Cross-Sectional Predictability of Stock Returns: Pre-World War I Evidence

FACULTY OF APPLIED ECONOMICS
DEPARTMENT OF ACCOUNTING AND FINANCE

DISSERTATION

Cross-Sectional Predictability of Stock
Returns:
Pre-World War I evidence

THESIS SUBMITTED IN ORDER TO OBTAIN THE DEGREE
OF
DOCTOR IN APPLIED ECONOMICS

Author: Supervisor:
Lord Mensah Prof. Dr. Jan Annaert

September, 2011

This research was funded by the Faculty of Applied Economics, from the 2007 to 2011
research grant for the Accounting and Finance Research group.

Cross-Sectional Predictability of Stock return: Pre-World War I Evidence
Lord Mensah, Antwerpen, Belgium, 2011
ISBN: 978-90-8994-044-5
Printed by: Universitas Antwerpen
Copyright 2011, Lord Mensah
All rights reserved. No part of this book may be reproduced or transmitted in any form
by any electronic or mechanical means (including photocopying, recording or
information storage and retrieval) without permission in writing from the author.

Doctoral Jury

Internal members
Prof. Dr. Jan Annaert (Supervisor)
Universiteit Antwerpen, Belgium
Department of Accounting and Finance
Jan.annaert@ua.ac.abe

Prof. Dr. Marc De Ceuster (chair/PhD committee chair)
marc.deceuster@ua.ac.be

Dr. Frans Buelens (PhD committee member)
Fran.buelens@ua.ac.be

Prof. Dr. Joseph Plasmans
Department of Economics
joseph.plasmans@ua.ac.be

External Members
Prof. Dr. Wim Janssens (PhD committee member)
Universiteit Hasselt, Belgium
Department of Marketing
wim.janssens@uhasselt.be

Prof. Dr. Laurens Swinkels
Erasmus School of Economics, Rotterdam, Netherlands
Vice President, Robeco Investment Solutions, Rotterdam
lswinkels@ese.eur.nl

Prof. Dr. Patrice Fontaine
EUROFIDAI and University of Grenoble 2, France
patrice.fontaine@eurofidai.org

ACKNOWLEDGEMENTS

In as much as writing a PhD thesis requires independent study and scientific research, the
success marked in this project depend on the environment I worked in and the people
around me. As a result, I may like to express my profound gratitude to everyone who
contributed to the success of this project both professionally and personally.
Primarily, I would like to thank my supervisor Prof. Dr. Jan Annaert for the confidence
pose in me, and giving me the opportunity to be part of the research project. You guided
me through the research and thought me how to select the literature to support my
arguments. In a mother-child scenario, I will say you know the right food combination
you will give to your child in order to mature quickly. I also appreciate how quick you
read my chapters and direct me for changes here and there. In short, this dissertation
would not be what it is now without your suggestions, encouragement and guidance.
Secondly, I would like to express my sincere gratitude to my PhD committee members
Prof. Dr. Mark DeCeuster, Prof. Dr. Wim Janssens and Dr. Frans Buelens. Your
comments and suggestions have been of immense value in completing this dissertation.
Dr. Frans Buelens, thank you very much for the updates you gave me on the data. In
addition, I am also thankful to the external doctoral jury members Prof. Dr. Joseph
Plasmans, Prof. Dr. Laurens Swinkels and Prof. Dr. Patrice Fontaine for taking your time to read
and discuss my work. Your comments and suggestions have been of great value to improve and
polish the chapters in this thesis.

Genuinely, I have enjoyed every moment at work for the past four years. On this note, I
want to thank all my colleagues from the Accounting and Finance department of the
University of Antwerp. Specifically, my sincere gratitude goes to the secretariat office
for their beaming smiles and friendly services. I want to thank the department head for
the annual outing, lunches and visit to museums in Antwerp. This allows me to release
stress at work for the past four years. I hope that we will stay in touch for the future.

