, on which the security payoffs are based, is capable of capturing the totality of states over which information is to be aggregated. If a complete market is the case, all private information nally becomes common knowledge in equilibrium, and thus any function of the private information can be directly evaluated by any trader or external observer [Feigenbaum et al., 2005]. However, as expected, this ideal scenario is almost non achievable in practice, as it generally requires a number of securities exponential to the number of variables of interest. In the usual case of incomplete markets, aggregation may occur as partial and, in this context, the study of market complementarity and its nal status of information aggregation arises as of interest. This research question will be addressed in detail across this Chapter, in accordance with the dynamics of aggregation towards equilibrium, namely the bounds of convergence time in the cases that aggregation succeeds.

4.2.2

Empirical studies

While any theoretical work, directly relevant to prediction markets modeling and the research questions posed at the beginning of this Chapter, is limited to Feigenbaum et al. [2005], Chen et al. [2006b,a], empirical studies on the topic consist a more extended pool of knowledge. Laboratory experiments stand by default as an attractive alternative for the study of markets, as they provide a controllable and easily customizable environment for the understanding of market properties. Section 2.5.1 already pointed out a dozen of references on such applications, while what follows is a short review of the most relevant to information aggregation experimental attempts so far.

4.2. BACKGROUND & RELATED WORK

83

Plott and Sunder [1982, 1988], pioneers of the Experimental Economics eld, were the rst to attempt such an experiment with human agents and prove the capacity of markets to aggregate information in experimental settings. Moreover, they demonstrated the importance of selected market structure to achieve aggregation. Forrest and Lundholm [1990] contributed that experience of agents is necessary for information aggregation in the case of heterogeneous preferences. Lundholm [1991] presented evidence on the decline of aggregation efciency at the presence of aggregate uncertainty, while OBrien and Srivastava [1991] introduced the negative impact aroused in the case of increased securities complexity. More details are provided by the extensive review of Sunder [1995], covering the totality of experimental work on information aggregation by the date of release. Most recent works include the one of Noeth et al. [1999] that demonstrated the existence of information traps and Barner et al. [2004] provided evidence on securities integrating information that arrives over time. Finally, Hanson et al. [2006] and Rietz [2005] offered some insights on the effects of manipulation and arbitrage, rather encouraging for the stability of aggregation. This non-exhaustive review would be rather incomplete without a reference to the speech of Plott [2001], introducing the Nobel Prize in Economics 2001 and highlighting in parallel the sensitivity of aggregation to the available instruments in the market. However, even if the number of relevant laboratory studies is rather extended, one also has to mention the signicant number of real world cases that study in practice information aggregation and the questions posed previously. Section 2.5.1 provided a critical number of references of applicative nature, among them ones referring to Iowa Electronic Markets, other political markets, markets on sports events or various other applications. However, and given the theoretical nature of the hereby Chapter, a more analytical review of this part of the literature remains out of interest.

84

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

4.3 Design of Markets


4.3.1 Model of markets

Model design essentially balances on the trade-off between realistic detail and meaningful simplicity. In the case of prediction markets, a generic model should be comprised by at least three indispensable components, namely information structure, market mechanism and trader behavior, as shown in Figure 4.1.

Figure 4.1 Breakdown Structure of Prediction Market Modeling. source: Chen et al. [2006b]

Such a model was rstly introduced by Feigenbaum et al. [2005], while Chen et al. [2006b] moved forward to further add aggregate uncertainty and Chen et al. [2006a] studied the process to equilibrium at the presence of rational Bayesian traders, instead of the default truth telling ones. The notion of aggregate uncertainty, rather typical in real world settings, occurs when the state of the world is still not fully determined, even if information of all traders is pooled together. The modeling suggested here leverages on these previous ones to also capture the use of a pari-mutuel market setting, in compliance with the mechanism developed in Chapter 3. As a result, its information structure and market mechanism differ essentially, while the core of assumptions on trader behavior

4.3. DESIGN OF MARKETS

85

remain nearly identical to the previous studies.

4.3.2

Information structure

Information structure sketches the models frame of reference, by dening its inputs and environment. In the under study case, information structure refers to the state space of the world, the information available to traders regarding the true state of the world and the way that the latter information relates to the true state of the world. In mathematical terms, let S denote the state space of the world, where s = (s1 , s2 , . . . , sn ) S is a state vector of m dimensions. Also, let n be the number of traders in the market. A common approach in modeling information structure of markets is using prior probability distributions for the state of the world and the information that traders possess [Chen et al., 2006b]. Following this, a common prior probability distribution P (s) : S [0, 1] regarding the state of the world is introduced for each trader, while her information space is X . In the initial state, each agent i = 1, . . . , n is privy to one bit of information xi (input bit) regarding the state of the world, while x = ( x1 , x2 , . . . , xn ) X denotes the information vector for all traders. Probability distribution of x, conditional on state of the world s, is common knowledge to all traders, Q( x |s) : X S [0, 1], none of which, however, has any specic information about the actual value of bits held by others. This common-prior assumption is typical in the economics literature, but does not not imply that all agents agree. To the contrary, because each agent has different information, the initial state of the system is in general a state of disagreement [Feigenbaum et al., 2005]. As a further simplication, a boolean restriction on both the state and the information variables (si s and xi s) is introduced. The resulted state space of the world is S = {0, 1}m , while X = {0, 1}n is the information space. The prior probability distribution of the state of the world is P (s) : {0, 1}m [0, 1] and Q( x |s) : {0, 1}m {0, 1}n [0, 1] is the conditional distribution of information. To make this framework more clear, assume an example with an one dimensional state space (i.e, s = s1 ) and a boolean state of the world (either 0 or 1). Conditional on s = 1(s = 0), the probability to get x1 = 1( x1 = 0) is 0.9, and the probability to get

86

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

x1 = 1( x1 = 0) is 0.1. If the trader i gets xi = 1, although she does not know the value of s for certain, she knows that, with probability 0.9, s equals 1. This uncertainty in individual information introduces aggregate uncertainty to the model, resulting in an uncertain true state of the world, even with pooled information. In models of prediction markets without aggregate uncertainty, market traders hold accurate information about the state of the world s. For example, Feigenbaum et al. [2005]s model species that trader i knows si in an information market with n traders. Pooling information of all traders together makes the state of the world, s, uniquely determined. In other words, an omniscient trader would always be able to deterministically dene the true value of the security. Their model can be viewed as a special case of the model deployed here and rstly introduced by Chen et al. [2006b], by setting m = n and xi = si with probability 1.

4.3.3

Market mechanism

Market mechanism provides the structure on which the various agents and strategies to be selected meet the input of the already dened information structure. The denition of market mechanism also prescribes the securities being traded and trading rules of the market. In the proposed framework, the market modeling targets at predicting the value of a function f (s), which is determined by the true state of the world s. The actual value of s and, as a result, f (s) itself, is revealed on a future time. The form of f is common knowledge to all traders. In compliance with the market structure deployed in Chapter 3, a pari-mutuel market setting is introduced, which differentiates the proposed model to the ones of Feigenbaum et al. [2005], Chen et al. [2006b]. Following this setting, two securities k and l are traded in the market, one for each binary outcome, whose payoff is contingent on the actual value of f (s). If f (s) is nally 1, then all money are redistributed to the traders who own the k security, while, if f (s) is 0, owners of l share the money invested in the market. Following Dubey et al. [1987a], Jackson and Peck [1999], Feigenbaum et al. [2005], Chen et al. [2006b], we model the market mechanism as a multiperiod Shapley-Shubik market game [Shapley and Shubik, 1977] with restrictions. The enhancement of the pari-mutuel

4.3. DESIGN OF MARKETS

87

setting with the Shapley-Shubik model results in a process that operates as follows: The market proceeds in synchronous rounds. In each round, each agent i submits a selected security vi and a quantity qi , or alternatively a selected security vi and an amount of money bi . Selling is feasible by default only for securities and quantities that an agent already owns, in other words short selling is not allowed. Each trading round could be dened by a vector v = (v1 , v2 , . . . , vn ), where vi is the security which trader i selected for this round, and a vector q = (q1 , q2 , . . . , qn ) of the quantities qi submitted for trade by each agent (negative values refer to selling). An alternative setting would require instead of the vector q of quantities, a vector b =

(b1 , b2 , . . . , bn ) of the amount of money bi an agent i offered in this specic round.


Sticking on the rst alternate, a further simplifying restriction introduced is that q1 = 1 for each agent i on each round. In other words, each agent buys just one stock of the security that she selects on each round; no extensive buying nor selling are feasible. This modeling assumption serves two analytical purposes. First, it ensures that there is forced trade in every round. Second, forcing qi = 1 for all i means that the total volume of trade, therefore and, in a way, also the impact of any trader on the clearing price are common knowledge. Implementing the coherent price function II of Section 3.6 and selecting sensitivity parameter b equal to 1, we can compute the clearing price of both securities as a function of the agents selections v. For example, security k is priced as pk = exp(qk )/ (exp(qk ) + exp(ql )), where qk , ql are the total quantities of securities k, l already bought, while the submissions of all traders in each round of the Shapley-Shubik market game are considered happening simultaneously, as one transaction. After each round, the new prices pk , pl and only these (not the traders bids) are publicly revealed. Agents then revise their beliefs according to any information garnered from the new price. The market then enters a new round. The process continues until an equilibrium is reached, after which prices remain stable from round to round.

4.3.4

Agent strategies

In order to draw formal conclusions about the price evolution process, assumptions on the agents behavior have to be made. In general, agent type determines the preferences of

88

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

an agent over different outcomes of the market, while the complete decision rule which denes the action an agent will select in every distinguishable state of the world is called agents strategy. Agents utility determines its preferences over its own strategy and the strategy of other agents, given its type, which in turn determines its base preferences over different outcomes in the world. The basic model of agent rationality in game theory is that of an expected utility maximizer [Parkes, 2001]. Such an agent will select a strategy that maximizes its expected utility, given its preferences over outcomes, beliefs about the strategies of other agents and structure of the game. In the simplistic case of traders behavior modeling handled for the hereby study, specifying traders risk preference, rationality and trading strategy describes sufciently the needed agents properties. Essentially it is assumed that, in agreement with Feigenbaum et al. [2005], Chen et al. [2006b], agents are risk-neutral, myopic and bid truthfully, rather than behave strategically: each agent in each round bids her current valuation of the probability that the securitys event turns true. Expectations are computed according to each agents probability distribution, which is updated via Bayes rule when new information becomes available by revealing the clearing prices. It is also assumed that it is common knowledge that all the agents behave in the specied manner.

4.4 Convergence Properties


Based on the model of markets in Section 4.3, we examine several convergence properties of prediction markets to answer the four research questions raised in Section 4.1.

4.4.1

Price Convergence

Given the under study use of market mechanisms as predictors of future events, it is of signicant interest to study the properties and stability of markets equilibrium, at which the price remains stable if no new information arrives. The rst question to ask regards markets convergence to equilibrium. McKelvey and Page [1986] and Nielsen et al. [1990] proved that, if the initial information partition of traders is nite and traders rene it through an iterative process, in which a market statistic of their expectation is made public in each period, then the market

4.4. CONVERGENCE PROPERTIES

89

converges to equilibrium in nite rounds. Moreover, they introduced some conditions to be satised by the market statistic, that result in identical conditional expectations of the event by each trader at equilibrium. In other words McKelvey and Page [1986] and Nielsen et al. [1990] proved that people who disagree with each other eventually reach an agreement using an analogous to the market trading process, while the properties set are fully satised by our prescribed market model. These lead to Theorem 1. Theorem 1. (Chen et al. [2006b]) Without the arrival of new information, a prediction market converges to an equilibrium in nite steps. At equilibrium, all traders have the same expectation about the value of f (s), which equals the equilibrium market price. Following Chen et al. [2006b], the initial information structure of the prediction market is dened as follows:

(, F, ) (a probability space)

(4.1)

0 0 P0 = P1 , . . . , Pn (initial information partitions)

(4.2)

h : Rn R (an aggregate function)

(4.3)

Pi0 is a nite partition of the probability space of trader i, while the random variable that the market tries to predict is x. The market proceeds in rounds and, on round t, for each individual i and any state , bit ( ) = E A| Pit ( )

(4.4)

is dened to be individual is expectation of the random variable A based on his current information partition and

t t bt ( ) = b1 ( ) , . . . , bn ( )

(4.5)

is the expectation vector for all agents. The dened information structure enables Theorems 2 and 3.

90

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

Theorem 2. (McKelvey and Page [1986], Nielsen et al. [1990]) Assume an initial information structure as in 4.1, 4.2 and 4.3. Assume the market proceeds in an iterative process such that: (a) In every round t, a market statistic t = h bt ( ) is made public; (b) Traders rene their information partitions according to the information brought by the market statistic; (c) Traders revise their next round expectation bit+1 s according to their new information partitions. Then, for all , there is a round T such that T is common knowledge at . Theorem 3. (McKelvey and Page [1986], Nielsen et al. [1990]) If the function h in 4.3 is stochastically regular, for any T, at which t = h bt ( ) becomes common knowledge, and for all , it must be the case that
t t t ( ) = . . . = bn ( ) = T b1 ( ) = b2

The requirements of the Theorems 2 & 3, as well as the initial information structure of 4.1, 4.2 and 4.3 are fully met by the market settings deployed in Section 3.6; as a result the theorems outcomes are directly applicable to the proposed market model. More specically, Theorem 2 suggests that, at some round T , the market price p T becomes common knowledge, in other words, the market reaches a situation where it cannot bring any new information to traders, while their information partitions can not be further rened. Thus, traders bids do not change in later rounds and market price remains stable. Theorem 2 introduced also that market achieves equilibrium after T rounds, while Theorem 3 proves that, at equilibrium, all traders happen to have the same expectation about the value of f (s), which equals to equilibrium market price.

4.4.2 Convergence Speed


According to Theorems 2 & 3, markets do converge to equilibrium, where traders expectations coincide, after T rounds. However, T is yet not uniquely determined and the question about convergences speed naturally arises. Theorem 4 comes into support.

4.4. CONVERGENCE PROPERTIES

91

Theorem 4. (Chen et al. [2006b]) A prediction market converges to an equilibrium after at most n rounds of trading. Let Sit be trader is knowledge possibility set at time t, namely the set of states in the space that trader i considers possible to be the true state at time t. Then, common knowledge possibility (St ) of all traders at time t is the set of states that are considered possible to be the true state by an outside observer who can only observe market prices without possessing any private information, given the common prior probability distribution P . In such a way, trading can be considered as a stepwise procedure of limiting the common knowledge possibility, while markets equilibrium occurs when the common knowledge possibility evolves into a set of one item. Using mathematical notation,

S 0 > S 1 > . . . > S T = S T +1 = . . .

(4.6)

where S0 denotes the whole sample space = S X = {0, 1}m {0, 1}n and T the round that market converges to the equilibrium. The time for an information market to converge to its equilibrium equals the number of rounds that an observer of the market takes to improve the common knowledge possibility set from S0 to S T . Feigenbaum et al. [2005] have shown that for any round t, the dimension of St is at least one dimension lower than that of St1 before the market equilibrium is reached. While the dimension of S0 is m + n, the dimension of S T at equilibrium in our market settings follows Chen et al. [2006b] and is at least m, due to the aggregate uncertainty. Hence, convergence to the equilibrium takes at most n rounds.

4.4.3

Best Possible Prediction

Another question of equally signicant value regards the quality and completeness of the best possible prediction that the proposed market setting can attain. Theorem 5. (Chen et al. [2006b]) The best possible prediction that a prediction market can make is the forecast at direct communication equilibrium. A direct communication equilibrium is dened as the equilibrium that is reached immediately when all market traders directly reveal their private information to each other

92

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

[Geanakoplos and Polemarchakis, 1982]. In other words, the market price at equilibrium equals the expectation of the security payoff conditional on all available information, i.e. E ( f (s)| x ). This property is rather counter-intuitive, as it describes the best possible prediction to be the result of complete aggregation of all private information distributed among traders. Example 1. Recall the simple market setting of subsection 4.3.2, with an one dimensional state space s1 and a boolean state of the world, either 0 or 1, each with probability 0.5, which is common knowledge to all traders. Assume also that the target function for prediction is f (s1 ) = s1 , thus 1, f ( s1 ) = 0, when s1 = 1 when s1 = 0

If f (s1 ) = 1, all money invested in the market are redistributed to security k and security l earns nothing, while if f (s1 ) = 0 the opposite is correct. Two traders exist in the market setting, each one of them privy to one bit of information, x1 & x2 , both following independently the identical probability distribution

P r ( xi = 0|s1 = 0) = 0.9, P r ( xi = 1|s1 = 0) = 0.1 P r ( xi = 0|s1 = 1) = 0.1, P r ( xi = 1|s1 = 1) = 0.9


Suppose true state s1 = 1 and both traders private information x1 = x2 = 1. Using Bayes Rule, we can calculate securitys k market price at the direct communication equilibrium: E( f (s1 )| x1 = 1, x2 = 1) =

= P r ( f (s1 ) = 1| x1 = 1, x2 = 1) = P r (s1 = 1| x1 = 1, x2 = 1) = P r ( x1 = 1, x2 = 1|s1 = 1) P r (s1 = 1) = P r ( x1 = 1, x2 = 1) P r ( x1 = 1, x2 = 1|s1 = 1) P r (s1 = 1) = = P r ( x1 = 1, x2 = 1|s1 = 1) P r (s1 = 1) + P r ( x1 = 1, x2 = 1|s1 = 0) P r (s1 = 0) 0.9 0.9 0.5 = 0.988. 0.9 0.9 0.5 + 0.1 0.1 0.5 =
Thus, the best possible forecast of f (s1 ) is 0.988. This implies that the true state of the world is indeed very likely to be s1 = 1, but there is uncertainty associated with the prediction. In that sense, the quality of best possible prediction is also constrained by the

4.4. CONVERGENCE PROPERTIES

93

amount of aggregate uncertainty. This uncertainty in individual information introduces aggregate uncertainty to our model, as the true state of the world is uncertain even with pooled information. For example, if we change the aforementioned probability distribution to

P r ( xi = 0|s1 = 0) = 0.6, P r ( xi = 1|s1 = 0) = 0.4 P r ( xi = 0|s1 = 1) = 0.4, P r ( xi = 1|s1 = 1) = 0.6


the expectation of securitys k market price at direct communication equilibrium would be E( f (s1 )| x1 = 1, x2 = 1) 0.692. Moreover, in the following extreme case of symmetrical to s1 probability distribution,

P r ( xi = 0|s1 = 0) = 0.1, P r ( xi = 1|s1 = 0) = 0.9 P r ( xi = 0|s1 = 1) = 0.1, P r ( xi = 1|s1 = 1) = 0.9


expected price at direct communication equilibrium equals E( f (s1 )| x1 = 1, x2 = 1) 0.500. In other words, in the latter case of xi being independent to s1 , best possible prediction simply reproduces the prior distribution of s1 . Finally, Theorem 5 and previous examples suggest that markets performance is directly related and limited to the quality of available to traders information. In other words, markets as predictive and information aggregation mechanisms cannot be expected to perform better than the totality of knowledge available to and by their participants.