Finally, the period of writing PhD thesis comes with moments of joy and doubt.
However, how to handle the stress depends on the people around you after office work. I
have been blessed with supportive family. Firstly, I owe an immense debt to my wife
Sefam Ekua Lord-Mensah. What you have done for the past four years goes beyond your
call to duty. I have come to understand that only few people are fortunate to have caring
and supporting wife like you. Completion of my PhD, gives me the opportunity to thank
you for the unconditional love and support you have given me, since we got married.
Second words go to my lovely kids Lakeisha Lord-Mensah, Zita Lord-Mensah and
Nhyira Lord-Mensah. Thank you for making my home a place of endless happiness.
Thank you for your patience and not been on your sight ten to eleven hours every day.
My final words go to all those who made my education a success by contributing
financially. Special thanks to my one and only Uncle Maxford Kwadwo Kwakye who
contributed to my senior high school education. I do appreciate your support, and I will
always remember you.

Lord Mensah

Table of Contents

CHAPTER 1 .................................................................................................................................. 1
1 INTRODUCTION, DATA DESCRIPTION AND SUMMARY ................................... 1
Introduction .................................................................................................................... 1 1.1
1.1.1 The aim of this Dissertation .................................................................................... 6
Data Description and Methodology ................................................................................ 7 1.2
1.2.1 The Risk-free rate and the Market Index .............................................................. 11
1.2.2 Methodology ......................................................................................................... 14
Summary of the chapters .............................................................................................. 20 1.3
1.3.1 Chapter 2 ............................................................................................................... 20
1.3.2 Chapter 3 ............................................................................................................... 21
1.3.3 Chapter 4 ............................................................................................................... 22
1.3.4 Chapter 5 ............................................................................................................... 23
CHAPTER 2 ................................................................................................................................ 25
2 ASSESSMENT OF BETA IN THE 19
th
CENTURY BSE ........................................... 25
Introduction and Literature Review .............................................................................. 25 2.1
Beta Coefficient Descriptive Statistics .......................................................................... 33 2.2
Beta Stability ................................................................................................................. 37 2.3
2.3.1 Blume and Vasicek stability adjustment techniques............................................. 39
Beta Bias ........................................................................................................................ 44 2.4
Impact of outlying observations on Beta ...................................................................... 47 2.5
Conclusion..................................................................................................................... 51 2.6
CHAPTER 3 ................................................................................................................................ 52
3 THE TEST OF CAPITAL ASSET PRICING MODEL (CAPM) AND THE SIZE
EFFECT IN 19
th
CENTURY BSE ............................................................................................. 52
Expected returns of portfolios sorted on betas ............................................................ 54 3.2

3.2.1 The Cross-Sectional Regressions ........................................................................... 59
Expected Returns, Beta, and the Size Effect ................................................................. 66 3.3
3.3.1 Fama-MacBeth Cross-Sectional Regressions to Test the Size Effect..................... 71
Conclusion..................................................................................................................... 79 3.4
CHAPTER 4 ................................................................................................................................ 80
4 DOES THE MOMENTUM EFFECT EXIST IN THE 19
TH
CENTRURY? ............. 80
Momentum Trading Strategies and their Returns ......................................................... 85 4.2
4.2.1 Expected Returns and Average Size of Quintile Portfolios ................................... 90
4.2.2 Momentum Profit within Size and Beta-based Subsamples ................................. 91
Seasonality and Sub-period Analysis of the Momentum Profit .................................... 94 4.3
Post holding Period Momentum Profits ........................................................................ 98 4.4
The Momentum profit and the Market State ............................................................. 103 4.5
Conclusion................................................................................................................... 106 4.6
CHAPTER 5 .............................................................................................................................. 109
5 THE COMBINED EFFECT OF DIVIDEND YIELD, SIZE, TOTAL RISK
AND MOMENTUM (1868-1913 EVIDENCE)....................................................................... 109
Introduction and Literature Review ............................................................................ 109 5.1
Measures of Characteristics ........................................................................................ 115 5.2
Descriptive Summary Statistics of the Characteristics ................................................ 118 5.3
5.3.1 Size-Dividend yield double sorts ......................................................................... 118
5.3.2 Momentum-Dividend yield double sorts ............................................................ 121
5.3.3 Total risk-Dividend yield double sorts ................................................................. 125
Average Excess Returns on Portfolio Sorts ................................................................. 130 5.4
The Cross-Sectional regressions .................................................................................. 136 5.5
5.5.1 Pervasiveness of the Cross-Sectional Relationships ........................................... 140
Conclusion ................................................................................................................... 143 5.6
6 CONCLUSION .............................................................................................................. 144
References .................................................................................................................................. 150
NEDERLANDSTALIGE SAMENVATTING ........................................................................ 157