4.4.4

Convergence to the Best Prediction or Not

Given the properties of markets convergence to equilibrium, where a market statistic becomes common knowledge, in n steps, as well as the direct communication equilibrium as markets upper limit of performance, the last question that arises regards the rate of markets convergence to this best possible prediction. Feigenbaum et al. [2005] proved that when target function takes the form of weighted threshold function 1 , the equilibrium price is guaranteed to equal to the value of the function. However, this appears not to be the case in general.
1A

function f : {0, 1}n {0, 1} is a weighted threshold function iff there are real constants w1 , w2 , . . . , wn

n such that f (x) = 1 iff i =1 wi x i 1

94

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

Theorem 6. (Chen et al. [2006b]) A prediction market is not guaranteed to converge to direct communication equilibrium. Example 2. Consider a prediction market with a two dimensional state space s = (s1 , s2 ) and a boolean state of the world, either 0 or 1 for each of s j , j = 1 or 2. The prior probability distribution of s for the two traders i = 1, 2 existing in the market is common and uniform, i.e. s = (s1 , s2 ) takes the values (0, 0), (0, 1), (1, 0), and (1, 1) each with probability 0.25. Target function for prediction is 1, f ( s1 , s2 ) = 0, when s1 = s2 otherwise

Two securities, k and l are traded in the market. In the case that f (s1 , s2 ) = 1, all money are redistributed to security k and security l earns nothing, while if f (s1 , s2 ) = 0 the opposite is correct. Each of the two existing traders is privy to one bit of information, x1 & x2 , both following independently the identical probability distribution, conditional on s

P r ( xi = 0|s1 = 0, s2 = 0) = 0.8, P r ( xi = 1|s1 = 0, s2 = 0) = 0.2 P r ( xi = 0|s1 = 0, s2 = 1) = 0.5, P r ( xi = 1|s1 = 0, s2 = 1) = 0.5 P r ( xi = 0|s1 = 1, s2 = 0) = 0.5, P r ( xi = 1|s1 = 1, s2 = 0) = 0.5 P r ( xi = 0|s1 = 1, s2 = 1) = 0.2, P r ( xi = 1|s1 = 1, s2 = 1) = 0.8
Using Bayes Rule, traders i expectation on price of security k can be computed, given that available information bit is xi = 1. E( f ( s ) | x i = 1) = P r ( f ( s ) = 1| x i = 1) =

P r ( x i = 1| f ( s ) = 1) P r ( f ( s ) = 1) = P r ( x i = 1) =
0.2+0.8 2

0.5 = 0.5. 0.5

If xi = 0, expectation of traders i on price of k will equally be E( f ( s ) | x i = 0) = P r ( f ( s ) = 1| x i = 0) =

P r ( x i = 0| f ( s ) = 1) P r ( f ( s ) = 1) = P r ( x i = 0) =
0.8+0.2 2

0.5 = 0.5. 0.5

4.4. CONVERGENCE PROPERTIES

95

In other words, traders expectations on price of security k (and l ) will be equal to 0.5, no matter of the price of xi each trader possess or what can infer to be possessed by the alternate trader. As a result and following the market mechanism proposed in subsection 4.3.3, in each round each of the traders will select either security k or l randomly, with equal probabilities assigned to each of them. This decision rule will not provide convergence at all, the best possible prediction included. To showcase that with an example, let true state be s = (1, 1) and both traders private information x1 = x2 = 1. Using Bayes Rule, securitys k market price at the direct communication equilibrium equals 2 E( f (s)| x1 = 1, x2 = 1) =

= P r ( f (s) = 1| x1 = 1, x2 = 1) = P r (s1 = s2 = 1| x1 = 1, x2 = 1) = P r ( x1 = 1, x2 = 1|s1 = 1) P r (s1 = 1) = P r ( x1 = 1, x2 = 1) (P r ( x1 = 1|s1 = 1) P r ( x1 = 1|s1 = 1) P r (s1 = 1)) P r (s1 = 1) = = P r ( x1 = 1, x2 = 1) 0.32 0.5 = 0.48. 0.33 =
For round one, given s and x would imply a 33% probability of both traders buying security k, or security l or buying different securities each. This decision rule is repeated in every round, no matter which the clearing prices of the previous one were, and nally results in ever increasing pairs of (qk , ql ) and no convergence or equilibrium. Example 3. Consider a variation of the previous example 2, where the probability distribution of traders information x, conditional on s is:

P r ( xi = 0|s1 = 0, s2 = 0) = 0.8, P r ( xi = 1|s1 = 0, s2 = 0) = 0.2 P r ( xi = 0|s1 = 0, s2 = 1) = 0.8, P r ( xi = 1|s1 = 0, s2 = 1) = 0.2 P r ( xi = 0|s1 = 1, s2 = 0) = 0.5, P r ( xi = 1|s1 = 1, s2 = 0) = 0.5 P r ( xi = 0|s1 = 1, s2 = 1) = 0.2, P r ( xi = 1|s1 = 1, s2 = 1) = 0.8
2 P r(x 1

= 1, x2 = 1) = P r ( x1 = 1, x2 = 0) = P r ( x1 = 0, x2 = 0) = 0.33 as the order of x1 , x2 does not

count.

96

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

Following Bayes, traders i expectation on price of security k if xi = 1 is E( f ( s ) | x i = 1) = P r ( f ( s ) = 1| x i = 1) =

P r ( x i = 1| f ( s ) = 1) P r ( f ( s ) = 1) = P r ( x i = 1) 0.2+0.8 0.5 = 1.7 1.7 0.69. =


4

If xi = 0, expectation of traders i on price of k will equally be E( f ( s ) | x i = 0) = P r ( f ( s ) = 1| x i = 0) =

P r ( x i = 0| f ( s ) = 1) P r ( f ( s ) = 1) = P r ( x i = 0) 0.8+0.2 0.5 = 2.3 2.3 0.38. =


4

Given the true state s = (1, 1) and traders private information x1 = x2 = 1, both traders will buy security k in rst round, resulting in clearing prices of pk = exp(3)/(exp(3) + exp(1)) 0.88 and pl = exp(1)/(exp(3) + exp(1)) 0.12. By observing these prices, however, each trader can infer that the information bit owned by the other trader is also equal to one. Their expectation of pk on second round then is 3 E( f (s)| x1 = 1, x2 = 1) =

= P r ( f (s) = 1| x1 = 1, x2 = 1) = P r (s1 = s2 = 1| x1 = 1, x2 = 1) = P r ( x1 = 1, x2 = 1|s1 = 1) P r (s1 = 1) = P r ( x1 = 1, x2 = 1) (0.8 0.8 0.5) 0.5 = 0.88 0.18 =

It is noted also that this expectation is equal to the direct communication equilibrium, as each trader knows the value of information bit available to herself and the other trader.
3 P r(x 1

= 1, x2 = 1) =

1 24

a,b,c,d P r ( x1 = 1|s1 = a, s2 = b) P r ( x2 = 1|s1 = c, s2 = d) = 0.18

{0,1}

4.5. DISCUSSION

97

It occurs that the market converges to the direct communication equilibrium, namely the best possible prediction, within one round of trading. If traders are forced to continue trading, they will select a security randomly in round two, but in the rounds to follow clearing prices will return to the direct communication equilibrium; the market will eventually oscillate around and converge at equilibrium. The previous examples, especially example 2, serve as prima facie evidence of Theorem 6 and the absence of guaranteed convergence to the best possible prediction. That seems to be the case when high symmetry among the prior probability distribution of the state of the world P (s) and the conditional probability distribution of traders information

Q( x |s) exists, resulting in random trading decisions and no revelation or inferences on


private information. However, this case should be considered as reasonably rare; in general the proposed market model performs sufciently well in aggregating information and shaping accurate predictions.

4.5 Discussion
4.5.1 Summary

Chapter 4, built on the market approach to forecasting unfolded in Chapter 2 and the market mechanism deployed in Chapter 3, essentially contributed a theoretical model of the market operation, which proved to be a proper tool for studying a set of properties, critical to the markets utility as an information aggregation mechanism. The model developed, as well as its most signicant results about the properties studied, are efciently described in Table 4.1, which follows hereafter. In a few words, the market process for information aggregation was shaped, following Feigenbaum et al. [2005], as a computation on distributed information and a simplied model of a prediction market was unfolded, that captures many essential aspects of real agent interaction. Within this model, various outcomes emerged regarding the quality and speed of markets efciency as a computational tool. More specically, it was proved that the under model market is guaranteed to converge to an equilibrium, at which all traders arrive at consensus about the market forecast. This convergence happens in a number of

98

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

Review of the Market Model Information structure State of the world: s {0, 1}m with common prior probability P (s) Trader information: n traders, trader i holds xi : x Market modeling

{0, 1}n with conditional probability distribution Q( x |s)


Market mechanism Market structure: Pari-mutuel market, securities
exp(qk ) , exp(qk )+exp(ql )

{k, l }, priced as pk =

where q quan-

tities of shares bought Pay-off: All money redistributed to the winning security, winner is dened according to f (s) The market is a restricted Shapley-Shubik market game Trader behavior Traders type: risk-neutral, myopic, bidding truthfully rather than strategically Traders probability distribution is updated after each trading round via Bayes rule Price convergence Convergence properties Convergence speed Best possible prediction Converge to consensus equilibrium in nite steps At most n rounds, where n is the number of traders Direct communication equilibrium, where p = E( f (s)| x ) Convergence to the best possible prediction Not guaranteed

Table 4.1 Review of the market model

4.5. DISCUSSION

99

rounds less or equal to the number of traders participating in the market. Moreover, it was showed that the best possible prediction the market setting is able to shape is the one of direct communication equilibrium, namely the expectation of an equivalent omniscient super-agent who possess all information available to all traders; however the market is not guaranteed to converge to this optimal state. The latter property stands as a critical vulnerability of the proposed market setting, providing a target for future contributions. Apart from that, the model -at the extent its abstractness can afford- appeared to validate the essential value of markets as a decision support tool for predictions and information aggregation and sufciently served the needs set as the core focus of this Chapter. Theoretical examination also provided a by default solid base for studying a market mechanism and its operational properties, however, the framework shaped each time is not easily molded or customized, in a way that leads to signicantly reducing the level of detail. Following that very need, the model described stands as an abstract departure from reality, a fact resulting in a strong recommendation to apply all of the conclusions of this Chapter with great caution, as long as mostly all of the assumptions made are not validated in real settings and actual prediction markets. The pari-mutuel market structure selected serves as a unique contribution to the literature, utilizing the price function introduced in Chapter 3. The results though proved to be quite similar to the ones of Chen et al. [2006b] who utilized a more traditional market setting with analogous aggregate uncertainty, an outcome providing further credence to the pari-mutuel mechanism and the price function in use. However the Shapley-Shubik nature of the market game introduces some limitations to the model. The utilized trading rule -of one share bought per round and agent- seems to be quite abstract to track any non-experimental prediction market setting. Moreover, while Theorems 1, 5 and 6 are sufciently robust, small perturbations in the selected price function of the restricted game may result to invalidating Theorem 4, as Chen et al. [2006b] points out. Furthermore, the assumption of truth-telling agents seems reasonable only when a large number of traders exists, as in the opposite scenario strategy shapes most of the trading decisions. However, the hypotheses taken indicate at the same time paths for future deployment of the proposed market setting, or the theoretical evaluation of prediction markets in general. Thoughts and directions are provided in the paragraph to follow.

100

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

4.5.2

Future work

Among the various paths for future contributions shaped out of this Chapter, a rough review would include the following ones as potential directions of interest, based on relaxing the assumptions of the proposed market setting. (a) information structure, state of the world: both the boolean nature and the common prior assumptions seem to be trivial, however a simple differentiation would result in a disproportionate increase in the complexity of the model. At the same time, studying the market behavior in the case of non boolean information structure or when that agents priors differ -signicantly or not- stands as a quite attractive research potential. Useful insights would also offer the possibility of some information bits to change during the markets operation. (b) information structure, trader information: various departures from the proposed setting could be set up, for example each trader may hold a fuzzy set of information, ranging from 0 to n information bits, next to the differentiation level a different state of the world could introduce. (c) market mechanism, market structure: if the pari-mutuel setting is taken for granted, it is important to explore the impact or robustness of the market setting according to the price function selected. The question is, do other price functions result to same or similar convergence properties, or not. (d) market mechanism, pay-off: the option of a losing money redistributed, instead of all money redistributed, as initially proposed by Pennock [2004], may be worth investigating. Also, as stated before, the Shapley-Shubik nature of the game implied some critical constraints, thus the use of a different setting is denitely of signicant value. (e) traders behavior and type: the study of the impact various traders strategies may imply to the given convergence properties remains a topic of crucial interest, as the assumption of them always bidding truthfully is rather naive, or may occur roughly when the number of traders is large enough.

4.5. DISCUSSION

101

(f) guaranteed convergence: the proposed setting was not successful in providing guaranteed convergence to the best possible prediction, namely direct communication equilibrium. The shaping of a market setting that satises this objective is a major milestone for future research attempts. In a nutshell, one may say that the market setting proposed could be visualized as a decision tree, where each of the assumptions and paths followed here introduces a separate branch and consists a single leaf, but at the same time the whole model setting introduces numerous options for further deployment, equal to the number of possible combinations of leaves. That said, it should be also considered that the theoretical approach to the problem is by nature rigid and increasingly complex, so one may need to turn to experimental approaches and simulations for a more extended study of this issues.

102

CHAPTER 4. THEORETICAL EV ALUATION OF PREDICTION MARKETS

Chapter 5

Empirical Analysis of Prediction Markets


5.1 Overview
Chapter 1 provided the generic context of the forecasting problem this Thesis attempts to address, while Chapter 2 described the market framework towards its solution. In Chapter 3, an algorithmic development for the efcient operation of this framework was suggested, with Chapter 4 providing theoretical evidence of the appeal and effectiveness of such a framework. The hereby Chapter takes the theoretical evaluation of the proposed market framework further, by empirically studying the virtues and vices of its application. A number of experiments in both academic and business environments are run to verify the applicability of the theoretical ndings, as well as to provide insights and suggestions for their effectual implementation. These experiments attempt to take advantage of the existing pool of knowledge aggregated from the signicant number of prediction markets empirical studies already performed. The analysis of the collected experimental results also serves as input to a number of structured guidelines for putting the proposed market framework in practical use. It needs to be mentioned though that the transition from theoretical topics to what may be considered as social science at its core comes at a price; there is a departure from denite results and proofs of theorems to suggested uses and best practices under dynamically varying settings. That said, the task to examine and 103

104

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

demonstrate the practicability of the theoretical ndings previously deployed is essentially performed in the hereby Chapter, followed by what may serve to be the most practical part of the Thesis, namely the proposed prediction markets implementation guide. In more detail, Section 5.2 provides a short summary of the various empirical studies for prediction markets already performed, further expanding on Section 2.4s ndings. Section 5.3 introduces the software platform developed and used for the purposes of this Thesis. Sections 5.4, 5.5, 5.6 summarize the designed specications of three among the various experiments run, along with analyzing and discussing their results. Leveraging on such experiments, Section 5.7 attempts to provide a generalized framework for the practical deployment of a prediction market implementation. Finally, Section 5.8 summarizes the outcomes of this empirical exploration, discussing the mechanisms drawbacks and benets, next to suggesting specic directions for its further study.

5.2 Review of Empirical Studies


5.2.1 Background

The evaluation of the proposed market mechanism deployed in Chapter 4 provided theoretical proof that the under study model is guaranteed to converge within nite steps to equilibrium, where price expectations become similar across traders and all of them agree on a market forecast. At the same time, it was shown that the arrival at direct communication equilibrium, a status at which the sum of information available to various traders becomes common knowledge, may occur, but it is not guaranteed. The experimental verication of both those theoretical outcomes is of primary importance, essentially judging the applicability of theoretical ndings into the real world. However, such experiments may only serve as an indication that the mechanism is capable of delivering the desired results, with such results remaining of varying difculty given the endemic uncertainty in a setting with real traders compared to one with articial agents. More specically, the existence of direct communication equilibrium is truly difcult to be validated in practice, if it is possible in a meaningful way at all, and the lack of any such reference -at least in authors knowledge- further supports this understanding. The

5.2. REVIEW OF EMPIRICAL STUDIES

105

verication of a generic equilibrium seems to be more feasible and indicated by, say, zero trading volume. However, in real world settings trading does not happen in serial rounds and the lack of liquidity may in fact suggest the failure of the selected market mechanism, or the lack of traders interest per se, which further obfuscates the issue. Theres another aspect though of a real world prediction market implementation that may be considered as equally interesting or inherently similar to the study of equilibrium, and that is the convergence of the market price to the actual forecasting target. As expected, the proposed market framework has yet to be tested in practice. That said, questions like the later one set above have been studied in a number of references, under different market settings, constituting the major part of the literature as demonstrated in Section 2.4. Politics and sports typically serve as the canvas for such studies, with other mostly business applications as well as laboratory experiments also contributing a signicant part in this segment of the literature (see Table 2.4). The equilibrium issue appears to be the dominant one across those references, mostly appearing under the term market accuracy. In other words, most studies attempt to assess the success of a prediction market in terms of the markets accuracy in predicting the real outcome of the markets question, well before this occurs. The earlier this happen and the more stable such a prediction is, the more accurate a market is being considered. In fact, this interpretation is not necessarily remote from an equilibriums denition, so for the needs of the hereby Chapter those questions will be treated equally. In this context, the related work further analyzed below will focus on summarizing the performance of prediction markets, as this is manifested in the existing corpus of literature.

5.2.2

Related Work

A big part of the related literature focuses on comparing the accuracy of prediction markets with the one of polls. As Pennock and Sami [2007] enunciates, a prediction market or a forecasting mechanism in general cannot take advantage of inexistent or unknown information, so its accuracy can only be evaluated in comparison to other such mechanisms. Adopting this view, a number of studies, especially in the political eld, have been performed with varying results. Kou and Sobel [2004], for example, suggests

106

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

after a number of experiments that market forecasts have smaller variance than poll-based forecasts, with the rst dominating the later in the case that polls are included in the information set available to the market participants. Plott [2000], Forsythe et al. [1992], Berg et al. [2000, 2003], among others, provide further experimental evidence that market forecasts can outperform polls, among other alternatives, while they provide unbiased, relatively accurate forecasts well in advance of outcomes [Berg and Rietz, 2003]. Gruca et al. [2003] takes this further to also provide some experimental evidence on markets accurately aggregating both private and public information reported in survey forecasts, with Lucas [1972], Grossman [1981] stating that price stands as a sufcient statistic for all information being available to traders in a market. Forsythe et al. [1999] also provides evidence that prediction markets perform quite well in the context of politics and elections, in spite on authors substantial evidence that traders behave on average far less rationally than what theory would suggest. Manski [2006] further veries the capacity of markets to efciently aggregate information. However, the author also mentions a number of negative issues with regards to the market operations, including market manipulation [Forsythe and Lundholm, 1990, Noeth and Weber, 2003] and the inability to settle on an equilibrium price [Anderson and Holt, 1997, Sharfstein and Stein, 1990]. Hahn and Tetlock [2006b] further adds on such problems which markets tend to suffer from, referring to manipulation [Hansen et al., 2004], illiquidity [Sunder, 1992], information traps [Camerer and K., 1991, Noeth et al., 1999] and lack of meaningful equilibria [Anderson and Holt, 1997, Scharfstein and Stein, 1990]. References such as the above eventually manifest the wide variety of issues arising from the empirical study of markets predictive operation. Signicant issues remain open and almost certainly cannot be but so, given the social nature embedded in the mechanisms core. In this context, what follows next is the description of the experiments performed taking advantage of the insights accumulated in the existing corpus of literature, the deployment of their results and the conclusions that can or cannot be drawn for the implementation of the proposed market mechanism.

5.3. THE ASKMARKETS PLATFORM

107

5.3 The askmarkets platform


For the required experiments to be facilitated, given the novelty of the prescribed market mechanism deployed in Section 3.6, a brand new software platform for prediction market was developed from scratch. Together with Efthimios Bothos [Bothos, 2010], who has also contributed a wide series of publications on the topic [Bothos et al., 2007, 2008a,b,c,d, 2009, 2010a,b, Bothos et al., 2011, Mentzas et al., 2011, Bothos et al., 2012], the author worked to create a customizable tool for prediction market deployment, that would follow business beyond academic rules to be utilized in broader than laboratory settings.

Figure 5.1 Homepage, screenshot

The platform, named askmarkets1 , took advantage of recent technologies to provide a


1 At

the time of writing, the platform can be accessed on the web at http://www.askmarkets.com

108

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

tool for the creation, participation and administration of marketplaces, market questions and stocks alike. Figure 5.1 provides a view of the platforms homepage; Figure 5.2 pictures an indicative view of a market trading page, while Figure 5.3 showcases the market creation process. The tool enjoyed some popularity and international coverage among early adopters, however it failed to address the entrepreneurial aspirations of its creators for a number of reasons. No matter of the irrational -from an academic standpoint- amount of effort it took, the platform efciently served the experimental needs of a researcher and was utilized in a series of cases, a number of which are deployed and analyzed below.