TABLES
Table 1.1: Summary Statistics for Risk-free rate and the Value-Weighted Index ......................... 13
Table 2.1: Beta coefficient descriptive statistics for the 15 estimated periods ............................ 34
Table 2.2: Average beta and average coefficient of determination of the size based sub-samples
....................................................................................................................................................... 36
Table 2.3: Weighted average of correlation and Spearman rank order correlation across
successive periods ......................................................................................................................... 38
Table 2.4: Measurement of regression tendency of estimated beta coefficient for individual
stocks ............................................................................................................................................. 40
Table 2.5: Predictive performance of Blume and Vasicek (Bayesian) procedures of estimating
beta ............................................................................................................................................... 42
Table 2.6: Modified Diebold-Mariano test statistics (p-value in parentheses) ............................. 44
Table 2.7: Dimson Aggregate Coefficient (AC) beta Adjustment .................................................. 45
Table 2.8: Comparison of the market model betas and the iterative reweighted least square
betas .............................................................................................................................................. 48
Table 2.9: Test of equal predictive accuracy between MM and IRLS models ............................... 50
Table 3.1: Time Series Mean (%), Standard Deviation (%), and Post-ranking Betas of Decile
portfolios formed from pre-ranking betas in Jan. 1868-Dec. 1913 .............................................. 58
Table 3.2: Average time series slopes from the Fama-MacBeth Cross-Sectional Regressions in
Jan. 1868-Dec. 1913 ...................................................................................................................... 61
Table 3.3: Average Time Series Slopes from Fama-French Cross-Sectional Regression in Jan.
1868-Dec. 1913 ............................................................................................................................. 64
Table 3.4: Sub-period look into estimated slopes and excess market returns ............................. 65
Table 3.5: Beta Estimate and Mean Excess Return for the BSE equally weighted size portfolios,
Jan. 1868- Dec. 1913 ..................................................................................................................... 68
Table 3.6: Equally weighted portfolios excess returns without the first-size decile group .......... 69
Table 3.7: Average time series slopes and intercept from the Fama-MacBeth cross-sectional
regression: Jan 1868-Dec. 1913 .................................................................................................... 73
Table 3.8: Average Time Series Slopes and Intercepts from the Fama-French Cross-Sectional
Regressions: Jan. 1868-Dec. 1913 ................................................................................................. 76