5.4 Experiment A
5.4.1 Design

The rst of the experiments conducted took place in an academic context. A number of graduate students from the School of Mechanical Engineering, National Technical University of Athens participated in a marketplace, both in the laboratory and from home, for a period that lasted for about two weeks. The subjects initially were not familiar with the notion and usage of prediction markets, stock markets or other forecasting mechanisms. A two-hour introductory session was conducted as a short course to forecasting in general, as well as to explain the concept behind the mechanism and the technicalities of the software platform. Participation was optional, with students participating and submitting a report on their approach and experience receiving up to 1 extra point in the courses total scale of 10. Participants were given the option to either trade in existing markets or create their own ones. The theme was the Athens 2004 Olympic Games and the Olympic Case Global Impact study that has been performed thereafter [Tziralis et al., 2006]. For the purposes of the experiment, which took place during November 2008, the students got access to a repository of related information, which, should sufcient time was allocated, could lead to the right answer in every market question. Moreover, before the trading started and access to the information repository was provided, participants were requested to provide

5.4. EXPERIMENT A

109

Figure 5.2 Market page, screenshot

110

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

Figure 5.3 Market creation page, screenshot

5.4. EXPERIMENT A

111

an uninformed answer on a scale from 1 to 10 to a small subset of yes/no questions. In this context, the scope of this experiment was twofold; rst, to compare the results of the questionnaires answers with the markets predictions and, more importantly, to study the convergence or the lack thereof of the markets results to the correct answers, as these were suggested by the already available data. It needs to be further claried here though that the term prediction market is rather inappropriate in the specic case of the hereby experiment, as data was already available and the answers were known in advance to each markets creator. Nevertheless, the case proved to be sufcient for the study of the core equilibrium issue, as this was deployed in Section 5.2.1; its analysis and results follow hereafter.

5.4.2

Analysis

More than 100 participants traded in more than 200 markets, the majority of which were created from the participants themselves. Almost four thousand transactions were facilitated. These mere numbers suggest the usability of the platform as sufcient, for the purposes of these experiments and beyond. Table 5.1 provides a comparison and summary of the results in 12 of those markets, the ones whose questions were also addressed in the questionnaire. To make results digestion easier, the correct answer in each one of the markets was positive (1.0). Column 1 provides the markets id, column 2 the average of the boolean answers (0/1) collected in the questionnaire, column 3 the market price of the correct answer before market closing and column 4 a qualitative record of market convergences speed to the nal result of column 3. Figure 5.4 provides an example of a market with slow convergence.

5.4.3

Results

Questionnaires results presented an average error of 0.33, compared to markets 0.23, while RMSE was 0.35 and 0.29 respectively. The improvement is a considerable one, however it comes at the cost of running a marketplace, compared to the simplicity of a traditional survey. The variance of convergence speed -half of the markets converged fastis also of interest, as markets performed better yet it took them a variable time to converge.

112

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

Figure 5.4 Example closed market page with slow convergence, screenshot

5.4. EXPERIMENT A

113

market id 1 2 3 4 5 6 7 8 9 10 11 12

survey 0.45 0.59 0.54 0.74 0.70 0.84 0.62 0.78 0.77 0.54 0.64 0.81

market price 0.6 0.9 0.7 0.7 1.0 0.6 1.0 1.0 0.6 0.5 0.7 0.9

convergence slow fast fast slow fast slow fast slow fast slow slow fast

Table 5.1 Experiment A, survey & market results The above results suggest market usage in the context of this experiment as successful. In contrast to the survey, every single market of the sample questions converged to the desirable outcome, however under varying accuracy and maturation time. Results showcased non-incremental improvements compared to a standard information aggregation method. On the other hand, the standard method provided results that might deem sufcient in a number of cases, so it eventually is a matter of choice, given the context of each problem at hand and the time, cost and accuracy limitations per case. In conclusion, this rst experiment suggests prediction markets and the prescribed mechanism as a premium tool for dynamic information aggregation.

114

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

5.5 Experiment B
5.5.1 Design

The second experiment was facilitated in a business environment. A Greek publicly traded company that manufactures, modies, distributes and supports a range of vehicles and other automotive machinery served as the context of the markets deployed. A number of employees, across various levels of the organizational chart, participated in a couple of markets with questions related to sales estimates of core company products. No training was conducted, apart from a short introduction to the prediction market concept and the platform itself to a select few of higher level executives. Most of the participants were actively engaged in the supply chain, being either salesmen or warehousemen. They thus had direct knowledge and input to contribute regarding the issues under study with the market mechanism. Participants did not have the option to create new or edit the existing markets and specic stocks. The experiment started in March 2009 and lasted until August 2009, a relatively long period that provided the opportunity to reect on the timely accumulation of information, or the lack thereof, by the market platform. No access to any further information was provided to participants due to the experiment, but the information that each one got exposed to due to her business as usual routine. In general, emphasis was put into avoiding any modications in the way of conducting business for the company, thus eliminating bias on the results. In this context, the companys administration decided to offer a prize to the market participant with the biggest portfolio value, but communicated this only after the closing of the marketplace. Considering the above characteristics, the scope of this experiment primarily was to assess the applicability of the market mechanism under study in a strictly professional environment, more specically its capacity to accumulate and broadcast information that is of business value for an enterprise on a forecasting basis and beyond. Secondarily, this specic experiment was also designed to study the comprehensiveness of the concept and the platform per se in an environment with divergent interests and incentives, compared to the previous academic one. Its analysis and results are provided below.

5.5. EXPERIMENT B

115

5.5.2

Analysis

265 participants, the majority of companys employees, were automatically registered in the marketplace and informed for its operation via a newsletter. Out of those, 32, or 12,1% made at least 1 transaction in one of the two markets operated. Figure 5.5 pictures the distribution of transactions per active user. 664 predictions were made in total. Both of the prediction markets operated focused on the pickup trucks market, due to this being of signicant value for the company.

Figure 5.5 Number of transactions per active user

Figure 5.6 provides a close look at the evolution of the various stocks prices in the rst one of the markets. This market focused on the prediction of the total market size, while the second one focused on the prediction of the market share for a brand owned by the company. Among the 9 contracts being traded in the rst market, the one indicating market size between 8 and 9 thousand vehicles was suggested from early on as the most probable outcome, retaining a probability higher than 45% for most of the duration of the markets operation. The correct result proved to be a volume between 7 and 8 thousand

116

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

vehicles.

Figure 5.6 Market A, stock prices evolution, graph

Figure 5.7 presents a screenshot of the nal results for the various answers provided as stocks in the second market under study. Among 11 available outcomes, the market eventually suggested the correct outcome as the most probable one (40% probability of occurrence), a few months before the nal data became available. The screenshot also presents the short historical data that became easily accessible to traders within the platform in the form of market description.

5.5.3

Results

The limited size of this experiment denitely cannot provide for generalizable outcomes. At the same time, however, the specic nature of the study suggests the attempt to articulate qualitative conclusions as of value. In the context of no specic incentives for participation and the lack of any active communication of the tool within the company, the participation rates presented above suggest the marketplaces acceptance and utilization

5.5. EXPERIMENT B

117

Figure 5.7 Market B, nal results, screenshot

118

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

as sufcient. The tool and the platform per se arose as potentially usable, even under the tough requirements of a strictly professional setup. The markets failed to predict the correct outcome in one of the two cases. However, even in this case results were sufciently close to the desirable outcome. Along with the previous one, this experiment essentially provides empirical evidence on the capacity of prediction markets and the algorithm introduced to converge into desirable results with signicant predictive and information aggregative value. To further support this, the manager being responsible for the projects implementation considered the aggregation of available information that was facilitated with the market tool as materially valuable and considerably improved, compared to more traditional judgmental or quantitative approaches that were being utilized within the company before.

5.6 Experiment C
5.6.1 Design

The third experiment took place in an online social context. A marketplace was built and customized for a popular Greek news portal, themed under the local 2009 legislative elections. Online visitors of the hosting website were invited to subscribe and participate in the marketplace, to contribute their information and beliefs, having their voice heard while shaping election result estimates. Participants did not receive any training on the concept or the mechanics of the marketplace and learned to use the platform by themselves instead. Only administrators were able to create new or edit existing markets and stocks. The experiment was initiated a couple of weeks before the October 4 elections, and it was actively promoted only during the last week. A small prize (a collectors item dvd) was provided to the participant ranked rst in the leaderboard by the election date, as a small extra incentive to participate. This experiment served as an attempt to study the dynamics and effectiveness of a prediction marketplace under an open social setting. Due to local legislation, it was prohibited to publish results of opinion polls for the last 15 days before election, a fact that did not provide the opportunity to compare the markets results to the ones of the polls in

5.6. EXPERIMENT C

119

the same period. However, it remains of interest to examine the capacity of the platform to attract a number of internet visitors and entice them to participate as traders. More importantly, the experiment may serve as another data point to the core issue of markets ability to reach equilibrium, as described in Section 5.2.1. The experiments analysis and results are provided below.

5.6.2

Analysis

45 traders participated and submitted 706 predictions in total across the 14 markets that were created. Half of those markets facilitated considerable trading volume and are studied below. Figure 5.8 provides an example of a market with relatively fast convergence to the desired outcome. It is clear, however, that the biggest part of the trading volume was facilitated during the second week of the markets operation, due to the reasons described above. Table 5.2 proceeds to summarize the results of the 7 markets that reached sufcient for the purposes of this study volume. Column 1 provides the markets id, column 2 the number of stocks (different answers) traded in the specic market (question), column 3 the nal price (upon market closing together with the elections) of the stock that proved to be correct, column 4 records if the markets most probable outcome was realized or not, and column 5 refers to the convergence speed to the most probable outcome. In this experiments case, convergence refers both to speed and uctuation of the results, so that fast convergence denotes a market that converged fast and without any uctuations to the result that was eventually suggested as the most probable forecast.

5.6.3

Results

The results above suggest the forecasting performance of the market mechanism as positive. In none of the cases above market prediction proved wrong. Moreover, the total volume of transactions facilitated and the number of participants serve to demonstrate the platforms sufcient functionality and convenience, also taking into account the lack of training and signicant incentives to the users. That said, the specic experiment was of poor value as a comparison tool to other forecasting mechanisms, as participants might have actually been trading on relatively secure information about the election results that was possibly

120

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

Figure 5.8 Example market, stock prices evolution, graph

market id 1 2 3 4 5 6 7

stocks (#) 2 2 3 2 3 2 2

nal price 0.86 0.54 0.51 0.53 0.57 0.81 0.72

correct prediction yes yes yes yes yes yes yes

convergence fast slow slow slow slow fast slow

Table 5.2 Experiment C, market results

5.7. A DEPLOYMENT FRAMEWORK

121

available a few days before the elections. At the same time, however, the experiment stands as another positive case where random web visitors were incentivized to participate and share their beliefs and available information in a context that enforces rational behavior and eventually works as a prediction tool. In other words, the hereby experiment suggests that the under study mechanism is capable of converging to the actual forecasting target, a fact that, according to Section 5.2.1, shall be considered as another soft empirical verication of the proposed market mechanisms capacity to reach equilibrium.

5.7 A Deployment Framework


5.7.1 Context

The previously described experiments, as well as a considerable number of smaller ones performed, will serve as data points in an attempt to shape a practicable framework for the deployment of prediction markets that this section will strive to provide. Following the gist of the hereby Chapter, such data points naturally cannot lead to universal principles that should be blindly followed in every prediction market implementation. On the other hand, they sufce to provide a practical guide instead, with some better-than-average practices that might be valuable to consider when deploying a market tool for forecasting purposes. In this context, such a section might actually be sitting on the edge of an academic Thesis purposes. Nevertheless, it is the belief of the author that this contribution will stand as one of the most useful parts of the dissertation on hand and is therefore duly included. The totality of experiments performed utilized the market mechanism proposed in Chapter 3 and the software platform described in Section 5.3. As a result, the good practices for deployment that will be provided will focus on those specications that remain customizable within the combination of the above, rather than a more generic framework covering the parts that were not modiable. More specically, no prediction market mechanism of the ones described in Section 3.2 or pay-off functions will be compared and non-play-money markets would not serve as an option (given that no experiment was facilitated as such, even if the mechanism can naturally operate with real money), among others. That said, some of the suggestions provided within this

122

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

Section may normally apply to prediction markets installations that go beyond the above combination of market mechanism and software platform, thus not strictly limiting the frameworks practicability to its pre-specied degrees of freedom. Notwithstanding the above limitations, before proceeding to the deployment framework itself, it is of value to repeat the core reasons why a market mechanism is capable of serving as an efcient tool for information aggregation and forecasting purposes. In short, and similar to what was analyzed in Section 2.3 before, the power of prediction markets derives from the very fact that a diverse number of agents are incentivized to continuously research, discover and reveal information [Hahn and Tetlock, 2006b, Wolfers and Zitzewitz, 2004a], while the market itself features an algorithm to aggregate such information and provide a sufcient statistic for the dynamic dissemination of such information on its entirety [Lucas, 1972, Grossman, 1981, Gruca et al., 2003]. At the same time however, as Hahn and Tetlock [2006b] aptly summarizes, there occur various issues in practice such as information traps [Camerer and K., 1991, Noeth et al., 1999], illiquidity [Sunder, 1992], manipulation [Forsythe and Lundholm, 1990, Noeth and Weber, 2003, Hansen et al., 2004], and lack of meaningful equilibria [Anderson and Holt, 1997, Scharfstein and Stein, 1990]. A deployment framework needs to take into account such capabilities and issues to respectively tap on the rst and limit the second; these provide the context for what is attempted in the paragraphs below.

5.7.2

Design

Eventually, the constituents of a prediction market are essentially two: its stocks and traders, along with the mechanism that brings them together. In the paragraphs that follow, simple directions for the effective design of each one of these elements will be provided. Before further delving into it though, it is important to highlight the generic rule of simplicity and intuitiveness that is highly recommended to be followed across all issues of prediction market design. With volume being the fuel of a markets efciency, it needs to be understood that for every design parameter resulting into a traders second thoughts before making a transaction, traders are likely to entirely skip their participation in the market. Taking this further, one may argue that the biggest obstacle to broader

5.7. A DEPLOYMENT FRAMEWORK

123

prediction market adoption so far was exactly that most attempts (including the authors venture) focused on educating the audience on what prediction markets are rst, and then requesting from users to participate. This proved to be a critical barrier to entry that most people were not sufciently concerned to pass, especially in cases performed far away from an academic lab. What is hereby suggested as a potentially more efcient solution is exploiting narratives and terminology that the vast majority of potential participants are already well familiar with, for example leveraging on the metaphors of stock markets or wagering.

5.7.3

Stocks

Moving further into the two core elements of market design and beginning with stocks, the right choice and clear denition of a forecasting goal that is converted into stocks is crucial for the success of a prediction market implementation. At the same time, this remains a topic that is often perceived as a secondary one, resulting in mediocre outcomes for the whole setting. With regards to intuitiveness, a number of properties that have been utilized within the proposed mechanism, as these were deployed in Section 3.4, further improve the comprehensiveness of such a tool, a factor that is expected to increase transactions volume. The experimental comparison of various market mechanisms on this base may actually provide for an appealing topic for future research. A naive yet common mistake in stocks design is that prediction targets are not naturally related to specic future events. Another one is poor description of the event that each stock refers to, leading to misconceptions and deviations between available information and realized transactions. As Wolfers and Zitzewitz [2004a] refer, it is important for stocks denition to be clear, easily understood and easily adjudicated. The textual description provided for a market and each one of its stocks should also strike the right balance at which random visitors will receive sufcient information to immediately understand the question at stake and participate, while more eloquent users will get triggered to also look for further information and return back to contribute it via trading in the various markets stocks. Stocks may refer to, among others, the prediction of an absolute number (i.e. sales

124

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

volume in a particular time period), a relative number (a % percentage, i.e. market share in a particular period) or the occurrence or not of a particular event (i.e. the completion of a project at a particular point in time) [Spann and Skiera, 2003a]. The authors limited experience suggests that it is better for the number of stocks to remain small, e.g. not more than 4 or 5 traded in each market. A small number of stocks improves the comprehensiveness of the market, while limiting behaviors such as favorite long-shot bias and others. Moreover, stocks referring to predictions of absolute numbers seem to be more easily comprehended compared to predictions of relative numbers. So, it is recommended that relative predictions are accompanied with quantitative descriptions (i.e. high (>80%) rather than 80% being the stocks description on the above market share example).

5.7.4

Participants

A market structure can contribute nothing but efciently aggregate the sum of information available to its traders at best. In other words, if no sufcient information is available to participants or exists at all, the market outcomes will respectively be of low value, if any. Assuming however that such information exists, thus a market setup is of value, the tasks that naturally arise are to a) nd those persons who possess or are capable to discover such information and b) incentivize and engage them to so do. This paragraph intends to summarize some practical experience on how to succeed in these tasks. Common sense and popular consciousness may be misinterpreted to suggest that, if you provide sufciently generous incentives, especially in the case of markets broadly open to the public, people holding relevant information will nd their way through to stumble across the market and participate to contribute their insights to it. Unfortunately, however, such beliefs rarely get reected in practice and low or poorly targeted participation ends up to be the core reason of the mechanisms failures. As a result and when it comes to participants, incentives design and optimization does not sufce. This goes at odds to what the majority of existing literature suggests, by focusing solely on the study of incentives and their optimization. Incentives alone are of limited impact if a targeted pool of participants is not aware of the markets existence,

5.7. A DEPLOYMENT FRAMEWORK

125

or their engagement remains low across the time window of markets operation. The study of such factors has probably little to do with operational research and mostly resides with social sciences and more specically marketing, however a few notes will be provided hereby as they are considered of particular value for the successful operation of a prediction marketplace. In short, a prediction market tends to be successful when incentives are not required for its efcient operation. At the same time and in the authors limited experience at least, the availability of incentives typically cannot reverse the lack of required volume from those who possess valuable information. This kind of an oxymoron might be against common knowledge from other mature domains, for example the stock market, however it was proven valid in every single experiment attempted and might be an interesting insight to keep in mind when designing a prediction marketplace. Lets take an example of a market with the forecasting goal of sales volume prediction, in which a number of salesmen are participating. For the market to become successful and its estimates valuable, salesmen are required to repeatedly check for current market prices and contribute their expectations in case these differ from the current consensus. This wont happen if incentives are, for example, monetary and of signicant height; after all they cannot be higher than, say, sales bonus per se, or else the market will be defeating its purpose. A market designers target should therefore be to turn the targeted audience to an active community of people who take pride in participating and having their say in the market. Within such a community, adequate competition dynamics will be organically created, turning the market itself into a daily topic of conversation. If the above apply, incentives will be of little value and, to give an example, ranking in a play-money leaderboard will probably sufce for participants to repeatedly search and commit information to the benet of markets efciency.

126

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

5.8 Discussion
5.8.1 Summary

This Chapter attempted to provide some empirical evidence on the practicability of the theoretical ndings of the previous Chapters 3 and 4. Along with the study of related empirical work, a customizable software platform implementing the proposed theoretical framework was developed from scratch and was utilized to perform a number of experiments in discrete environments. After the platforms description, three of those experiments were studied in detail, in an academic, business and social setting respectively. All three experiments consistently suggested that -no matter its variable performancethe proposed market mechanism is capable of converging into equilibrium. Within the limitations of the particularly social aspect of such experiments, a practical framework for the deployment of a prediction market implementation was also described, providing some directions for researchers and practitioners alike to take into account when designing similar experiments.

5.8.2

Future work

The empirical nature of this Chapter provides for a practically limitless pool of directions for future research, most of them being of qualitative nature and probably not having much to do with a Thesis in Operations Research. That said, the hereby Thesis contributed nothing but scratching the surface of what might be studied with regards to the proposed market mechanism and its empirical examination. A profound extension, as mentioned earlier in this Chapter, would be to compare the performance of the proposed mechanism to other, more established ones. The difculties for such an endeavor would be considerable though, since the replication of a social experiment is hard if feasible at all. However, a number of useful conclusions might be drawn out of such a comparison, no matter its approximative nature. Another direction might be the experimentation with a real-money market setting. Results are not expected to differ though, as the algorithm is fully appropriate for such usage and features a number of desirable characteristics for monetary purposes, while

5.8. DISCUSSION

127

incentives are suggested to be a less important parameter for a markets success compared to what is typically perceived. What seems to remain a challenge, not limited to the academic scope of the document at hand, is the close study of markets usage and effectiveness in real-world cases where there is commitment in decision making, repeatedly and entirely based on the markets outcomes. The author has condence in actual, real-world usage being the catalyst for the signicant progress of the research eld and beyond. Furthermore, it is also to be noted that the askmarkets platform remains available to anyone willing to take these ideas and experiments further.

128

CHAPTER 5. EMPIRICAL ANALYSIS OF PREDICTION MARKETS

Chapter 6

Advanced Prediction Markets Concepts: Event Detection


6.1 Overview
This Chapter presents a divergence from the remaining part of the Thesis, in studying an advanced topic that may be considered detached from the ow of the work presented so far. Data from a well-functioning marketplace -not related to the proposed market mechanism- served as input for an attempt to study the capacity of prediction markets to operate as a platform for event detection, on top of the core functions discussed so far. The required theoretical background is provided on why such a function makes sense, before the suggested process of volatility analysis is investigated through some experimental data and conclusions are derived out of it. More specically, the Chapter is organized as follows. Section 6.2 introduces the concept of event detection and the rationale behind its attempted connection to prediction markets. Section 6.3 presents background concepts and state-of-the-art information on text mining in market prediction and volatility analysis. Then it proceeds to describe techniques for analyzing time series of prediction markets in Section 6.4, including the pre-estimation, parameters and post-estimation of a Garch model analysis. Section 6.5 sets the framework for a model of volatility that serves to analyze information ows and detects events, while Section 6.6 provides details for the experiments performed and Section 6.7 129

130CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

deploys their results. Finally, Section 6.8 summarizes the outcomes of the Chapter and suggests directions for future work.

6.2 Introduction
Topic Detection and Tracking (TDT) stands in the last decade as a prosperous and promising research eld [Allan et al., 1998, Allan, 2002a]. The vision of a robust system that would monitor news streams and reliably alert an analyst to new and interesting events happening has fueled a large number of research initiatives and seems today more tangible than ever. Out of the various progresses made and tasks that comprise the core of the TDT portfolio, emerges the principal assumption of information organization by event, rather than by subject. This Chapter focuses on the primary task of event detection (or clustering). An event is typically dened as something that happens at a particular time and place. Then, event detection arises straightforwardly as the task of monitoring news corpus to discover stories that discuss a previously unidentied event [Yang et al., 1998]. The importance and wide applicability of a reliable solution to this task has fueled a number of research attempts [Allan, 2002b], with the majority of them focusing on analyzing the original news sources. However, the rapid growth of news streams available at the information age makes this task resource-intensive, if yet viable. The latter fact highlights the need for a meta-mechanism that aggregates information from news corpus by nature, to serve as input for such an analysis. Markets seem proper to handle this role. Information aggregation is a function taking place inherently in all of well-established nancial market institutions, including stock [Grossman and Stiglitz, 1980a], futures [Jackwerth and Rubinstein, 1996a] and betting markets [Thaler and Ziemba, 1988a]. Following these ndings, a considerable number of papers attempt to correlate news with stock returns, in order to detect emerging stories, or, more usually, trying to utilize the counterpart of the problem to foresee the uctuation of stock prices [Mittermayer and Knolmayer, 2006]. Among them, intelligent information triage of MacSkassy et al. [2001] focuses on one-hour stock returns and suggests labeling a story as important if the return of any stock associated with the story in the hour after the story appeared is greater than one

6.2. INTRODUCTION

131

standard deviation from the norm. Also, Lavrenko et al. [2000] proposed the co-mining of stock time series and news stories, by introducing piecewise linear tting to the time series, assignment of labels to the linear trends and nally the extraction of highly indicative language patterns to be reverted for trend forecasting. While these contributions testify that signicant price changes could be perceived as a signal of an important event, they also suggest that there remains a lot to be explored. Firstly, if information aggregation is the reason behind switching into markets for information retrieval, then a more native host of this market function might be more appropriate. In this context, following the denition provided in Section 2.3, prediction markets arise by their very nature as a unique t to facilitate such a task of tracking information ows for event detection. As Pennock et al. [2002] puts it, perhaps the relationship between price and information is no more clear than in betting markets. Furthermore, the plurality of topics encompassed in prediction markets -often empowered by user generated content- is typically greater than the limited set of themes usually covered by other market institutions. In this work prediction markets are used to enable detection of events through relevant information ows while focusing on political prediction markets, namely markets of outcome tied to specic political events, such as the democratic nominee of 2008 and the winning party of the US elections. Secondly, regarding event detection, the most relevant to this research topic approaches have gone as far as correlating a simple change in market prices or returns with a signal of an event. This naive approach, while beautiful on its simplicity, relies on the assumption that market volatility remains constant through the study period. However, this research assumption rarely depicts reality and the true functioning of markets. The so-called volatility clustering property [Mandelbrot, 1963], namely the tendency of large price changes to be followed by similarly large changes in price, of either sign, and its opposite functioning, is rather typical in market institutions [Cont, 2005]. This observation downsizes the applicability of pure time-series of prices as an input to text mining analysis; big price changes in periods of high volatility should not suggest signicant events. This concern is addressed by introducing and formulating a GARCH model for the variance of prediction markets prices, which is designed to deal with exactly this set of issues [Engle, 2001]. After that, the timing of news in conjunction with market volatility is

132CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

studied to arrive at conclusions about the relation of volatility and event clustering.

6.3 Background & Related Work


6.3.1 Text mining & stock market prediction

Financial markets are a native outlet for machine learning techniques. The interest in forecasting the future evolution of prices is strong and implicit, while traditionally such attempts were fueled by time series analysis, made solely on technical and fundamental past data. However, most relevant sources are naturally available in textual format, as news stories stream into the web in large volumes and a nearly real-time manner. The logical assumption that arises is that the study of the plethora of textual information sources might give much better predictions. The need to explore this assumption was recently addressed by the deployment of tools and techniques able to handle and automate the task of Knowledge Discovery from Textual sources (KDT), a process widely known as text mining. Karanikas and Theodoulidis [2002] use the term KDT to indicate the overall process of turning unstructured textual data into high level information and knowledge, while the term Text Mining is used for the step of the KDT process that deals with the extraction of patterns from textual data. By extending the denition of knowledge discovery in databases (KDD) given by Fayyad et al. [1996], Knowledge Discovery in Text (KDT) could be dened as the non-trivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in unstructured textual data. No matter of the actual denition, a rational attempt to implement such techniques for predicting the movements of stock prices would require a proven relationship among the release of relevant news stories and the uctuation of prices. Indeed, previous studies suggest that such an assumption is valid. Mitchell and Mulherin [1994] report that salient news has a positive impact on absolute price changes, Berry and Howe [1994] prove that there exists a signicant positive relationship between news arrivals and trading volume, while Chan [2001] further conrm the hypothesis, by proving that both salient political and economic news have impact on measures of trading activity, including return volatility, price volatility, number of shares traded, and trading frequency.

6.3. BACKGROUND & RELATED WORK

133

In light of both such ndings and the recent advances of the elds of KDD and KDT in particular, mining textual documents for stock market prediction stands as an emerging topic. Prior attempts include Wuthrich et al. [1998], Fawcett and Provost [1999], Lavrenko et al. [2000], Thomas and Sycara [2000], MacSkassy et al. [2001], Peramunetilleke and Wong [2002], Seo et al. [2002], Gidofalvi and Elkan [2003], Fung et al. [2002, 2005], Mittermayer [2004] and Schumaker and Chen [2006]. While all of these attempts are of rather explorative nature and by no means represent a standardized approach, their common ground is that they all imply pure stock prices in their reasoning and process tree. Among them, Wuthrich et al. [1998] provide predictions only for time instances on which at least 0.5% price changes are made, and Gidofalvi and Elkan [2003] scored news articles based on relative movement of the stock price during a selected window of inuence. In same direction, Fawcett and Provost [1999] label an event as interesting whether there is a 10% change in stock price which can be triggered by the content of relevant news stories, while Mittermayer [2004] correlates a number of categorization rules to news releases, according to categories associated to a specic impact on the stock prices. Moreover, MacSkassy et al. [2001] focus on one-hour stock returns and suggest labeling a story as important if the return of any stock associated with the story in the hour after the story appeared is greater than one standard deviation from the norm, and Lavrenko et al. [2000] propose the comining of stock time series and news stories, by introducing piecewise linear tting to the time series of pure prices. It clearly appears that so far text mining approaches for stock market prediction are essentially solely based on the study of simple changes in market prices, or -in few casesreturns. This fact raises a concern and a need for a more realistic treatment, as well as a strong research potential. Such concerns are addressed with a further analysis hereafter.

6.3.2

Volatility analysis

Time series of stock prices share some unique characteristics. Excess volatility, heavy tails, absence of autocorrelations in returns, volatility clustering and correlation of volatility with trading volume are rather typical phenomena appearing in market institutions [Cont, 2005] and further prove the distinctiveness and need for special treatment of such time

134CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

series. Out of these properties, volatility clustering is of specic essence in this case. Introduced and popularized by Mandelbrot [1963] who stated that large changes tend to be followed by large changes, of either sign, and small changes tend to be followed by small changes, this property actually downsizes the value of pure time series of stock prices or returns as input of text mining processes. In other words, big price changes in periods of high volatility shall not suggest signicant events; in any case, such signals should not be treated as a robust indicator for event detection or market prediction. A quantitative manifestation of this phenomenon is that, while returns themselves are uncorrelated, absolute returns or their squares display a positive, signicant and slowly decaying autocorrelation function. A simple display of daily log-returns would typically introduce the volatility clustering feature by highlighting sustained periods of high or low volatility. A plot of autocorrelation of returns would also introduce insignicant autocorrelation among various lags, while, by contrast, the autocorrelation function of absolute returns would initially remain positive and slowly decay to zero [Cont, 2005]. This latter observation remains remarkably stable across asset classes and time periods and is regarded as a typical manifestation of volatility clustering, while similar behavior is also observed for the autocorrelation of squared returns [Bollerslev et al., 1992]. The volatility clustering feature nally indicates that asset returns are not independent across time; on the other hand the absence of linear autocorrelation suggests that their dependence is nonlinear. This observation intrigued many researchers and nally lead to the introduction of the GARCH and stochastic volatility models by Engle [1995, 2001], designed to deal with just this set of issues. The simplest of such models, GARCH(1,1) is yt = c + t2 = + G1 t21 + A1 where , G1 and A1 are positive, G1 + A1 < 1 and
t t

2 t 1

(6.1)

iid(0, 1). Estimations of GARCH(1,1)

on stock and index returns usually yield G1 + A1 very close to 1 [Bollerslev et al., 1992]. For this reason the volatility clustering phenomenon is sometimes called a GARCH effect; one should keep in mind however that volatility clustering is a non-parametric

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

135

property and is not intrinsically linked to a GARCH specication [Cont, 2005]. After all, the GARCH(1,1) model arises as the simplest and most robust of the family of volatility models [Engle, 2001]. In the following a volatility analysis and GARCH modeling of prediction markets time series is introduced to perform as input of the text mining process and nally extract either alarms of important events or trends for market prediction. Afterwards, the timing of the news in conjunction with market volatility is studied to arrive at conclusions about the relation of volatility and event clustering.

6.4 Time Series Analysis of Prediction Markets


In what follows an attempt to prescribe techniques for the analysis and predictive modeling of prediction markets time series is provided, focusing on both price, variance and entropy movements prediction. The walkthrough is implemented through the use of actual market data, regarding political contracts on US presidential nominee elections1 .

6.4.1

Machine Learning Techniques

Predictive modeling using machine learning techniques is mostly based on trial and error, rather than being a concrete process. This experimental nature of the approach applies to all steps of the procedure, from the selection and formation of input data to the nal choice of the training algorithm. In this case, experimentation was needed at rst to decide on the input data to be fed into the machine learning algorithm to forecast the under study variable. Data available contained only the past time series of prices and volume for each contract, however this is not rather restrictive in terms of options, as the variety and usefulness of existing transformations in past time series data remains huge. After extended experimentation, the input variables that were nally used are the ones described in Table 6.1, while no further attribute selection were performed on this set of variables. For compliance purposes, the same input structure was utilized for all contracts. Moreover, a variety of target outputs was tried, in various combinations of the
1 acquired

from InTrade, http://www.intrade.com

136CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Table 6.1 SVMs Input Variables input variable logreturns percentiles past prices volume current price details 1, 2, 3, 4, 5, 10, 15, 30, 60 days 1, 2, 3, 4, 5, 10, 15, 30, 60 days 1, 2, 3, 4, 5, 10, 15, 30, 60 days last day last day

Table 6.2 SVMs Output Variable output form time span returns format details 1,3,7,15,30 days returns or logreturns normal or binary

characteristics documented in Table 6.2. All these were nally implemented for a bunch of prediction market stock time series, which are described in detail below. Among the plethora of available machine learning algorithms for the shaping of prediction models of a time series evolution, equally wide experimentation suggested the use of Support Vector Machines (SVM). SVM consist a recent advance in the Machine Learning eld, an algorithm proposed by Vapnik [1995] to address the two-class pattern recognition problem using the risk minimization principle. SVM are repeatedly reported to obtain very accurate results in text classication, while they outperform many other techniques such as neural networks and Naive Bayes [Joachims, 1998, Yang and Liu, 1999]. This latter observation makes attractive the use of the SVM algorithm for predictive modeling of time series, an attractiveness that seems to be conrmed in our case, too. The freely distributed machine learning software package Waikato Environment for Knowledge Analysis (WEKA; [Witten and Frank, 2005]) and its SVM classier were utilized. For the experiments performed, 7 political contracts of about 800 to 1000 time series instances each were performed, holding the last 30% of instances for validation and testing purposes. A sequential minimal optimization algorithm for training SVM

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

137

Table 6.3 SVM errors CC 0.9818 MAE 1.419 RMSE 2.1713 RAE 16.12% RRSE 17.38%

linear regression models on pre-normalized data [Smola and Scholkopf, 1998, Shevade et al., 1999] was applied. Its performance was then assessed by calculating the Correlation Coefcient, Mean Absolute Error, Root Mean Square Error, Relative Absolute Error and Root Relative Square Error on the non-overlapping and more recent data of the test set. Table 6.3 presents the results of running SVMs for the 1-day returns of the contract 2008.DEM.NOM.CLINTON. These numbers may introduce some rather satisfactory conclusions, at least in terms of the height of the computed errors. However, the closer inspection of the predicted time series versus the actual one may suggest another, less convenient, view. Figure 6.1 shows the results for the test set of the Clinton contract. Errors are low, but the graph spotlights a discouraging pattern: predictions just follow the previously known prices, in practice the predicted time series is very close to the actual one, but with a time lag equal to the predictions time window.

6.4.2
6.4.2.1

Garch analysis
Pre-estimation analysis

No matter how attractive SVM or other machine learning algorithms are, there exists no well-documented and universally acceptable way to validate that a predicted time series is equal or close enough to the best feasible one. To address this need and comparably assess the predictability of a time series, a switch into more classical statistical techniques was attempted. In the following, the prescriptions of Section 6.3.2 are leveraged to deploy a GARCH-based process for the predictive analysis of a nancial time series. A typical nancial time series GARCH analysis procedure is comprised of preestimation analysis, parameter estimation and post-estimation analysis. What follows is an extended walkthrough to the process using as an example the same contract described previously, regarding the democratic nomination for the 2008 elections. An econometric analysis typically starts (and nishes) with a plot. Plotting the actual

138CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Figure 6.1 SVM predictions vs actual results, Clinton.

time series could provide useful insights on the nature and patterns underlying the under study variable. However, it is typical for nancial time series a plot of returns or differences to provide a clearer view of the actual patterns. As a result, transforming prices into returns series is usually suggested. Next typical step regards checking for correlation in the original or transformed series. Figures 6.2, 6.3 & 6.4 present sample autocorrelation diagrams (ACF) of prices, differences and returns accordingly, along with the upper and lower standard deviation condence bounds. Equally useful are plots of partial sample autocorrelation function of the under study time series. Individual ACF values may present large variances, or also appear to be highly autocorrelated [Box et al., 1994]. Such phenomena are highlighted in the time series of actual prices, as for example Figure 6.2 introduces. While this observation does not undermine the value of such graphs as preliminary identication tools of the broad correlation characteristics of the returns, it clearly suggests that such deductions should be handled with care. The gures also suggest that, while the ACF of prices appears to be

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

139

Figure 6.2 Autocorrelation of prices series, Clinton, 2004Q4.

140CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Figure 6.3 Autocorrelation of differences series, Clinton, 2004Q4.

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

141

of considerable value and dying out slowly, ACFs of differences and returns are nearly identical and indicate a more stable pattern of small correlation.

Figure 6.4 Autocorrelation of returns series, Clinton, 2004Q4.

These latter graphs provide little evidence that any correlation structure in the conditional mean is actually required. However, the ACF of the squared returns may still indicate a signicant correlation and persistence in the second-order moments. Figure 6.5 provides the ACF plot of the squared returns and seems to support such a hypothesis. While returns themselves are nearly uncorrelated, the variance process exhibits some correlation, even at low extent, of a form that might suggest non-stationarity. The need for quantifying these correlations is clear and statistics come in support. The visual qualitative procedures described previously can be supplemented by formal hypothesis tests, such as the Ljung-Box-Pierce Q-test [Box and Pierce, 1970, Ljung and Box,

142CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Figure 6.5 Autocorrelation of squared returns series, Clinton, 2004Q4.

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

143

1978] and Engles ARCH test [Engle, 1982]. The Ljung-Box-Pierce test is performed to test whether the ACF of a time series signals a signicant departure from randomness or not. The null hypothesis for the test is that there exists no correlation and under that hypothesis the Q-test statistic is asymptotically 2 distributed [Box et al., 1994]. The Lbq-value is calculated by: Q k = T ( T + 2)
k

i =1

ri2 Ti

where T is the number of samples, k is the number of lags and ri the ith autocorrelation. If Qk is large then the probability that the process has uncorrelated data decreases. The null hypothesis for the test is that there exists no correlation and under that hypothesis, Qk is 2 with k degrees of freedom. Engles approach tests whether the residuals from an ordinary least squares regression exhibit conditional heteroskedasticity or not. The process begins with saving the residuals t of a regression, followed by a new regression of u t 2 on a constant and its own m-lagged u values. The occurring R2 value of the last regression is then asymptotically 2 -distributed t is i.i.d. N0, 2 ). Conrmation with m degrees of freedom under the null hypothesis that u of the null hypothesis implies that the time series is a random sequence of Gaussian disturbances, in other words no ARCH effect exists. Both tests normally return nearly identical outputs, so in the following we implement the Ljung-Box-Pierce test to quantify the suggestions of the previous visual analysis. We used MATLAB Garch toolbox [Inc., 2004] to perform this task and the results are presented in Tables 6.4 & 6.5 for various time lags (in days). H is a Boolean decision ag, where H = 0 suggests accepting the null hypothesis and H = 1 implies that signicant correlation exists, based on the 0.05 level of signicance. Ljung-Box-Pierce Q-test results seem to partially validate the visualizations prescriptions. Returns series present signicant correlation in the early lags (less than 10 days), as Figure 6.4 suggests, while such deductions are limited for the few and earliest of lags (not more than 5 days) regarding the squared returns series, a conclusion that may not seem perfectly coherent with Figure 6.5.

144CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Table 6.4 Ljung-Box-Pierce test on returns series, Clinton 2004Q4 lag 5 10 15 20 H 1.0000 1.0000 0 0 p-Value 0.0014 0.0275 0.1601 0.4249 Statistic 19.6930 20.1904 20.3169 20.5351 Critical Value 11.0705 18.3070 24.9958 31.4104

Table 6.5 Ljung-Box-Pierce test on square of returns series, Clinton 2004Q4 lag 5 10 15 20 H 1.0000 0 0 0 p-Value 0.0420 0.2740 0.6564 0.9018 Statistic 11.5206 12.1669 12.2978 12.3950 Critical Value 11.0705 18.3070 24.9958 31.4104

6.4.2.2

Parameter estimation

No matter what the time width is or the extent of correlation documented above, such results make the presence of heteroscedasticity clear and this in turn indicates the appropriateness of GARCH modeling. Time series (of returns) were eventually modeled by using the parsimonious GARCH(1,1) model, as described in Section 6.3.2. Parameters estimation is implemented using maximum likelihood estimation with the MATLABs maximization subroutine garcht. Running the algorithm for the under study contract returned a ARMAX (0,0,0) model for the mean and a GARCH(1,1) model for the variance, in compliance with the previous tests suggestions, while the values of model parameters in equation 6.1 are described in Table 6.6. By substituting these estimates in the denition of the default model, the estimation process implies that the constant conditional mean/ GARCH(1,1) conditional variance model that best ts the observed data is the following:

yt = 0.0030515 + t2 = 0.0010123 + 0.19403

2 t 1

(6.2)

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

145

Table 6.6 GARCH model estimated parameters, Clinton 2004Q4 Parameter C K GARCH(1) ARCH(1) Value -0.0030515 0.0010123 0 0.19403 Standard Error 0.015178 0.0010569 1.0595 0.31895 T-Statistic -0.2010 0.9579 0.0000 0.6083

It occurs that the model contains no moving average parameters for describing the variances evolution and is actually an ARCH model, as the estimated value of G1 parameter equals zero. 6.4.2.3 Post-estimation analysis

A GARCH (or ARCH) analysis needs to continue after model parameters estimation. In the following, some typical steps are sketched, continuing the analysis of the Clinton contract. In addition to the parameter estimates and standard errors, of value is the inspection of the residuals (innovations) and conditional standard deviations (sigmas), versus the observed returns. Figure 6.6 presents the time series and relationship occurred. While the sets width is very small to draw secure conclusions, it is noticeable that the series of innovations and returns present nearly identical behavior. The pointers from the last gure may suggest further examination. The plot of standardized innovations is clearly of value. In our case, Figure 6.7 validates the stability and absence of clustering phenomena, while Figure 6.8 further conrms the absence of correlation in the time series of squared standardized innovations. Of particular interest is also the comparison of Figure 6.5, which provides the ACF plot of the squared returns versus the ACF of the squared standardized innovations of Figure 6.8. In our case, this comparison comes to testify the sufciency of heteroscedasticity modeling of raw returns. As previously, a last step for conrming the absence of autocorrelation in the squared standardized innovations includes performing a Ljung-Box-Pierce test. Table 6.7 comes in

146CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Figure 6.6 Innovations, conditional standard deviations and returns series, Clinton, 2004Q4.