Table 3.9: Average Time Series Slopes and Intercepts from the Fama-French Cross-Sectional
Regressions without the First-Size Decile: Jan. 1868-Dec.1913 .................................................... 78
Table 4.1: Profitability of momentum Strategies on BSE (Jan.1868-Dec. 1913) ........................... 89
Table 4.2: Average Returns and Average Size of Quintile Momentum Portfolios ........................ 91
Table 4.3: Portfolio Returns of the Momentum Strategies with Size and Beta Subsamples ......... 92
Table 4.4: Seasonality in momentum profits ................................................................................ 95
Table 4.5: Sub-period Analysis of Momentum Profit .................................................................... 97
Table 4.6: Long Horizon Momentum Profits ............................................................................... 100
Table 4.7: The Momentum profit and the Market State ............................................................ 104
Table 5.1: Summary statistics for Size-Dividend double-sorts .................................................... 119
Table 5.2: Summary statistics for Momentum-Dividend yield double sorts .............................. 123
Table 5.3: Summary statistics of total risk- dividend yield double sorts .................................... 126
Table 5.4: Annual Time Series Average of the correlation between the entire characteristic and
Average return ............................................................................................................................ 129
Table 5.5: Equal and Value-weighted portfolios excess returns (%) of double-sorted
characteristics ............................................................................................................................. 131
Table 5.6: Cross-Sectional Regression of Excess Returns on Dividend Yield, Size, Total Risk and
Momentum ................................................................................................................................. 138
Table 5.7: Cross-Sectional regression of Excess Returns on Dividend Yield, Size, Total Risk and
Momentum of Size subsamples .................................................................................................. 141

FIGURES
Figure 1.1: Number of listings on the 19
th
Century Brussels Stock Exchange ................................. 9
Figure 1.2: Total market capitalization of the BSE ........................................................................ 10
Figure 1.3: Evolution of the value weighted market index in the two sub-periods of our study . 12
Figure 2.1: The graph of the average beta of each period for large stocks and small stocks ....... 36
Figure 2.2: Plot of average market model betas and IRLS betas for stocks with outlier
observation less than 4 ................................................................................................................. 49
Figure 3.1: Number of stocks in our selection criteria for the entire period of the pre-world war I
SCOB data ...................................................................................................................................... 55

Figure 3.2: Sixty months moving average of the cross-sectional slopes and excess market returns
using Dimson beta estimates ........................................................................................................ 65
Figure 3.3: Size Portfolio betas ...................................................................................................... 70
Figure 4.1: Number of common stocks in our sample for the momentum studies...................... 86
Figure 4.2: Time line of sample periods ........................................................................................ 86
Figure 4.3: Average returns of the momentum profit in all calendar months ............................. 95
Figure 4.4: Cumulative Returns for Five years after portfolio formation ................................... 101
Figure 5.5.1: Percentage of Zero-dividend paying stocks and their Relative Market Capital: 1868-
1913 ............................................................................................................................................. 116

1

CHAPTER 1
1 INTRODUCTION, DATA DESCRIPTION AND SUMMARY

More than fifty variables have been used to predict returns. The overall picture remains murky, because
more needs to be done to consider the correlation structure among the variables, use a comprehensive set
of controls and discern whether the results survive simple variations in methodology
1
.

Introduction 1.1
The issue about why stock returns differ in the cross-section from one another at a
particular time has been a hot topic of financial research for the past decades. The capital
asset pricing model (CAPM) proposed by Sharpe (1964), Lintner (1965) and Mossin
(1966) seems to provide an adequate description of the cross-section of stock returns
until the 1980s. The model postulates a linear relationship between expected returns and
the covariance between the market portfolio returns and the returns of an asset (beta).
This implies that the asset beta with respect to the market portfolio is sufficient to
determine its expected returns. However, since the 1980s, documentation on deviations
from the model (anomalies) has been an extremely active area of research. Either the
anomalies represent inadequacies in the CAPM, inefficiencies in the market (profit
opportunities) or a data snooping bias. If the presence of an anomaly truly indicates the
inadequacy of the CAPM, then factors other than beta can predict stock returns. This
implies that anomalies would continue to exist before and after their discovery. While
many have accepted the factors other than beta to predict returns, others believe that they
were discovered out of luck and are due to data snooping bias (Lo and MacKinlay

1
Subrahmanyam, Avanidhar, 2010, The cross-section of expected stock returns: What have we learnt from
the past twenty-five years of research?, European Financial Management 16, 27-42.
2