6.4. TIME SERIES ANALYSIS OF PREDICTION MARKETS

147

Figure 6.7 Standardized innovations of returns series, Clinton, 2004Q4.

148CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Figure 6.8 Autocorrelation of the squared standardized innovations of returns series, Clinton, 2004Q4.

6.5. VOLATILITY MODELING FOR DETECTING INFORMATION FLOWS

149

Table 6.7 Ljung-Box-Pierce test on square of standardized innovations series, Clinton 2004Q4 lag 5 10 15 20 H 0 0 0 0 p-Value 0.9996 1.0000 1.0000 1.0000 Statistic 0.1458 0.3282 0.3940 0.4244 Critical Value 11.0705 18.3070 24.9958 31.4104

support of the previous suggestions, as the Q-test indicates rejection ( H = 0 of the null hypothesis of correlation with signicance of near 100%, a result much improved over the ones of the preestimation analysis that further validates the explanatory power of the selected GARCH model. The time series analysis techniques described in this Section do not aim only at time series of prices or returns. Their predicting power could easily be extended in time series of variance or entropy and such applications and their accuracy in predicting future movements of various prediction markets contracts stand as an appealing direction for future study.

6.5 Volatility Modeling for Detecting Information Flows


Section 6.4 highlighted the applicability of a GARCH approach for the modeling of prediction market time series, specically the time series of returns. Moreover, evidence was provided on the existence of volatility clustering in such data, a result that conrms the necessity of volatility modeling before any effective event detection takes place. The current Section attempts to leverage on these outcomes, by contributing a process for modeling time series of volatility and nally utilizing the latter to perform news prediction and analysis tasks.

150CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

6.5.1

Volatility Modeling and Prediction

Volatility of a time series, usually of a nancial instrument, is typically dened and numerically computed as the sample standard deviation =
T 1 (r t )2 T 1 t =1

where rt is the return on day t, and is the average return over the T -day period. According to what was previously deployed, volatility does not follow a random walk across time, but exhibits different behavior in discrete chunks of time, an observation that seems to be replicated in prediction markets. Moreover, the GARCH(1,1) model proved to be a good solution in terms of explanatory power for capturing the dynamics of a prediction markets process, including the ones of volatility. The previous observations naturally lead to the proposal of a GARCH(1,1) model for the time series of volatility of prediction markets contracts. Such an approach seems to be compatible to both the theoretical concerns and the empirical results presented before, lending credence to a capable and accurate model, at least in comparison to existing alternatives. Moreover, this modeling approach offers the useful insights of separate estimation for the mean value and variance, which may prove to be convenient in this case. Given the selected model and available data, prediction is almost straight-forward, yet though computationally expensive. For example, given n days of data, the prediction for day n + 1 requires a GARCH model tted to the available data and then the use of this model for estimating the n + 1 price, but the prediction for n + 2 requires another model, using the n + 1 instances and so forth.

6.5.2

Analyzing Variance Movements

The time series of predicted variance now stands as an advanced and very competent input for event detection. The effect of volatility clustering is absorbed within the time series occurred, a fact that bypasses the problematic deployed in Section 6.3 about currently available techniques and gives further weight to the analysis that follows. However, at the same time, the materialized volatility could also be computed by the given data, which leads to the availability of two time series as valuable input signals for

6.6. EXPERIMENTAL SETTINGS

151

news analysis. The absolute difference of the expected minus the materialized volatility is a simple approach providing even deeper insight, as it counterbalances potential deviations and elevates unexpected events from obscurity. Thus, the analysis so far has contributed an improved way of isolating time instances of signicant information value, by computing the absolute difference of the expected -via a GARCH model- volatility minus the materialized volatility of a prediction market contract. The question that now arises concerns the potential relation of these signals to real events. This in turn highlights the need for selecting a time series being sufciently descriptive of the news stream relevant to a contracts description. The text mining process of all news streams available for the topics and keywords of relevance to the under study contract emerges as the optimal and almost ideal solution. However, more simplistic approaches may also be adequate to serve this target. The costudy of time series metrics of market volatility and news may nally lead to the following two categories of contributions, which differ according to which part of information is considered known. If the volume of relative news is increased, that may provide evidence on future market movements, while text in that case serves as a leading indicator and the contract volatility and price will change to adopt the news occurred (news analysis). On the other hand, if the volatility index alone reaches a peak, this stands as evidence on potentially signicant news emerging, and in that case text functions as a lagging indicator (news prediction). Both scenarios are of value and of various applications, for example detecting uncertainty reducing or inducing events may be helpful to market traders or decision makers. Combinatorial markets may also be used to investigate such events further.

6.6 Experimental Settings


While the previous Sections may proved to be theoretical enough, the ones to follow will bring back the scope of this Chapter to its empirical basis. To testify the applicability of the proposed approach, signicant efforts were put to collecting and analyzing actual prediction markets data and studying it in parallel with data descriptive of the relative news streams. The hereby Section describes the data and processes followed, while

152CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Section 6.7 deploys the results occurred. Data on 116 prediction markets contracts from InTrade were collected for the dates from 21-Oct-2004 till 28-Jul-2008. All these contracts were selected according to some properties and limitations, for example having over $1,000 as a total transaction volume in the under study period and this period being longer than a month. For each one of these contracts and each one of their instances, the procedure deployed in Section 6.5 was utilized, namely a GARCH(1,1) model was tted and then exploited for predicting the volatility of the following day. Also, the actual volatility for each day and contract was computed and nally the absolute difference of the expected minus the materialized volatility was known. Next, all the instances (days) of the latter time series and each contract were ranked to nd the dates of the biggest convergence of actual versus predicted volatility. A small number of dates were selected, nally contributing in a set of 276 days / contracts, for the given 116 contracts. For each instance of this set, a period of three days before the actual instances date until three days after that date was put into focus. For example, for the instance Feb 23, 2008 / Democratic Party Candidate to Win 2008 Presidential Election, the period from Feb 20 until Feb 26 was studied. The approach selected for gathering information about the relevant news streams was a rather naive one. More specically, for each contract one to ve keywords were selected (in the contract of the previous example, the keywords selected were democrat OR democrats OR democratic, win, president OR presidential and elections, and so on), while, for each set of keywords and date, the number of occurrences of the specic set of keywords in Google News Archive2 was tracked. This procedure nally resulted in a capable set of important volatility signals and relevant important news occurrences, providing the needed canvas for testifying the approach deployed in the previous Sections. Results of the procedure in the data set described are provided in the Section to follow.
2 http://www.news.google.com/newsarchive

6.7. EXPERIMENTAL RESULTS

153

Table 6.8 Example of hits in news archive found, typical contract date news hits -3 40 -2 64 -1 40 Feb 23, 2008 28 +1 28 +2 32 +3 43

Table 6.9 Metrics of average results for all contracts day hits diff z-score day hits diff z-score -3 12.86 1.58 0.069 +1 11.91 0.54 -0.042 -2 11.81 0.43 -0.037 +2 12.03 0.67 -0.039 -1 11.01 -0.44 0.022 +3 12.08 0.72 0.009 0 11.42 0 0.016

6.7 Experimental Results


Results collected were in the form of Table 6.8, with regards to the example mentioned before, for the instance Feb 23, 2008 / Democratic Party Candidate to Win 2008 Presidential Election. This specic example suggests the use of news as a leading indicator for market volatility, as the spike in news hits occurred two days before the day 0 of the spike in materialized volatilitys deviation from the expected one. However, a signicant number of instances demonstrated invert behavior, with the text and news following the spike of volatility. Such an analysis was performed for all nally available instances, with Table 6.8 presenting the average results for the actual number of hits, difference from day 0 and the relative z-score. It is also pointed that the averages presented occurred only from the contracts with a total number of hits bigger than 10 across the 7 days studied, as contracts with less than 10 hits were removed for accuracy reasons. The average contract and date instance studied tends to suggest a pattern of decreasing rate of relevant news the day before the error spike of expected volatility occurs, which is followed by a smaller but growing increase of relevant news occurring the days after the

154CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

spike.

6.8 Discussion
6.8.1 Summary

The latter Table 6.9 may leave the reader perplexed, or dissatised. In general, no clear and signicant enough trend seems to be revealed, about either the predictive or the analytical power of volatility modeling in relation with the stream of news. However, this tends to be the result for instances of both leading and lagging nature, with regards to the impact of text news in volatility, across the data set under focus. The proposed approach to volatility modeling essentially contributed the ability to point out and highlight such phenomena, of both types, with a signicantly improved condence rate, compared to the totality of the existing approaches. This Chapter presented a unique approach to important event detection. The problem seems to be trapped so far in the low informational value of pure nancial time series, which nally serves as a signal of poor quality for the under study purpose. An extension of focus was attempted into the yet almost untouched informational value of prediction markets that served as the input to the performed analysis. Moreover, focus was put into bypassing the problem of volatility clustering, while leveraging on the potential of the series of volatility. Advanced modeling techniques were put in use, in a process that was extensively showcased using a prediction market contract and resulted in a modeled time series that fully taps the informative potential of volatility. Further processes were proposed to essentially contribute accurate signals for important events detection, by comparing the expected versus the materialized volatility. The parallel consideration of these signals with proper and relevant streams of news is nally able to highlight signicant events, of either predictive or analytical to the markets prices nature. The whole approach was captured at its essence, while also deployed, analyzed and assessed using a rather extensive data set of prediction markets contracts, which came into support to the empirical nature of the study while shaping out various ways for further deployment.

6.8. DISCUSSION

155

6.8.2

Future work

The process presented remains far from complete. Actually, it essentially contributes little but paving the way to what the author hopes to become a viable alternative for event detection. A number of assumptions followed in this Chapter may be challenged, while the directions that are opened are practically endless, some of them going beyond a purely academic point of view. Among the various options for future study, the rst to mention is the rudimentary nature of the utilized metrics for news streams. A simple archive search for a set of keywords leaves much to be desired and the introduction of text mining approaches will improve signicantly the quality of the results; these remain an appealing area for further exploitation. Moreover, the data set studied may be extended to include more contracts, more instances and a wider time window to focus, while a classication of instances may provide some better insights on the patterns occured.

156CHAPTER 6. ADV ANCED PREDICTION MARKETS CONCEPTS: EVENT DETECTION

Chapter 7

Conclusions
7.1 Summary
The dissertation on hand has attempted to contribute in a number of research directions with regards to prediction markets. This section will provide a summary of the contributions made on each one of the research questions initially made in Table 1.1. After Chapter 1s problem statement and introduction, Chapter 2 discussed the conceptual framework of why and how prediction markets work, as well as the fundamental concepts of their operation. It then proceeded to review 155 research papers, namely the totality of journal articles, book chapters, conference proceedings and dissertations that were -to authors knowledge- until recently available. The breadth of this review enabled the study of prediction markets evolution across time and highlighted the almost exponential increase in publications volume. Moreover, a classication of literature was provided to categorize references in those related to a) prediction markets description, b) theoretical work, c) applications, d) law & policy and the subcategories thereof. This literature review and its classication aspires to serve as a contribution of value for beginners or experts diving in the topic of prediction markets. Chapter 3 stands as the core contribution of the Thesis. It began to describe the mechanisms being utilized across the various established market institutions in general. Then, it proceeded to study in detail the available mechanisms for the operation of prediction markets and highlight their various shortcomings. In response to the latter, a set 157

158

CHAPTER 7. CONCLUSIONS

of comprehensive mathematical properties for a coherent prediction market price function were introduced, providing a solid solution to the aforementioned issues. On top of that, a couple of functions satisfying these properties were presented and their compatibility to the suggested framework was validated. Essentially, this Chapter contributed what might be described as a coherent hybrid of dynamic pari-mutuel market and market scoring rules, the two dominant prediction market mechanisms so far. Chapter 4 discussed the theoretical evaluation of the proposed market mechanism. It started to describe the generic problem of information aggregation in market institutions and the theoretical background of its study. Then it proceeded to set a theoretical market framework, modeling the various aspects of information structure, market mechanism and trader behavior for the study and evaluation of the proposed market mechanism. Under the specied assumptions, the model provided evidence on the mechanisms capacity to converge to a consensus equilibrium with nite steps. However, it also demonstrated that the proposed market mechanism is not guaranteed to converge to the best possible prediction, namely to an equilibrium where all information available to traders becomes common knowledge across them. The deployed model and its application contribute a structured approach to the theoretical evaluation of a prediction market institution, along with demonstrating some capacities of the proposed mechanism. Chapter 5 deliberated about the empirical assessment of the proposed prediction market mechanism. A generic software platform, developed from scratch to rstly implement the suggested mechanism, was described. Then, a number of experiments performed in various academic, professional and social contexts were discussed. All experiments -no matter their variable performance- provided evidence on the capacity of markets to reach equilibrium, namely converge to suggest the actual forecasting target as the markets most probable outcome. Leveraging on the experience of the range of experiments facilitated, effort was then put into shaping a practical framework for prediction markets deployment, providing insights and recommendations for the setup and operation of prediction markets applications. Eventually, taking into account the rst real world implementation of the proposed market mechanism that was facilitated and discussed, this Chapter stands as an innitesimal contribution to the fast growing literature of prediction markets applications. Chapter 6 provided a look at a more advanced concept of prediction markets study.

7.2. FINAL REMARKS

159

More specically, the value of prediction markets as a tool for event detection was for the rst time investigated and discussed. A review of the state-of-the-art, with a focus on text mining and volatility modeling, highlighted the limitations of the approaches explored so far and the potential t of tapping into prediction markets as input for event detection. A GARCH model for prediction markets time series analysis was shaped to describe their underlying volatility. Following that, a structured approach for the analysis of information ows in parallel with market volatility was suggested and analyzed, eventually enabling event detection. Experimental evidence on the practical applicability of the proposed approach was also provided with an extended data set. After all, the Chapter contributes a novel approach to important event detection, highlighting its potential performance while suggesting new directions for event detection research and practice.

7.2 Final Remarks


There are two kinds of forecasts, goes the popular saying, those which are wrong, and those which are lucky. Probably for a reason and due to experts fault, there is a growing disconnect between forecasts and trust, across practitioners and the society at large. The more complex or dark black-box forecasting approaches get, the less people count on them, eventually replacing algorithms with judgments when making a decision. This might sound naive at rst, however nature is after all governed by a few simple rules and the more one distances herself from such facts, the less things tend to work out. This Thesis attempted an 180 degree turn from the dominant narrative of increasing computational complexity during the last decades. It regressed towards a simpler, transparent and arguably more elegant, yet increasingly sophisticated approach to forecasting. Markets are institutions that organically work, while by nature performing the core forecasting functions of information discovery, aggregation and processing when operating in an endless variety of contexts. The vision of contributing towards markets coming of age as a primary forecasting tool fueled the delivery of this work. Extensive effort was put in studying the concept of prediction markets in depth and width, from literature to theory to experiments to practice. A number of optimistic outcomes were met, drawing in bright colors forecastings divine promise. Yet at the

160

CHAPTER 7. CONCLUSIONS

same time, as sciences knowledge tree never grows old, for each answer provided, more questions impatiently emerge. And ironically, should this Thesis innitesimally serve the literature it helped to emerge, only future will tell.

Bibliography
Michael Abramovicz. The law-and-markets movement. American University Law Review, 49:327431, 1999. Michael Abramovicz. Trial by market: A thought experiment, 2004. Michael Abramovicz. Predictive decisionmaking. Virginia Law Review, 92(1):69122, 2006a. article George Predictive Decisionmaking. Michael Abramovicz. Deliberative information markets for small groups. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 101125. AEI Press, Washington D.C., 2006b. Michael B. Abramowicz. Information markets, administrative decisionmaking, and predictive cost-benet analysis. University of Chicago Law Review, 71:9331020, 2004. A. R. Admati and P. Peiderer. Viable allocations of information in nancial markets. Journal of Economic Theory, 43:76115, 1987. Shipra Agrawal, Erick Delage, Mark Peters, Zizhuo Wang, and Yinyu Ye. A unied framework for dynamic prediction market design. Operations research, 59(3):550568, 2011. M. Ali. Probability and utility estimates for racetrack bettors. Journal of Political Economy, 85:803815, 1977. James Allan. Introduction to Topic Detection and Tracking, chapter 1, pages 116. Kluwer Academic Publishers, Massachusetts, 2002a. 161

162

BIBLIOGRAPHY

James Allan. Topic Detection and Tracking: Event-Based Information Organization. Kluwer Academic Publishers, Norwell, MA, 2002b. James Allan, Jaime Carbonell, George Doddington, Jonathan Yamron, and Yang Yiming. Topic detection and tracking pilot study: Final report. In Proceedings of DARPA Broadcast News Transcription and Understanding Workshop, pages 194218, 1998. Beth E. Allen. Generic existence of completely revealing equilibria for economies with uncertainty when prices convey information. Econometrica, 49(5):11731199, 1981. Beth E. Allen. Strict rational expectation equilibria with diffuseness. Journal of Economic Theory, 27(20-46), 1982. Morris Altman. The nobel prize in behavioral and experimental economics: a contextual and critical appraisal of the contributions of daniel kahneman and vernon smith. Review of Political Economy, 16(1):341, 2004. L. Anderson and C. Holt. Information cascades in the laboratory. American Economic Review, 87:847862, 1997. Bernd H. Ankenbrand and Caroline Rudzinski. Description analysis of markets. Technical Report techreport George Description & Analysis of Markets. Bernd H. Ankenbrand and Caroline Rudzinski. Taxonomy for information markets, 2005b. Witten/Herdecke University. Werner Antweiler and Thomas W. Ross. The 1997 ubc election stock market. Canadian Business Economics, 6:1522, 1998. Aristotle. Politics. translated by E. Barker. Oxford University Press, London, 1972. Scott J. Armstrong. Forecasting with econometric methods: Folklore versus fact. The Journal of Business, 51(4):549564, 1978. Scott J. Armstrong. Principles of Forecasting. Kluwer Academic Publishers, Dordrecht, The Netherlands, 2001.

BIBLIOGRAPHY

163

Scott J. Armstrong. Findings from evidence-based forecasting: Methods for reducing forecast error. International Journal of Forecasting, 22:583598, 2006. Kenneth J. Arrow. The role of securities in the optimal allocation of risk-bearing. Review of Economic Studies, 31(2):9196, 1964. Alison Hubbard Ashton and Robert H. Ashton. Aggregating subjective forecasts: Some empirical results. Management Science, 31(12):14991508, 1985. Robert J. Aumann. Agreeing to disagree. The Annals of Statistics, 4(6):12361239, 1976. Martin Barner, Francesco Feri, and Charles R. Plott. On the microstructure of price determination and information aggregation with sequential and asymmetric information arrival in an experimental asset market, 2004. Roy Batchelor and Pami Dua. Forecaster diversity and the benets of combining forecasts. Management Science, 41(1):6875, 1995. Glenn Bean. Decision market predictions: Separating efciency and bias, 2005. Klaus Beckmann and Martin Werding. passauer wahlborse: Information processing in a political market experiment. Kyklos, 49:171205, 1996. Tom W. Bell. Gambling for the good, trading for the future: The legality of markets in science claims. Chapman Law Review, 5:159180, 2002. Tom W. Bell. The prediction exchange: Progress in promoting the sciences and useful arts, 2006. Joyce Berg, Robert Forsythe, and Thomas A. Rietz. What makes markets predict well? evidence from the iowa electronic markets, 1996. Joyce Berg, Robert Forsythe, Forrest Nelson, and Thomas Rietz. Results from a dozen years of election futures market research, 2000. Joyce Berg, Forrest Nelson, and Thomas Rietz. Accuracy and forecast standard error of prediction markets, 2003.