(1990)). The possible ways that data snooping bias can occur are (i) when researchers
continuously test the properties of a data set or the outcomes of other studies on data set
(ii) form predictive models based on the characteristics of the previous results and (iii)
test the power of their models on the same data set. As a result, any anomaly found
might appear to be valid within the data set, but they would have no statistical
significance outside the sample from which it was discovered. Indeed, Schwert (2003)
documents that anomalies often seem to disappear, reverse, or attenuate after they have
been documented and analyzed in the academic literature. The problem can be addressed
by using data from markets that have not been searched exhaustively, or by making
predictions using periods that are new to asset pricing research. The situation remains
murky, as more needs to be done to distinguish between the data snooping hypothesis
and the persistence of the characteristics other than beta. Many studies have tried to
differentiate between the two possibilities, but most of them concentrate on the post-
World War I data, usually limited to the USA. On the international front, Haugen and
Baker (1996) find some degree of commonality in the characteristics that are most
important in determining comparative expected returns among different stocks. Because
of this, testing the asset pricing models on a new data set provides an obvious way to
distinguish between the data snooping hypothesis and the persistence of the
characteristics identified to predict returns. This is to test whether the anomaly exists in a
new and independent sample. This dissertation fills these gaps in the literature by
introducing an independent data set and a different set of characteristics to test the cross-
sectional predictability of stock returns.
3

Several characteristics have been studied in the literature to predict stock returns. Among
them are size (market capitalization: measured as price times shares outstanding), book
to market value ratio, momentum, return reversal, price earnings ratio, dividend yield,
accruals, illiquidity, net issues, asset growth, profitability, total risk (idiosyncratic risk),
etc. Subrahmanyam (2010) categorizes the origins of the predictive characteristics based
on four principles. These are:
- Theoretical motivation based on risk-return model variants.
- Informal Wall Street wisdom (such as value and size investing)
- Predictors originated from the behavioral biases of investors.
- Models that include market frictions (illiquidity).
The theoretically motivated risk/return models conduct a test to see whether a higher
return has been associated with higher risk (measured as beta). Black, Jensen and Scholes
(1972), Sharpe and Cooper (1972) and Fama and MacBeth (1973) find support for the
CAPM in the 70s. On the contrary, Fama and French (1992) find that their data do not
appear to support the CAPM. The traditional CAPM argues that only market risk should
be incorporated into asset prices and command compensation. However, Malkiel and Xu
(2006) document that the CAPM may not hold when some investors are not able to hold
the market portfolio, due to various reasons such as transaction costs. Investors inability
to hold market portfolios will force them to care about the total risk, not simply the
market risk, as implied by the CAPM. Hence, idiosyncratic risk would be priced in the
market. In effect, there is a positive relationship between idiosyncratic risk and average
returns. On the contrary, Ang, Hodrick, Xing and Zhang (2006) find a negative
relationship between idiosyncratic risk and expected returns. The negative relationship
4

between idiosyncratic risk and expected returns remains debatable, since no theoretical
framework has been established to determine the source. Idiosyncratic risk may also
limit arbitrageurs from exploiting mispricing opportunities on the market. That is, for
high idiosyncratic risk stocks, it is a challenge to execute arbitrage activity that is free
from idiosyncratic risk; hence, such stocks are more likely to trade at a price far from
their fundamental values. Doukas, Kim and Christos (2010) find a positive relationship
between mispricing and idiosyncratic risk, which is consistent with the limited arbitrage
argument. The arbitrage theory posits that mispricing can persist whenever the cost of
arbitrage exceeds the benefit. In addition, as a specific example, McLean (2010) finds a
strong reversal mispricing in high idiosyncratic risk firms.
The Wall Street wisdom characteristics are just found by chance or motivated by
informally appealing to the knowledge of finance professionals. The premise of these
characteristics is not based on any prior theoretical reasoning. For example, Basu (1977)
documents the negative relationship between price/earnings ratio and abnormal returns.
He partially based his assertion on the notion that recommending stocks based on
price/earnings ratio is common on the Wall Street. Similarly, Banz (1981) documents
that stocks with low market capitalization outperforms those with high market
capitalization. The literature on the informal Wall Street characteristics is given a
magnificent boost by Fama and French (1992). They document the importance of size
and book/market value in the cross-section of expected stock returns. In addition, they
show the CAPM is not supported in their data. Fama and French (1993) build on the role
of size and value to postulate that a three factor model based on factors formed on size,
book/market value characteristics and the market returns can explain expected returns.
5

On the return predictions based on past performance, Jegadeesh and Titman (1993)
document the prediction of three to twelve months of past returns. They find that stocks
or portfolios that have performed well in the past 3 to 12 months will continue to do so in
the next 3 to 12 months (momentum). However, the source of the momentum effect is
subject to debate. Conrad and Kaul (1998) and Bulkley and Nawosah (2009) argue that
the effect is mainly due to the cross-sectional variation in expected returns. On the
contrary, Jegadeesh and Titman (2001) do not only address the data-mining critique in
explaining momentum effect but also document that models of investor behavioral bias
offer a good explanation of the momentum effect. Their argument is because momentum
profit is due to delay in overreactions that are eventually reversed. However, they
indicate the support for the behavioral model should be tempered with caution. Others
believe momentum effect can be explained by the state of the market (Cooper, Gutierrez
and Hameed (2004) and Chabot, Ghysels and Jagannathan (2009)).
The characteristics derived from behavioral biases or cognitive challenges are based on
informal arguments about investor overreaction and underreaction to information. The
premise of the behavioral notion is that the conventional financial theory ignores how
investors take decisions. Daniel, Hirshleifer and Subrahmanyam (1998) argues that if an
investor places more weight on his private information signal, it causes stock price
overreaction. As information that is more public arrives, the prices move closer to their
fundamental values. The overreaction correction pattern is consistent with a long run
negative autocorrelation in stock returns (long-term reversal in returns). On the other
hand, if the investor begins with unbiased beliefs about his private information signal, he
tends to see a new public information signal as a confirmation of his private signal. This
6

suggests that the public can also trigger further overreaction to the preceding private
signal, thereby causing momentum in security prices. In effect, the biased self-attribution
implies the short-run momentum and long-term reversal. With the notion of investor
reactions to value, Cooper, Gulen and Schill (2008) show that growth in book assets are
cross-sectionally related to future returns, and the implication is that investors underreact
to information in the time series of balance sheets.
Investors require compensation for market friction (illiquidity). Hence, market friction is
a predictor of stock returns. The most difficult issue in relating market illiquidity to
expected returns is the measure of illiquidity. There is a pack of measures in the literature
related to illiquidity. Some of these include the bid-ask spread, absolute return to the
dollar trading ratio, relationship between price changes and order flows, share turnover,
the proportion of zero returns, market capitalization, etc. Amihud and Mendelson (1986)
find a significant premium for the bid-ask spread measure. In a reaction to the possible
problem associated with providing a consistent liquidity measure for all markets, Amihud
(2002) proposes price impact measures, defined as the absolute value of stock returns
divided by the dollar volume. He finds a positive relationship between his illiquidity
measure and average return. From emerging markets, Bekaert, Harvey and Lundblad
(2007) use the proportion of zero returns as a measure of illiquidity to establish a positive
relationship with expected returns.
1.1.1 The aim of this Dissertation
In this dissertation, we study the robustness of the cross-sectional patterns in stock
returns, using the completely independent database of the Brussels Stock Exchange (BSE
from here onwards) in the 19
th
century and the first few years of the 20
th
century. Apart
7