164

BIBLIOGRAPHY

Joyce E. Berg and Thomas A. Rietz. Longshots, overcondence and efciency on the iowa electronic market, 2002. Joyce E. Berg and Thomas A. Rietz. Prediction markets as decision support systems. Information Systems Frontiers, 5(1):7993, 2003. Joyce E. Berg and Thomas A. Rietz. The iowa electronic markets: Stylized facts and open issues. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 142169. AEI Press, Washington D.C., 2006. incollection George The Iowa Electronic Markets: Stylized Facts and Open Issues. Joyce E. Berg, Robert Forsythe, and Thomas A. Rietz. The iowa electronic market. In Dean Paxson and Douglas Wood, editors, Blackwell Encyclopedic Dictionary of Finance, pages 111113. Blackwell, Oxford UK, 1997. Ole Jakob Bergfjord. Prediction markets as a tool for management of political risk, 2006. Michael Berlemann and Nelson Forrest. Forecasting ination via electronic markets. results from a prototype market, 2002. Michael Berlemann and Carsten Schmidt. Predictive accuracy of political stock markets: Empirical evidence from a european perspective, 2001. T.D. Berry and K.M. Howe. Public information arrival. Journal of Finance, 49(4):13311346, 1994. Peter Bohm and Joakim Sonnegard. Political stock markets and unreliable polls. Scandinavian Journal of Economics, 101(2):205222, 1999. Tim Bollerslev, Ray Y. Chou, and Kenneth F. Kroner. Arch modeling in nance: A review of the theory and empirical evidence. Journal of Econometrics, 52(1-2):559, 1992. Oleg Bondarenko and Peter Bossaerts. Expectations and learning in iowa. Journal of Banking and Finance, 24:15351555, 2000. Efthimios Bothos. Social Computing with Information Markets. PhD thesis, School of Electrical Engineering, National Technical University of Athens, 2010.

BIBLIOGRAPHY

165

Efthimios Bothos, Konstantinos Christidis, Dimitris Apostolou, and Gregoris Mentzas. Information market based recommender systems fusion. In 2nd International Workshop on Information Heterogeneity and Fusion in Recommender Systems (HetRec 2011), 5th ACM Conference on Recommender Systems, Chicago, IL, USA, October 27 . Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Evaluating new ideas with prediction markets. In 22nd European Conference on Operational Research, EURO 2007, KM workshop - Applications, Methods and Practices, Stream: Knowledge Management, 2007. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Idea selection and information aggregation markets. In IEEE International Engineering Mnagement Conference, Estoril, Portugal, June 28-30 2008a. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. A collaborative information aggregation system for idea management. In ICIW, editor, Proceedings of the Third International Conference on Internet and Web Applications and Services, pages 289296, 2008b. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Idem: A prediction market for idea management. In Stefan Luckner Christof Weinhardt and Jochen Stober, editors, Designing E-Business Systems. Markets, Services, and Networks, volume 22, pages 113. ISSN 1865-1348, 2008c. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Collaborative idea management with an information aggregation system. In O. Cunningham and M. Cunningham, editors, Collaboration and the Knowledge Economy: issues, Applications and Case Studies, eChallenges e-2008 Conference, pages 691698, Stockholm, Sweden, 22-24 October 2008d. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Collective intelligence for idea management with internet-based information aggregation markets. Internet Research, 19(1):2641, 2009. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Agent based information

166

BIBLIOGRAPHY

aggregation. In AAMAS-2010, 9th International Conference on Autonomous Agents and Multiagent Systems, Toronto, Canada, May 10-14 2010a. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Using social media to predict future events with agent-based markets. IEEE Intelligent Systems, pages 5058, November-December 2010b. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Citizen engagement with information aggregation markets. In A. Macintosh E. Tambouris and H. de Bruijn, editors, ePart 2011, LNCS 6847, pages 274285, 2011. Efthimios Bothos, Dimitris Apostolou, and Gregoris Mentzas. Collective intelligence with web-based information aggregation markets: The role of market facilitation in idea management. Expert Systems with Applications, 39(1):13331345, 2012. Brian L. Boulier and H. O. Stekler. Predicting the outcomes of national football league games. International Journal of Forecasting, 19(2):257270, 2003. G.E.P. Box and D. A. Pierce. Distribution of residual autocorrelation in autoregressive integrated moving average time series models. Journal of the American Statistical Association, 65(1509-1526), 1970. G.E.P. Box, G.M. Jenkins, and G.C. Reinsel. Time Series Analysis: Forecasting and Control. Prentice Hall, third edition edition, 1994. Glenn Boyle and Steen Videbeck. A primer on information markets, 2005. Aseem Brahma, Mithun Chakraborty, Sanmay Das, Allen Lavoie, and Malik MagdonIsmail. A bayesian market maker. In Proceedings of the 13th ACM Conference on Electronic Commerce, pages 215232. ACM, 2012. David Bray, Karen Croxson, and William Dutton. Information markets: Feasibility and performance. 2008. Peter J. Brockwell and Richard A. Davis. Introduction to time series and forecasting. Springer, 1996.

BIBLIOGRAPHY

167

Georg Brueggelambert. Information and efciency in political stock markets: using computerized markets to predict election results. Applied Economics, 36:753768, 2004. Georg Brueggelambert and Arwed Crueger. Election markets: Experiences from a complex market experiment. In Friedel Bolle and Marco Lehmann-Waffenschmidt, editors, Surveys in Experimental Economics: Bargaining, Cooperation and Election Stock Markets, pages 167191. Physica-Verlag, Heidelberg, 2002. Ryan Bubb. From uncertainty to risk? problems with using prices in prediction markets to improve decision making, 2005. C. Camerer and Weigelt K. Information mirages in experimental asset markets. Journal of Business, 64:463493, 1991. Nicholas Chan, Ely Dahan, Adlar Kim, Andrew Lo, and Tomaso Poggio. Securities trading of concepts, 2002. Nicholas T. Chan, Blake LeBaron, Andrew W. Lo, and Tomaso Poggio. Agent-based models of nancial markets: A comparison with experimental markets. Technical Report Paper No.124, MIT, 1999. Nicholas T. Chan, Ely Dahan, Andrew W. Lo, and Tomaso Poggio. Experimental markets for product concepts, 2001. T. Chan. Articial Markets and Intelligent Agents. PhD thesis, MIT, 2001. PhD Thesis. Kay-Yut Chen, Leslie R. Fine, and Bernardo A. Huberman. Forecasting uncertain events with small groups. In 3rd ACM conference on Electronic Commerce, pages 5864, Tampa, Florida, USA, 2001. Kay-Yut Chen, Leslie R. Fine, and Bernardo A. Huberman. Predicting the future. Information Systems Frontiers, 5(1):4761, 2003. Kay-Yut Chen, Leslie R. Fine, and Bernardo A. Huberman. Eliminating public knowledge biases in information-aggregation mechanisms. Management Science, 50(7):983994, 2004. Qi Chen, Itay Goldstein, and Wei Jiang. Price informativeness and investment sensitivity to stock price, May 1, 2007 2007.

168

BIBLIOGRAPHY

Yiling Chen. Markets as an Information Aggregation Mechanism for Decision Support. PhD thesis, The Pennsylvania State University, 2005. Yiling Chen and David M Pennock. Designing markets for prediction. AI Magazine, 31(4): 4252, 2010. Yiling Chen and Jennifer Wortman Vaughan. A new understanding of prediction markets via no-regret learning. In Proceedings of the 11th ACM conference on Electronic commerce, pages 189198. ACM, 2010. Yiling Chen, Chao-Hsien Chu, Tracy Mullen, and David M. Pennock. Information markets vs. opinion pools: an empirical comparison. In 6th ACM conference on Electronic commerce, Vancouver BC, Canada, 2005. ACM Press. Yiling Chen, Chao-Hsien Chu, and Tracy Mullen. Predicting uncertain outcomes using information markets: Trader behavior and information aggregation. New Mathematics and Natural Computation, 2(3):281297, 2006a. Yiling Chen, Tracy Mullen, and Chao-Hsien Chu. An in-depth analysis of information markets with aggregate uncertainty. Electronic Commerce Research, 6(2):201221, 2006b. Miriam A. Cherry and Robert L. Rogers. Markets for markets: Origins and subjects of information markets. Rutgers Law Review, 58(2):339, 2006. Martin Cihak. The 2007 nobel prize in economics: Mechanism design theory. Czech Journal of Economics and Finance, 58(1-2):8289, 2008. R. T. Clemen. Incentive contracts and strictly proper scoring rules. Test, 11:195217, 2002. R. Cont. Long Memory in Economics, chapter Volatility clustering in nancial markets: empirical facts and agent-based models. Springer, Berlin, 2005. Thomas Copeland and D. Friedman. The effect of sequential information arrival on asset prices: An experimental study. Journal of Finance, 42:763797, 1987. Thomas Copeland and D. Friedman. Partial revelation of information in experimental asset markets. Journal of Finance, 46:265295, 1991.

BIBLIOGRAPHY

169

Thomas Copeland and D. Friedman. The market value of information: Some experimental results. Journal of Business, 65:241266, 1992. Ely Dahan and John R. Hauser. The virtual customer. Journal of Product Innovation Management, 19(5):332353, 2002. Sandip Debnath, David M. Pennock, C. Lee Giles, and Steve Lawrence. Information incorporation in online in-game sports betting markets. In 4th ACM conference on Electronic commerce, pages 258259, San Diego, CA, USA, 2003. ACM Press. Cary Deck and David Porter. Prediction markets in the laboratory. Journal of Economic Surveys, 2013. Pradeep Dubey, John Geanakoplos, and Martin Shubik. The revelation of information in strategic market games: A critique of rational expectations equilibrium. Journal of Mathematical Economics, 16:3105137, 1987a. Pradeep Dubey, John Geanakoplos, and Martin Shubik. The revelation of information in strategic market games : A critique of rational expectations equilibrium. Journal of Mathematical Economics, 16(2):105137, 1987b. Nobel E-Museum. The bank of sweden prize in economic sciences in memory of alfred nobel 2002. http://www.nobel.se/economics/laureates/, 2003. Matthew Einbinder. Information markets: Using market predictions to make todays decisions. Virginia Law Review, 92(1):149186, 2006. Edmund Eisenberg and David Gale. Consensus of subjective probabilities: The pari-mutuel method. Annals of Mathematical Statistics, 30(165-168), 1959. R. Engle. ARCH models. Oxford University Press, Oxford, 1995. Robert Engle. Garch 101: The use of arch/garch models in applied econometrics. Journal of Economic Perspectives, 15(4):157168, 2001. Robert F. Engle. Autoregressive conditional heteroskedasticity with estimates of the variance of united kingdom ination. Econometrica, 50:19872007, 1982.

170

BIBLIOGRAPHY

Robert S. Erikson and Christopher Wlezien. Are political markets really superior to polls as election predictors?, April 5 2006. Ronald Fagin, Yoram Halpern, Joseph Y.and Moses, and Mosche Y. Vardi. Reasoning About Knowledge. MIT Press, Cambridge, MA, 1996. Eugene F. Fama. Efcient capital market: A review of theory and empirical work. Journal of Finance, 25(383-417), 1970. Fang Fang, Maxwell Stinchcombe, and Andrew Whinston. putting your money where your mouth is - information elicitation mechanism design when agents privately know their quality of information, 2005. T. Fawcett and F. Provost. Activity monitoring: Noticing interesting changes in behavior. In S. Chaudhuri, D. Madigan, and U. Fayyad, editors, Proceedings of the 5th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 5362, San Diego, California, August 15-18 1999. ACM Press. U. M. Fayyad, G. Piatetsky-Shapiro, and P. Smyth. Advances in Knowledge Discovery and Data Mining., chapter From Data Mining to Knowledge Discovery: An Overview, pages 134. AAAI / MIT Press, Cambridge, Massachusetts, 1996. Joan Feigenbaum, Lance Fortnow, David M. Pennock, and Rahul Sami. Computation in a distributed information market. Theoretical Computer Science, 343(1-2):114132, 2005. Stephen Figlewski. Subjective information and market efciency in a betting market. The Journal of Political Economy, 87(1):7588, 1979. Robert Fildes. Quantitative forecasting-the state of the art: Econometric models. The Journal of the Operational Research Society, 36(7):549580, 1985. Peter Filzmaier, Maria Beyrl, Florian Hauser, and Juergen Huber. Wahlboersen als interdisziplinaeres instrument der sozialforschung - erfahrungen aus wahlen in oesterreich. SWS Rundschau, 3/2003:387410, 2003. Nelson Forrest and R. Lundholm. Information aggregation in an experimental market. Econometrica, 58(309-347), 1990.

BIBLIOGRAPHY

171

Robert Forsythe and R. Lundholm. Information aggregation in an experimental market. Econometrica, 58:309347, 1990. Robert Forsythe, T. R. Palfrey, and C. Plott. Asset valuation in an experimental market. Econometrica, 50:537567, 1982. Robert Forsythe, Forrest Nelson, George R. Neumann, and Jack Wright. Anatomy of an experimental political stock market. American Economic Review, 82(5):11421161, 1992. Robert Forsythe, Forrest Nelson, George Neumann, and Jack Wright. The 1992 iowa political stock market: September forecasts. The Political Methodologist, 5:1519, 1994. Robert Forsythe, Murray Frank, Vasu Krishnamurthy, and Thomas W Ross. Using market prices to predict election results: the 1993 ubc election stock market. Canadian Journal of Economics, 28(4a):770794, 1995. Robert Forsythe, Murray Frank, Vasu Krishnamurthy, and Thomas W Ross. Markets as predictors of election outcomes: Campaign events and judgement bias in the 1993 ubc election stock market. Canadian Public Policy, 24(3):329351, 1998. Robert Forsythe, Thomas A. Rietz, and Thomas W. Ross. Wishes, expectations and actions: a survey on price formation in election stock markets. Journal of Economic Behavior Organization, 39:83110, 1999. Lance Fortnow, Joe Kilian, David M. Pennock, and Michael P. Wellman. Betting booleanstyle: A framework for trading in securities based on logical formulas. Decision Support Systems, 39(1):87104, 2004. James H. Fowler. Elections and markets: The effect of partisanship, policy risk, and electoral margins on the economy. The Journal of Politics, 68(1):89103, 2006. article George Elections and Markets: The Effect of Partisanship, Policy Risk, and Electoral Margins on the Economy. Simon French. Group consensus probability distributions: a critical survey. Bayesian Statistics, 2:183202, 1985.

172

BIBLIOGRAPHY

Daniel Friedman. How trading institutions affect nancial market performance: Some laboratory evidence. Economic Inquiry, 31:410435, 1993. G. P. C. Fung, J. X. Yu, and W. Lam. News sensitive stock trend prediction. In M.S. Chen, P.S. Yu, and Liu B., editors, Proceedings of the 6th Pacic-Asia Conference (PAKDD) on Advances in Knowledge Discovery and Data Mining, Lecture Notes in Computer Science, volume 2336, pages 481493, Taipei, Taiwan, May 06-08 2002. Springer-Verlag. G. P. C. Fung, J. X. Yu, and H. Lu. The predicting power of textual information on nancial markets. IEEE Intelligent Informatics Bulletin, 5(1):110, 2005. John M. Gandar, William H. Dare, Craig R. Brown, and Richard A. Zuber. Informed traders and price variations in the betting market for professional basketball games. The Journal of Finance, 53(1):385401, 1998. Xi Gao, Yiling Chen, and David M Pennock. Betting on the real line. In Internet and Network Economics, pages 553560. Springer, 2009. Xi Alice Gao and Yiling Chen. An axiomatic characterization of continuous-outcome market makers. In Internet and Network Economics, pages 505514. Springer, 2010. John Geanakoplos and Heracles M. Polemarchakis. We cant disagree forever. Journal of Economic Theory, 28(1):192200, 1982. Christian Genest and James V. Zidek. Combining probability distributions: A critique and an annotated bibliography. Statistical Science, 1(1):114148, 1986. G. Gidofalvi and C. Elkan. Using news articles to predict stock price movements. Technical report, Department of Computer Science and Engineering, University of California, San Diego, 2003. Steven Gjerstad. Risk aversion, beliefs, and prediction market equilibrium, 2005. Andreas Graefe and J Scott Armstrong. Comparing face-to-face meetings, nominal groups, delphi and prediction markets on an estimation task. International Journal of Forecasting, 27(1):183195, 2011.

BIBLIOGRAPHY

173

Kesten C Green and J Scott Armstrong. Demand forecastingevidence-based methods. 2012. S. J. Grossman. An introduction to the theory of rational expectations under asymmetric information. Review of Economic Studies, 48(4):541559, 1981. Sanford J. Grossman and Joseph E. Stiglitz. On the impossibility of informationally efcient markets. The American Economic Review, 70(3):393408, 1980a. Sanford J. Grossman and Joseph E. Stiglitz. On the impossibility of informationally efcient markets. The American Economic Review, 70(3):393408, 1980b. Thomas S. Gruca. The iem movie box ofce market: Integrating marketing and nance using electronic markets. Journal of Marketing Education, 22(1):514, 2000. Thomas S. Gruca, Joyce Berg, and Michael Cipriano. Do electronic markets improve forecasts of new product success?, 2001. Thomas S. Gruca, Joyce Berg, and Michael Cipriano. The effect of electronic markets on forecasts of new product success. Information Systems Frontiers, 5(1):95105, 2003. Thomas S. Gruca, Joyce E. Berg, and Michael Cipriano. Consensus and differences of opinion in electronic prediction markets. Electronic Markets, 15(1):1322, 2005. Refet S. Gurkaynak and Justin Wolfers. Macroeconomic derivatives: An initial analysis of market-based macro forecasts, uncertainty and risk. In Christopher Pissarides and Jeffrey Frankel, editors, NBER International Seminar on Macroeconomics. MIT Press, Cambridge, 2006. Robert W. Hahn and Paul C. Tetlock. Harnessing the power of information: A new approach to economic development, 2004. Robert W. Hahn and Paul C. Tetlock. Big ideas: The markets last frontier. Milken Institute Review, 7(1):8389, 2005a. article George http://ssrn.com/abstract=821450. Robert W. Hahn and Paul C. Tetlock. Using information markets to improve public decision making. Harvard Journal of Law Public Policy, 29(1):214289, 2005b. article George Using Information Markets to Improve Public Decision Making.

174

BIBLIOGRAPHY

Robert W. Hahn and Paul C. Tetlock. Making development work. Policy Review, 132:2739, 2005c. Robert W. Hahn and Paul C. Tetlock. Introduction to information markets. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 112. AEI Press, Washington D.C., 2006a. incollection George Introduction to Information Markets. Robert W. Hahn and Paul C. Tetlock. A new approach for regulating informa-

tion markets. Journal of Regulatory Economics, 29(3):265281, 2006b. article George http://dx.doi.org/10.1007/s11149-006-7399-z. Jan Hansen, Carsten Schmidt, and Martin Strobel. Manipulation in political stock markets - preconditions and evidence. Applied Economics Letters, 11(7):459463, 2004. R. Hanson. Consensus by identifying extremists. Theory and Decision, 44(3):293301, 1998. Robin Hanson. Could gambling save science? encouraging an honest consensus. In Eight International Conference on Risk and Gambling, London, 1990a. Robin Hanson. Market-based foresight: a proposal. Foresight Update, 10:14, 1990b. Robin Hanson. More market-based foresight. Foresight Update, 11:11, 1991. Robin Hanson. Idea futures: Encouraging an honest consensus. Extropy, 3(2), 1992. Robin Hanson. Disagreement is unpredictable. Economic Letters, 77:365369, 2002a. Robin Hanson. Logarithmic market scoring rules for modular combinatorial information aggregation, 2002b. Robin Hanson. Combinatorial information market design. Information Systems Frontiers, 5 (1):107119, 2003a. Robin Hanson. Shall we vote on values, but bet on beliefs?, 2003b. Robin Hanson. Designing real terrorism futures, 2005.