from the quality of the data set, it allows us to study the influence of strong varying
conditions in the economic and institutional environment on stock returns. The period of
the data set avoids the data mining critique. With this data,
- We will test the validity of the CAPM.
- In addition, we test whether other characteristics can explain the cross-sectional
variation in stock returns.
As of the time of writing this dissertation, accounting and transaction data has not been
digitalized for the 19
th
century BSE, so we will not be able to investigate the
predictability of accounting-related characteristics on stock returns. However, we
investigate whether size, momentum (short run past returns), total risk (firm specific risk)
and dividend yield (an asset value' indicator like the book-to-market value ratio) can
cross-sectionally predict returns.
With the importance of illiquidity, in the absence of volume and transaction data, we
follow Bekaert et al. (2007) to measure illiquidity as a proportion of zero returns in the
last 30 months. However, our unreported results indicate that zero-return illiquidity
measure is strongly correlated with market capitalization (also a proxy for liquidity) and
stock price level.
Data Description and Methodology 1.2
It is worthy to reiterate that empirical research on the stock market return predictability
has a long tradition. The large body of the financial literature in this area uses data from
the Center for Research in Security Prices (CRSP). The CRSP data consist of
comprehensive and accurate historical returns for all stocks listed on the NYSE, the
8

Amex and the NASDAQ stock markets in the USA. Researchers make use of the long
time series (NYSE from 1926 to date, Amex and NASDAQ from 1962 and 1973,
respectively to date) and high quality financial data to carry out research about general
equilibrium asset pricing models and predictable patterns in returns, among others.
However, the obsessed and continuous search for predictable patterns in a single data set
will likely reveal an interesting (spurious) pattern (Lo and MacKinlay (1990)). Since one
of the most studied quantities on the market to date is the stock return, the tests on
financial asset pricing models seem especially at risk. Recent anomalies detected in
empirical studies call for researchers to suggest a modification to standard economic
theories about asset prices. However, the CRSP data are mostly used in asset pricing
research due to the non-availability of reliable and independent data sources. There is a
considerable danger of inducing data mining/snooping bias when a single data set is used
repeatedly. This may attenuate the reliability of statistical analysis on the data. Data
from financial markets of other countries are normally used to investigate asset return
predictability. However, common characteristics have been identified that can predict
returns on contemporary markets.
To reduce the doubts regarding inferences of the test of asset pricing predictability, we
analyze a completely new and unique historical data set of stock returns from the BSE
during the 19
th
century and the beginning of the 20
th
century. The data set was
constructed at the University of Antwerp in Belgium (Studiecentrum voor Onderneming
en Beurs (SCOB)). The data for our study start from 1832 and ends at 1914, just before
the outbreak of World War I. During the World War I period, the BSE was closed and
this can be regarded as a natural breaking point of the long time series of the stock
9

returns. The BSE was considered one of the biggest markets in the world at that time,
because Belgium was one of the first nations on the European continent to become
industrialized (see Van der Wee (1996) and Neymarck (1911)). On the industrial output
per head ladder, Belgium stood second after Britain in 1860, and third in 1913, after the
UK and the USA (see Bairoch (1982)). During this period, highly developed banking
system coupled with liberal stock market regulations attracted a great deal of domestic
and foreign capital in Belgium. In confirmation, Van Nieuwerburgh, Buelens and
Cuyvers (2006) document that the development of the financial sector, accompanied with
the stock market-based financing of firms, played an important role in the economic
growth of 19
th
century Belgium.
Figure 1.1: Number of listings on the 19
th
Century Brussels Stock Exchange

Source: Annaert, Buelens and De Ceuster (2004)
Figure 1 indicates the number of listings on the 19
th
century BSE. The figure depicts the
rising popularity of the BSE. The bold line shows the number of common stocks, bonds
10

and preferred stocks. The evolution of common stock listings is represented with the
dashed line. It is clear from the figure that the BSE almost exclusively lists common
stocks until the mid-1850s. The importance of the common stocks declines to about 65%
just before the World War I. To illustrate the international attraction of the BSE, the
faint line shows the common stock listings of foreign companies. About one fifth of the
common stock listings from the late 1870s are for foreign companies. The number of
common stock listings on the BSE is quite important, even if we limit our attention to
Belgian common stocks. From 1868 onwards, at least 100 Belgian common stocks are
listed, gradually increasing to about 600 just before the World War I.
Figure 1.2: Total market capitalization of the BSE