BIBLIOGRAPHY

175

Robin Hanson. Foul play in information markets. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 126141. AEI Press, Washington D.C., 2006a. incollection George Foul Play in Information Markets. Robin Hanson. Impolite innovation: The technology and politics of terrorism futures and other decision markets. In Eric Patashnik and Alan Gerber, editors, Promoting the General Welfare: American Democracy and the Political Economy of Government Performance. Brookings Institution Press, Washington D.C., 2006b. incollection George Impolite Innovation: The Technology and Politics of Terrorism Futures and Other Decision Markets. Robin Hanson. Decision markets for policy advice. In Eric Patashnik and Alan Gerber, editors, Promoting the General Welfare: American Democracy and the Political Economy of Government Performance. Brookings Institution Press, Washington D.C., 2006c. incollection George Decision Markets for Policy Advice. Robin Hanson and Ryan Oprea. Manipulators increase information market accuracy, 2004. Robin Hanson, Ryan Oprea, and David Porter. Information aggregation and manipulation in an experimental market. Journal of Economic Behavior and Organization, 60(4):449459, 2006. Florian Hauser. Die presse online-wahlboerse 2002 - eine nanzwirtschaftliche betrachtung. Masters thesis. Florian Hauser and Juergen Huber. Systematic mispricing in experimental markets evidence from political stock markets. In International Conference on Finance, Kopenhagen, Denmark, 2005. Friedrich Hayek. The use of knowledge in society. American Economic Review, 35:519530, 1945. Martin F. Hellwig. On the aggregation of information in competitive markets. Journal of Economic Theory, 22(3):477498, 1980.

176

BIBLIOGRAPHY

T. Hill, L. Marquez, M. OConnor, and W. Remus. Articial neural network models for forecasting and decision making. International Journal of Forecasting, 10:515, 1994. Bengt Holmstrom and Jean Tirole. Market liquidity and performance monitoring. The Journal of Political Economy, 101(4):678709, 1993. MathWorks Inc. GARCH Toolbox Users Guide. MATLAB, version 6.5 edition, 2004. M. Jackson and J. Peck. Asymmetric information in a strategic market game: reexamining the implications of rational expectations. Economic Theory, 13(603-628), 1999. Jens Carsten Jackwerth and Mark Rubinstein. Recovering probability distributions from option prices. The Journal of Finance, 51(5):16111631, 1996a. Jens Carsten Jackwerth and Mark Rubinstein. Recovering probability distributions from option prices. The Journal of Finance, 51(5):16111631, 1996b. Ben Jacobsen, Jan Potters, Arthur Schram, Frans van Winden, and Joergen Wit. (in)accuracy of a european political stock market: The inuence of common value structures. European Economic Review, 44:205230, 2000. Jeevan Jaisingh, Shailendra Mehta, and Alok Chaturvedi. An experimental study of information markets. In The 7th INFORMS Conference on Information Systems and Technology, San Jose, California, USA, 2002. Richard C. Jebb. Sophocles: the plays and fragments, part VII: the Ajax. University Press, 1896. T. Joachims. Text categorization with support vector machines: Learning with many relevant features. In The European Conference on Machine Learning. Springer, 1998. James Jordan and Roy Radner. The nonexistence of rational expectations equilibrium: A robust example. Department of Economics, University of Minnesota, 1979. J. H. Kagel and A. E. Roth, editors. Handbook of Experimental Economics. Princeton University Press, 1995. Ajit Kambil and Eric van Heck. Making Markets: How Firms Can Design and Prot from Online Auctions and Exchanges. Harvard Business School Press, Boston, MA, 2002.

BIBLIOGRAPHY

177

H. Karanikas and B. Theodoulidis. Knowledge discovery in text and text mining software. Centre for Research in Information Management (CRIM), UMIST University, UK., 2002. Eiichiro Kazumori. Foundations of information aggregation mechanisms. In 2nd Workshop on the Economics of Peer-To-Peer Systems, Harvard University, 2004. Frederic Koessler, Charles Noussair, and Anthony Ziegelmeyer. Individual behavior and beliefs in experimental parimutuel betting markets, 2005. S. G. Kou and Michael E. Sobel. Forecasting the vote: A theoretical comparison of election markets and public opinion polls. Political Analysis, 12:277295, 2004. Joel T. Krueger and Kenneth N. Kuttner. The fed funds futures rate as a predictor of federal reserve policy. Journal of Futures Markets, 16(8):865879, 1996. Bettina Kuon. Typical trading behavior in the german election markets 1990, 1991. Victor Lavrenko, Matt Schmill, Dawn Lawrie, Paul Ogilvie, David Jensen, and James Allan. Mining of concurrent text and time series. In Proceedings of the 6th ACM SIGKDD Intermational Conference on Knowledge Discovery and Data Mining, Workshop on Text Mining, pages 3744, Boston, MA, August 20-23 2000. ACM Press. Michael Lawrence, Paul Goodwin, Marcus OConnor, and Dilek Onkal. Judgmental forecasting: A review of progress over the last 25 years. International Journal of Forecasting, 22(3):493518, 2006a. Michael Lawrence, Paul Goodwin, Marcus OConnor, and Dilek Onkal. Judgmental forecasting: A review of progress over the last 25years. International Journal of Forecasting, 22(3):493518, 2006b. John Ledyard, Robin Hanson, and Takashi Ishikida. An experimental test of combinatorial information markets, 2005. John O. Ledyard. Designing information markets for policy analysis. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 3766. AEI Press, Washington D.C., 2006.

178

BIBLIOGRAPHY

Andrew Leigh and Justin Wolfers. Competing approaches to forecasting elections: Economic models, opinion polling and prediction markets, 2006. G. M. Ljung and G.E.P. Box. On a measure of lack of t in time series models. Biometrika, 66:297303, 1978. Andrew W. C. Lo. Market efciency: stock market behaviour in theory and practice. Edward Elgar, London, 1997. R. E. Lucas. Expectations and the neutrality of money. Journal of Economic Theory, 4(2): 103124, 1972. R. Lundholm. What affects the efciency of the market? some answers from the laboratory. The Accounting Review, 66:486515, 1991. H. Lutkepohl. Introduction to Multiple Times Series Analysis. Springer, Berlin, Germany, 1993. Jeffrey K. MacKie-Mason and Michael P. Wellman. Automated markets and trading agents. In Tesfatsion Leigh and L. Judd Kenneth, editors, Handbook of Computational Economics, volume 2, pages 13811431. Elsevier, Amsterdam, The Netherlands, 1 edition, 2006. Sofus A. MacSkassy, Hirsh Hayim, Provost Foster, Sankaranayanan Ramesh, and Dhar Vasant. Intelligent information triage. In Proceedings of the 24th Annual International Conference on Research and Development in Information Retrieval (SIGIR), pages 318326, New Orleans, 2001. ACM. Spyros Makridakis, Steven Wheelwright, and Rob J. Hyndman. Forecasting: methods and applications. John Wiley Sons, New York, 1998. B.B. Mandelbrot. The variation of certain speculative prices. Journal of Business, 36:392417, 1963. Benoit Mandelbrot. Forecasts of future prices, unbiased markets, and martingale models. The Journal of Business, 39(1):242255, 1966. Bernard Mangold, Mike Dooley, Rael Dornfest, Gary W. Flake, Havi Hoffman, Tejaswi Kasturi, and David M. Pennock. The tech buzz game. IEEE Computer, 38(7):9497, 2005.

BIBLIOGRAPHY

179

Henry G Manne. Insider trading: Hayek, virtual markets, and the dog that did not bark, 2005. Charles F. Manski. Interpreting the predictions of prediction markets. Economic Letters, 91: 425429, 2006. David McAdams and Thomas W. Malone. Internal markets for supply chain capacity allocation, 2005. Daniel L. McFadden. Economic choices. Nobel Prize Lecture, December 2000. Richard D. McKelvey and Talbot Page. Common knowledge, consensus, and aggregate information. Econometrica, 54(1):109127, 1986. Adam Meirowitz and Joshua A. Tucker. Learning from terrorism markets. Perspectives on Politics, 2(2):331335, 2004. Gregoris Mentzas, Dimitris Apostolou, Efthimios Bothos, and Babis Magoutas. Information markets for social participation in public policy design and implementation. In Social Innovation and Social Media (SISoM2011), Barcelona, Spain, July 21 2011. Paul R. Milgrom and Nancy Stokey. Information, trade and common knowledge. Journal of Economic Theory, 26:1727, 1982. M.L. Mitchell and J.H. Mulherin. The impact of public information on the stock market. Journal of Finance, 49(3):923950, 1994. M. A. Mittermayer. Forecasting intraday stock price trends with text mining techniques. In Proceedings of the 37th Annual Hawaii International Conference on System Sciences (HICSS), volume 3, page 30064.b, Big Island, Hawaii, January 05-08 2004. IEEE Computer Society. Marc-Andr e Mittermayer and Gerhard F. Knolmayer. Text mining systems for market response to news: A survey. Working Paper No 184, Institute of Information Systems, University of Bern, August 2006. Gerald Murauer. Vergleich von Wahlprognosen durch Meinungsumfragen und Bewertung durch Boersen anhand der Oberoesterreichischen Landtagswahl 1997. Diplomarbeit, Technische Universitaet, 1997.

180

BIBLIOGRAPHY

J. Muth. Rational expectations and the theory of price movements. Econometrica, 29(6): 315335, 1961. Lars Tyge Nielsen, Adam Brandenburger, John Geanakoplos, Richard D. McKelvey, and Talbot Page. Common knowledge of an aggregate of expectations. Econometrica, 58(5): 12351238, 1990. Noam Nisan. Algorithmic Game Theory, chapter Introduction to Mechanism Design (for Computer Scientists), pages 209242. Number 9. Cambridge University Press, 2007. Markus Noeth and Martin Weber. Information aggregation with random ordering: Cascades and overcondence. Economic Journal, 113:166189, 2003. Markus Noeth, Colin F. Camerer, Charles R. Plott, and Martin Weber. Information aggregation in experimental asset markets: Traps and misaligned beliefs, 1999. T. Norvig. Consensus of subjective probabilities: A convergence theorem. Annals of Mathematical Statistics, 38:221225, 1967. John OBrien and Sanjay Srivastava. Dynamic stock markets with multiple assets: An experimental analysis. Journal of Finance, 46(5):18111838, 1991. Kenneth Oliven and Thomas A. Rietz. Suckers are born but markets are made: Individual rationality, arbitrage, and market efciency on an electronic futures market. Management Science, 50(3):336351, 2004. Dilek Onkal-Atay, Mary E. Thomson, and Andrew C. Pollock. Judgemental forecasting. In Michael P. Clements and David F. Hendry, editors, A Companion to Economic Forecasting. Wiley-Blackwell, 2008. Gerhard Ortner. Experimentelle Aktienmaerkte als Prognoseinstrument: Qualitaetskriterien der Informationsverarbeitung in Boersen am Beispiel Political Stock Markets. PhD thesis, Universitaet Wien, 1996. Gerhard Ortner. Forecasting markets - an industrial application, part i, 1997. Gerhard Ortner. Forecasting markets - an industrial application, part ii, 1998.

BIBLIOGRAPHY

181

Gerhard Ortner, A Stepan, and J Zechner. Political stock markets. the austrian experiences. Zeitschrift fuer Betriebswirtschaft, (4/95):123135, 1995. Abraham Othman, Tuomas Sandholm, David M Pennock, and Daniel M Reeves. A practical liquidity-sensitive automated market maker. In Proceedings of the 11th ACM conference on Electronic commerce, pages 377386. ACM, 2010. Marco Ottaviani and Peter Norman Sorensen. Aggregation of information and beliefs in prediction markets, 2005. Alexander Pagon. Polls and markets in the 2004 presidential election: A risk premium to explain deviations between polling predictions and market prices, 2005. David C. Parkes. Iterative Combinatorial Auctions: Achieving Economic and Computational Efciency. PhD thesis, University of Pennsylvania, Department of Computer and Information Science, 2001. David L. Passmore, Evin D. Cebeci, and Rose M. Baker. Market-based information for decision support in human resource development. Human Resource Development Review, 4, 2005a. David L. Passmore, Evin D. Cebeci, and Rose M. Baker. Potential of idea futures markets in educational technology. The ICFAI Journal of Derivatives Market, October:5666, 2005b. David M. Pennock. Aggregating Probabilistic Beliefs: Market Mechanisms and Graphical Representations. PhD thesis, University of Michigan, 1999. David M. Pennock. A dynamic pari-mutuel market for hedging, wagering, and information aggregation. In 5th ACM conference on Electronic commerce, pages 170179, New York, USA, 2004. ACM Press. http://doi.acm.org/10.1145/988772.988799. David M. Pennock and Rahul Sami. Algorithmic Game Theory, chapter Computational Aspects of Prediction Markets. Cambridge University Press, 2007. David M. Pennock and Michael P. Wellman. Representing aggregate belief through the competitive equilibrium of a securities market. In 13th Conference on Uncertainty in Articial Intelligence (UAI-97), pages 392400, Providence, RI, USA, 1997.

182

BIBLIOGRAPHY

David M. Pennock and Micheal P. Wellman. A market framework for pooling opinions, 2001. David M. Pennock, Steve Lawrence, C. Lee Giles, and Finn Aerup Nielsen. The power of play: Efciency and forecast accuracy in web market games, 2000. David M. Pennock, Steve Lawrence, C. Lee Giles, and Finn Aerup Nielsen. The real power of articial markets. Science, 291:987988, 2001a. David M. Pennock, Steve Lawrence, Finn Arup Nielsen, and C. Lee Giles. Extracting collective probabilistic forecasts from web games. In 7th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 174183, San Francisco, California, 2001b. David M. Pennock, Sandip Debnath, Eric J. Glover, and C. Lee Giles. Modeling information incorporation in markets with application to detecting and explaining events. In 18th Conference on Uncertainty in Articial Intelligence (UAI-2002), pages 405413, Alberta, Canada, 2002. Morgan Kaufmann. D. Peramunetilleke and R. K. Wong. Currency exchange rate forecasting from news headlines. In the 13th Australasian Databases Conference (ADC 2002), Melbourne, Australia, 2002. Mark Peters. Convex Mechanisms for Pari-Mutuel Markets. PhD thesis, Stanford University, 2009. Mark Peters, Anthony Man-Cho So, and Yinyu Ye. A convex parimutuel formulation for contingent claim markets. NSF grant DMS, 30661, 2005. Mark Peters, Anthony Man-Cho So, and Yinyu Ye. Pari-mutuel markets: Mechanisms and performance, volume 4858, pages 8295. Springer, Heidelberg, wine, lncs edition, 2007. Robert S. Pindyck and Daniel L. Rubinfeld. Econometric Models and Economic Forecasts. McGraw-Hill, 1981. C. Plott and S. Sunder. Rational expectations and the aggregation of diverse information in laboratory security markets. Econometrica, 56(1085-1118), 1988.

BIBLIOGRAPHY

183

Charles R. Plott. The application of laboratory experimental methods to public choice. In S. Russell Clifford, editor, Collective Decision-Making: Applications from Public Choice Theory, pages 137160. Johns Hopkins Press for Resources for the Future, Baltimore, Md., 1979. Charles R. Plott. Markets as information gathering tools. Southern Economic Journal, 67 (1-15), 2000. Charles R. Plott. Equilibrium, equilibration, information and multiple markets: From basic science to institutional design. In Nobel Symposium on Behavioral and Experimental Economics, Stockholm, Sweden, 2001. Charles R. Plott and Kay-Yut Chen. Information aggregation mechanisms: Concept, design and implementation for a sales forecasting problem, 2002. Charles R. Plott and Shyam Sunder. Efciency of experimental security markets with insider information: An application of rational expectations models. Journal of Political Economy, 90:663698, 1982. Charles R. Plott, Joergen Wit, and W.C. Yang. Parimutuel betting markets as information aggregation devices: experimental results. Technical Report Social Science Working Paper 986, California Institute of Technology, April 1997. Charles Polk, Robin Hanson, John Ledyard, and Takashi Ishikida. The policy analysis market: an electronic commerce application of a combinatorial information market. In 4th ACM conference on Electronic commerce, pages 272273, San Diego, CA, USA, 2003. ACM Press. Roy Radner. Rational expectations equilibrium: Generic existence and the information revealed by prices. Econometrica, 47(3):655678, 1979. Paul W. Rhode and Koleman S. Strumpf. Manipulating political stock markets: A eld experiment and a century of observational data, 2006. Jill Rickershauser. Judging the consequential events of the 2004 presidential campaign: Prediction markets, volatility, and campaign effects, 2006.

184

BIBLIOGRAPHY

A.Thomas Rietz. Behavioral mis-pricing and arbitrage in experimental asset markets, 2005. Juan Rodriguez Poo. Computer-Aided Introduction to Econometrics. Springer, 2003. Richard Roll. Orange juice and weather. The American Economic Review, 74(5):861880, 1984. E. S. Rosenbloom and William Notz. Statistical tests of real-money versus play-money prediction markets. Electronic Markets, 16(1):6369, 2006. Richard N. Rosett. Gambling and rationality. Journal of Political Economy, 73(6):595607, 1965. Leonard J. Savage. Elicitation of personal probabilities and expectations. Journal of the American Statistical Association, 66(336):783801., 1971. D. Scharfstein and J. Stein. Herd behavior and investment. American Economic Review, 80: 465479, 1990. Mark J. Schervish. A general method for comparing probability assessors. The Annals of Statistics, 17(4):18561879, December 1989. Carsten Schmidt and Axel Werwatz. How accurate do markets predict the outcome of an event? the euro 2000 soccer championships experiment, 2002. Jared M. Schrieber. The Application of Prediction Markets to Business. Master of engineering in logistics, MIT, 2004. R.P. Schumaker and H. Chen. Textual analysis of stock market prediction using nancial news articles. In 12th Americas Conference on Information Systems (AMCIS), Acapulco, Mexico, August 4-6 2006. Young-Woo Seo, Joseph Giampapa, and Katia Sycara. Text classication for intelligent portfolio management. Robotics Institute, Carnegie Mellon University, May 2002. Emile Servan-Schreiber, Justin Wolfers, David M. Pennock, and Brian Galebach. Prediction markets: Does money matter? Electronic Markets, 14(3):243251, 2004.

BIBLIOGRAPHY

185

L. Shapley and M. Shubik. Trading using one commodity as a means of payment. Journal of Political Economy, 85:937968, 1977. D. Sharfstein and J. Stein. Herd behavior and investment. American Economic Review, 80: 465479, 1990. S.K. Shevade, S.S. Keerthi, C. Bhattacharyya, and K.R.K. Murthy. Improvements to smo algorithm for svm regression. Technical Report CD-99-16, Control Division Dept of Mechanical and Production Engineering, National University of Singapore, 1999. Scott P. Simkins. Promoting active-student learning using the world wide web in economics courses. Journal of Economic Education, Summer 1999:278291, 1999. Bernd Skiera and Martin Spann. Opportunities of virtual stock markets to support new product development. In Saaenke Albers, editor, Cross-functional Innovation Management, pages 227242. Verlag Gabler, Wiesbaden, 2004. Eric Smith, J. Doyne Farmer, Laslo Gillemot, and Supriya Krishnamurthy. Statistical theory of the continuous double auction. Quantitative Finance, 3(6):481514, 2003. Michael A. Smith, David Paton, and Leighton Vaughan Williams. Market efciency in person-to-person betting. Economica, 2005. Alex J. Smola and Bernhard Scholkopf. A tutorial on support vector regression. Nc2-tr1998-030, NeuroCOLT2 Technical Report Series, 1998. Janet A. Sniezek. A comparison of techniques for judgmental forecasting by groups with common information. Group Organization Management, 15(1):519, 1990. Eric Snowberg, Justin Wolfers, and Eric Zitzewitz. Information (in)efciency in prediction markets. In Leighton Vaughan Williams, editor, Information Efciency in Financial and Betting Markets. Cambridge University Press, Cambridge, UK, 2005. Erik Snowberg, Justin Wolfers, and Eric Zitzewitz. Partisan impacts on the economy: Evidence from prediction markets and close elections, 2006. Wayne W. Snyder. Horse racing: Testing the efcient markets model. Journal of Finance, 33 (4):11091118, 1978.

186

BIBLIOGRAPHY

Arina Soukhoroukova and Martin Spann. New product development with internet-based information markets: Theory and empirical application. Regensburg, Germany, 2005. inproceedings George New Product Development with Internet-Based Information Markets: Theory and Empirical Application. Arina Soukhoroukova and Martin Spann. Informationsmaerkte. Wirthschaftsinformatik, 48 (1):6164, 2006. Martin Spann and Bernd Skiera. Internet-based virtual stock markets for business forecasting. Management Science, 49(10), 2003a. Martin Spann and Bernd Skiera. Taking stock of virtual markets. how can internet-based virtual stock markets be applied or business forecasting and other forecasting issues? Operations Research Management Science Today, 30(5):2024, 2003b. Martin Spann and Bernd Skiera. Einsatzmoeglichkeiten virtueller boersen in der marktforschung. Zeitschrift fuer Betriebswirtschaft, 74:2548, 2004. S. Sunder. The Handbookof Experimental Economics, chapter Experimental Asset Markets, pages 445500. Princeton University Press, 1995. Shyam Sunder. Markets for information: Experimental evidence. Econometrica, 60:667695, 1992. Cass R. Sunstein. Group judgments: Deliberations, statistical means, and information markets, 2004. Cass R. Sunstein. Infotopia: How Many Minds Produce Knowledge. Oxford University Press, USA, 2006a. Cass R. Sunstein. Deliberation and information markets. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 67100. AEI Press, Washington D.C., 2006b. Cass R. Sunstein. Neither hayek nor habermas. Public Choice, 134:8795, 2008.

BIBLIOGRAPHY

187

James Surowiecki. The Wisdom of Crowds: Why the Many Are Smarter Than the Few and How Collective Wisdom Shapes Business, Economies, Societies and Nations. Doubleday, New York, 2004. Paul Tetlock and Robert Hahn. Optimal liquidity provision for decision makers, 2006. Paul Tetlock, Robert Hahn, and Donald Lien. Designing information markets for decision making, 2005. Paul C. Tetlock. How efcient are information markets? evidence from an online exchange, 2004. Richard H. Thaler and William T. Ziemba. Anomalies: Parimutuel betting markets: Racetracks and lotteries. The Journal of Economic Perspectives, 2(2):161174, 1988a. Richard H. Thaler and William T. Ziemba. Anomalies: Parimutuel betting markets: Racetracks and lotteries. The Journal of Economic Perspectives, 2(2):161174, 1988b. J.D. Thomas and K. Sycara. Integrating genetic algorithms and text learning for nancial prediction. In A. A. Freitas, W. Hart, N. Krasnogor, and J. Smith, editors, Data Mining with Evolutionary Algorithms, Proceedings of the Genetic and Evolutionary Computing Conference (GECCO), pages 7275. Las Vegas, Nevada, July 8-12 2000. G. Tziralis, A. Tolis, I. Tatsioopulos, and K. Aravossis. Environmental Economics And Investment Assessment, chapter Economic aspects and the sustainability impact of the Athens 2004 Olympic games, pages 2133. WIT Press, 2006. Georgios Tziralis and Ilias Tatsiopoulos. Prediction markets: Aspects on transforming a business game into a valuable information aggregation tool. In 3rd Future Business Technology Conference,, pages 3539, Athens, Greece, 2006. Georgios Tziralis and Ilias Tatsiopoulos. Prediction markets: An extended literature review. The Journal of Prediction Markets, 1:7591, 2007. Gerrit H Van Bruggen, Martin Spann, Gary L Lilien, and Bernd Skiera. Prediction markets as institutional forecasting support systems. Decision Support Systems, 49(4):404416, 2010.

188

BIBLIOGRAPHY

V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, 1995. Springer, 1995. Hal R. Varian. The arbitrage principle in nancial economics. Journal of Economic Perspectives, 1(2):5572, 1987. Sun-Chong Wang, Sai-Ping Li, Chung-Ching Tai, and Shu-Heng Chen. Statistical properties of an experimental political futures market, 2005. Jennifer H Watkins and Marko A Rodriguez. A survey of web-based collective decision making systems. In Evolution of the Web in Articial Intelligence Environments, pages 243277. Springer, 2008. Martin Weitzman. Utility analysis and group behavior: An empirical study. Journal of Political Economy, 73(1):1826, 1965. Leighton Vaughan Williams. Models, markets, polls and pundits: a case study of information efciency. In Leighton Vaughan Williams, editor, Information Efciency in Financial and Betting Markets. Cambridge University Press, Cambridge, UK, 2005. Robert L. Winkler. Probabilistic prediction: Some experimental results. Journal of the American Statistical Association, 66(336):675685, 1971. H.I. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and Techniques. Number 2. Morgan Kaufmann, San Francisco, California, 2005. Justin Wolfers and Andrew Leigh. Three tools for forecasting federal elections: Lessons from 2001. Australian Journal of Political Science, 37(2):223240, 2002. Justin Wolfers and Eric Zitzewitz. Prediction markets. Journal of Economic Perspectives, 18 (2):107126, 2004a. Justin Wolfers and Eric Zitzewitz. Using markets to inform policy: The case of the iraq war, 2004b. Justin Wolfers and Eric Zitzewitz. Prediction markets in theory and practice. In Steven Durlauf and Lawrence Blume, editors, The New Palgrave Dictionary of Economics, page forthcoming. McMillan, 2nd edition, 2006a.

BIBLIOGRAPHY

189

Justin Wolfers and Eric Zitzewitz. Five open questions about prediction markets. In Robert W. Hahn and Paul C. Tetlock, editors, Information Markets: A New Way of Making Decisions, pages 1336. AEI Press, Washington D.C., 2006b. Justin Wolfers and Eric Zitzewitz. Interpreting prediction market prices as probabilities, 2006c. George Wright and Peter Ayton. Judgmental Forecasting. John Wiley Sons, 1987. B. Wuthrich, D. Permunetilleke, S. Leung, V. Cho, J. Zhang, and W. Lam. Daily stock market forecast from textual web data. In IEEE International Conference on Systems, Man, and Cybernetics, volume 3, pages 27202725, San Diego, California, October 11-14 1998. IEEE Press. Annie Yang. Information markets: A tool for effective decision making, 2005. Y. Yang and X. Liu. A re-examination of text categorization methods. In The 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 4249, 1999. Yiming Yang, Tom Pierce, and Jaime Carbonell. A study of retrospective and on-line event detection. In Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, pages 2836, Melbourne, Australia, August 24-28 1998. G. Zhang, E. Patuwo, and M. Hu. Forecasting with articial neural networks: The state of the art. International Journal of Forecasting, 14:3562, 1998.

190

BIBLIOGRAPHY

Appendix A

Revision
A.1 Overview
The Thesis was essentially completed in 2008, yet its defence took place in June 2013. In the meanwhile, prediction markets have emerged to become a broadly accepted forecasting mechanism, at least according to mentions like the couple ones Armstrong, a leading authority in the forecasting eld, recently co-authored [Lawrence et al., 2006b, Green and Armstrong, 2012]. At the same time, the volume of relevant publications largely follows the trends mentioned in Chapter 2. This Appendix presents a rough attempt to update the contents of this Thesis, from the completion of its underlying research until the time of its submission. This is not an attempt to provide an update at an extent similar to the literature survey under Chapter 2. Instead, effort is put in summarizing recent ndings and, most importantly, validating the novelty of the contributions across this Thesis and especially Chapter 3. The contents of this Appendix are as follows: Section A.2 summarizes the recent trends and volume of prediction market literature in general, next to providing an update to the generic acceptance of markets as a forecasting mechanism, or its lack there of. Section A.3 studies the latest developments on the eld of prediction markets mechanisms, under the prism of the contributions of Chapter 3. Finally, Section A.4 concludes the outcomes of this Appendix.

191

192

APPENDIX A. REVISION

A.2 Trends
Prediction markets as a research eld are, similarly to what proposed in Chapter 2, largely growing 1 . This trend seems to apply in both theory and applications, in an ever growing canvas of different areas. A detailed study of the totality of recent publications remains out of the scope of this revision. However, a closer look will be attempted in updating the status quo of prediction markets as a forecasting mechanism, next to comparing it to other well established alternatives. A number of relevant publications will be reviewed below, while the outcomes of their study will be summarized right after. Bray et al. [2008] introduced performance and feasibility as metrics equally important to accuracy for a prediction market operation. The authors proceed to compare markets to other information aggregation alternatives, including one or more expert opinions, surveys and group deliberation. Expectably, results are mixed; not only is the feasibility of the various approaches context-sensitive, but even where all alternatives are feasible, the optimal choice of mechanism, as judged by the balance of costs and benets, depends on the setting. No matter the facilitated comparisons, the authors argue, the literature has offered little to assess the relative costs of implementation and operation, along with risk factors, of prediction markets as opposed to expert opinion. Their eld evidence eventually suggests the perceived legitimacy of markets as low, while accuracy and capacity to update results in real-time is high. Markets also are reported to stand as a tool of high set up and operation costs compared to other approaches, and their content does not always remain self-explanatory, yet there still exist some positive non-replicable effects out of their operation. Watkins and Rodriguez [2008] provided a survey of web-based collective decision support systems, comparing prediction markets to other, more or less novel approaches. These approaches include document ranking, folksonomy, recommender system, vote system, wiki and open source software, and are compared on the basis of their common
1 For

example, as a rough indication, Google Scholar (http://scholar.google.gr/scholar) suggests the

existence of 2940 articles referring to prediction markets between 2008 and 2013, while only 655 before that. It is to be noted though that not all of those are related to prediction markets, as these are dened in this Thesis.

A.2. TRENDS

193

purpose, namely structuring individual input to generate meaningful aggregate decisions. A taxonomy is presented, organized into four classes, namely problem space, implementation, individual and collective features. The characteristics of markets, among the under study approaches, are deployed and eventually the authors conclude that, similarly to the efforts above, each system has particular benets, applications and disadvantages, while none can be characterized as a superior choice. Van Bruggen et al. [2010] compared the performance of prediction markets with more traditional judgment-based approaches, namely the approach of a single informant and a weighted average of up to six informants. The number of participants in both markets and traditional approaches is selected small on purpose, so as to provide insights on the performance of markets in such settings that have not been studied in great extent, as opposed to cases with much larger participation. Results suggest prediction markets to outperform traditional approaches in settings where each informant possesses unique information. In other cases, markets performed no worse than their alternatives. In this context, markets are suggested for usage even with quite small group sizes, especially under situations of high information-heterogeneity. Graefe and Armstrong [2011] compared prediction markets to face-to-face meetings, nominal groups and delphi, within a university laboratory for a simple estimation task. The three more structured approaches performed expectably better than face-to-face meetings, yet the differences in accuracy were small. Moreover, each methods accuracy was affected by the difculty of questions, as in hard-to-estimate questions meetings outperformed nominal groups and delphi. What was equally interesting though, was the study of participants perception of the tools they used, since this is decisive for the acceptance of a methods results. In this assessment, prediction markets performed poor. Due to lack of personal communication and relative difculty to understand, markets were not among the favorite choices for participants, a fact that may hinder further adoption of the mechanism, while at the same time highlights the opportunity in providing more appealing software solutions for non-experienced traders. Finally, Deck and Porter [2013], addressing the mechanisms increasing popularity, surveyed laboratory experiments that provide part of the scientic justication for using prediction markets. While markets are indeed proved to be far from a panacea, further

194

APPENDIX A. REVISION

evidence is provided that they stand as a viable approach for aggregating and disseminating information. In Lawrence et al. [2006b], summarizing 25 years of progress in the eld of judgmental forecasting, the authors mention prediction markets as the rst choice in combining forecasts, which recently arises as a quite active research eld. In Green and Armstrong [2012], summarizing comparative testing of various forecasting approaches in recent decades, prediction markets are suggested as one of the mechanisms one should employ in cases where quantitative data is scarce.

A.3 Developments
Chapter 3 summarized the state of the art at the time of writing for proper prediction market mechanisms. It then proceeded to propose a set of properties providing for a coherent price function for a dynamic pari-mutuel market, along with a couple of such functions. These contributions were made on the basis of the existing research on the topic which proved limited related to the directions under focus, more specically due to the lack of any similar reference on a price functions framework, on top of the couple of functions Pennock [2004] originally proposed. The following paragraphs will provide a summary of recent research developments that bring some relevance to the above topic, with the target to stand for extra evidence on the contributions novelty, or its lack there of. In general, recent literature suggests that the logarithmic market scoring rule has emerged as the de facto standard for a prediction market mechanism [Brahma et al., 2012, Chen and Pennock, 2010]. This observation arises as valid in the majority of both applications and theoretical work recently conducted. On top of the arguments already provided as market scoring rules advantages against dynamic pari-mutuel in Chapter 3, the author argues that the lack of appropriate price functions has manifested itself in real-world usage attempts of the mechanism, the unsatisfying properties of which may have contributed their own part in the dynamic pari-mutuel mechanisms limited acceptance. Nevertheless, a number of publications are in various capacities related to this and other prediction market mechanisms in general, and will be carefully summarized in the paragraphs to follow. Peters et al. [2005, 2007], Peters [2009] proposed a sequential convex pari-mutuel

A.3. DEVELOPMENTS

195

mechanism, a sequential version of a call auction problem which is solved by convex optimization. Essentially, the mechanism stands as a market maker solving in each step a set of optimization functions, to determine whether and when to accept orders and how to update prices of available contracts. The approach attempts to match the benets of popular prediction market mechanisms, yet it features some attributes that render its appeal practically limited. For example, the market maker does not guarantee the acceptance of each bid, while each time the computation of new prices requires solving a complex system of non-linear equations. In this context, this proposed mechanism has not emerged as dominant, yet it stands as a competitive alternative. More interestingly, in the context of Peters et al. [2007], an empirical comparison of the performance of prediction market mechanisms is presented, using the share-ratio price function in the dynamic pari-mutuel contestant. Among ten simplistic data sets of 500 available orders each and within varying results, the dynamic pari-mutuel setting demonstrated improved price stability, compared to the market scoring rules and the sequential convex pari-mutuel ones. Gao et al. [2009] introduced innite and continuous outcome generalizations of call markets, market scoring rules and dynamic pari-mutuel mechanisms. The authors continued to provide evidence that the market makers matching problem can be solved in polynomial time for call markets, while market scoring rules suffer from unbounded loss in both cases of continuous and innite outcomes. Under the dynamic pari-mutuel approach, loss for the market maker remains bounded to the initial cost for operating the market. However, in the continuous-outcome case, it is proved that a trader may incur losses even if her selected interval includes the true outcome. The continuous and innite outcome variations may be quite useful in practice, granted the wide applicability of such cases, however the added complexity remains signicant and its trade-off should be examined in each case. Agrawal et al. [2011] introduced a unied framework for dynamic pari mutuel market design. In practice, the proposed approach has little to do with dynamic pari-mutuel markets, as in this case pari mutuel refers to the generic fact that the winners are generally paid out by the stakes of the losers, which applies also to market scoring rules and sequential convex mechanisms. The theoretical approach attempted to provide a

196

APPENDIX A. REVISION

foundation for unifying existing market makers like market scoring rules, sequential convex pari-mutuel, cost function and utility based market makers, namely pretty much all available approaches but dynamic pari-mutuel. More specically, the authors demonstrate a generalized version of the sequential convex pari-mutuel mechanism as a unifying convex optimization framework of the above market makers. It is also shown that the sequential convex pari-mutuel framework is equivalent to market scoring rules, as any proper scoring rule can be formulated as a sequential convex pari-mutuel model. No further mention was made with regards to dynamic pari-mutuel markets. Building on Gao et al. [2009]s work, Gao and Chen [2010] dened the cost functional based market makers, a general class of automated market maker mechanisms for continuous-outcome events, generalizing the cost function based market makers for nite-outcome events to handle continuous-outcome spaces. In this framework, a cost functional is dened as valid if instantaneous prices are non-negative and form a probability distribution over the outcome space. For a valid cost functional, the authors proceeded to dene a set of desirable properties, namely differentiability, increasing monotonicity and positive translation invariance. They also provided evidence that no cost functional based market maker for continuous-outcome events can satisfy all three properties simultaneously. No matter the limitations of its nal results and the different context compared to the coherency framework deployed in Chapter 3, this research effort is the rst and only one found across literature that attempts to provide a set of properties that a market maker, or a price function, should comply with, to stand for a well-operating prediction market mechanism. Chen and Vaughan [2010] demonstrated, among other ndings, the elegant mathematical connection between market scoring rules, cost function based prediction markets and no-regret learning. More specically, the authors showcased a mapping between a mildly restricted class of convex cost function based markets and a mildly restricted class of strictly proper market scoring rules. In other words, any regular, it was proved that a strictly proper market scoring rule with differentiable scoring functions can be translated as a convex cost function based market, which can be solved with standard methods. Moreover, granted that the market probabilities are non-zero, the market scoring rule and corresponding cost function based markets are proved to be equivalent. It is however to

A.3. DEVELOPMENTS

197

be noted that the dynamic parimutuel market falls outside the scope of the authors effort. Othman et al. [2010] extended the logarithmic market scoring rules to provide for a liquidity sensitive market maker. More specically, a major shortcoming of the original logarithmic market scoring rules is their limited ability to adapt to market liquidity levels, due to liquidity being dened by a constant set a priori, before the market operation starts. Moreover, higher liquidity implies an increase to the market makers worst-case loss. The proposed variant automatically adjusts liquidity according to the existing level of activity within the market, while it can also ensure an arbitrarily small loss in the worst case scenario, next to a positive prot in most other cases. That said, the new market maker also comes with a few shortcomings, including its increased complexity and the fact that market prices can sum to greater than one. Brahma et al. [2012] introduced a bayesian market maker, comparing its performance to logarithmic market scoring rules. The proposed market maker for binary or continuous 0-1 outcomes provides liquidity by adapting its spread based on current uncertainty level, which enables small spreads when close to equilibrium, while remaining adaptive to market shocks. Experimental evaluations of the algorithm with both articial agents and real traders suggest some benets against market scoring rules, including liquidity and quick adaptation, however there is also some theoretical risk of substantial loss. A property that suggested as broad across market makers highlights an inherent trade-off between adaptation and fast convergence towards equilibrium, versus potential loss. Chen and Pennock [2010], similarly to the target of this Appendix, provided a survey to a number of mechanisms created to elicit predictions, essentially drawing the stateof-the-art for prediction markets mechanisms. Reviewing recent research on designing automated market makers that provide sufcient market liquidity, the authors highlight the following categories as appropriate for facilitating the operation of a prediction market institution: market scoring rules, convex cost-function-based market makers and sequential convex parimutuel mechanism ([Peters et al., 2007, Agrawal et al., 2011]), utility-based market makers and dynamic parimutuel market. In this reference, market scoring rules are repeatedly highlighted as the de facto standard, while, at the same time, no mention is being made on dynamic pari mutuel market price functions or any similar developments on this front.

198

APPENDIX A. REVISION

A.4 Discussion
Section A.2 provided some further evidence that prediction markets enjoy increasing acceptance as a forecasting mechanism. Moreover, markets stand for an appealing alternative, subject to various requirements, for the forecasting problem at hand. That said, it is reiterated that this or any other mechanism cannot universally serve as the preferred contestant and its appropriateness should be investigated on a case by case basis. Furthermore, to the best of the authors knowledge, references mentioned in Section A.3 summarize recent ndings with regards to appropriate prediction market mechanisms. The amount of research performed is denitely growing, however, among the various new market makers and other ndings being introduced, the paragraphs above shall provide sufcient evidence that the contributions of Chapter 3, and therefore Chapters 4, 5 still apply.

Index
agent utility, 88 AI, Articial intelligence, 3 articial markets, 14 betting markets, 9 bookmaker, 32 call market auction, 31 causal methods, 2 common knowledge, 81 common-prior assumption, 85 complete market, 82 conditional logit model, 61 continuous double auction (CDA), 31 Continuous Double Auction with Market Maker (CDAMM), 32 convergence, 46, 49 decision markets, 13 differentiability, 46, 49 DPM share-ratio price function, 42 Dynamic Pari-mutuel Market(DPM), 39 efcient market hypothesis, 81 election stock markets, 14 electronic markets, 13 event detection, 130 expected utility maximizer, 88 experimental economics, 83 nancial markets, 9 forecasting, 2 fundamental economics equation, 77 futures markets, 9 GARCH model, 134 hedging, 9 idea futures, 14 incomplete markets, 82 information aggregation, 9 information markets, 13 injection, 46, 49 input bit, 85 investment, 9 judgmental methods, 3 Knowledge Discovery from Textual sources (KDT), 132 logarithmic scoring rule, 39 market accuracy, 105 Market Maker (MM), 32 Market Scoring Rules (MSR), 36 mechanism design, 77 model design, 77 money-ratio function, 42 monotonicity, 46, 49 MSR cost function, 38 pari-mutuel market (PMM), 33 political stock markets, 14 prediction markets, 11 proper scoring rules, 37 rational expectations, 80 rational expectations equilibrium (REE), 80 reexiveness, 46, 49 Scoring Rules, 36 Shapley-Shubik market game, 86 speculation, 9 state space, 85 stock markets, 9 summation, 46, 49 Support Vector Machines (SVM), 136 199

200

INDEX

taxonomy, 18 text mining, 132 thin market, 32 time-series methods, 2 Topic Detection and Tracking (TDT), 130 virtual markets, 13 volatility, 150 volatility clustering, 134