Source: Annaert, Buelens and De Ceuster (2004)
It is obvious from Figure 2 that increasing number of common stock listings increases the
market capitalization of the BSE. Notably, the financial liberalization from the early
1970s onwards reflects in the growth of the total market capitalization. Annaert, Buelens
and De Ceuster (2004) shows that in the 19
th
century stock market capitalization was
11

27% of gross national product in 1846, 57% in 1880 and steadily growing to 80% by the
end of the period of our study.
Annaert et al. (2004) classified all companies listed on the BSE into five categories,
based on the geographical location of the major production facilities and the companys
country of residence. However, in this research, we restrict ourselves to the analysis of
data from the securities of Belgium owned companies. The stocks on the BSE are well
diversified across industries such as transportation, mining and extraction, financials,
utilities and industrials. The BSE provides monthly data on stock prices, dividends and
the number of shares outstanding for more than 1000 different companies officially
quoted on the exchange. This enables the computation of the market capitalization and
total returns for individual common stocks on the exchange. The total returns of the
common stocks on the BSE were computed by considering dividends, stocks splits,
mergers, delisting and other capital operations. The data permit us to test the robustness
of the predictive power of characteristics on common stock returns, and the analysis of
the anomalies identified in the existing literature on asset pricing. All available
information begins in January and ends in December of each year.
1.2.1 The Risk-free rate and the Market Index
We use the annualized short rate converted to a monthly rate as a proxy for the risk-free
rate. The short rate is based on the commercial rate from the official quotation list of the
Antwerp Stock Exchange. The rate is extracted from the newspapers in the period of our
study.
2
For the market index, we employ the all share monthly value weighted index
constructed by Annaert et al. (2004). We concentrate on the value-weighted index as it

2
Journal du Commerce, dAnvers, LAvenir, Moniteur des Intrts Matriels and Het Handelsblad.

12

mirrors the return evolutions of the investable assets' universe. The value-weighted index
is constructed by considering the market capital of the individual stocks in the index
portfolios. The index represents the total returns of the stock market investable assets.
Figure 1.3: Evolution of the value weighted market index in the two sub-periods of our study

It is clear from Figure 1.3 that, 100 Belgian Franc invested in the index would have
grown to about 500 Belgian Franc at the end of the period of strict regulation and
industrial revolution (Jan.1832-Dec.1867). Evidence from the Figure depicts a market
0
50
100
150
200
250
300
350
400
450
500
1830 1835 1840 1845 1850 1855 1860 1865 1870
Value-weighted index evolution in the period of industrial
revolution: Jan. 1832-Dec.1867
0
200
400
600
800
1000
1860 1870 1880 1890 1900 1910 1920
Value-weighted index evolution in the period of
deregulation and expansion: Jan. 1868-Dec.1913
13

drop in 1948, the year of industrial revolution across Europe. In the period of the
deregulation and expansion period, 100 Belgian Franc invested in the beginning of the
period would have grown to about 1000 Belgian Franc at the end of the period (Jan.1868-
Dec.1913).
Table 1.1: Summary Statistics for Risk-free rate and the Value-Weighted Index

Table 1.1 shows the time series summary statistics of the risk-free rate and the value-
weighted index. As indicated in Annaert et al. (2004), the period from 1832 to 1913 is
divided into two sub-periods based on the environment in which the BSE was operating.
These were the periods of the industrial revolution and high regulation (1832-1867) and
the period of deregulation and expansion (1868-1913). We show the summary statistics
in the entire period and both sub-periods. The average risk-free rate, as well as the
average value weighted return for the entire period, is almost the same in the two sub-
periods. On the other hand, the standard deviation of the value-weighted index is high in
the first period as compared to the second period. The high standard deviation and
Standard 1st Order
Mean Deviation Skewness Kurtosis Min Median Max Autocorr.
Risk-free 0.26% 0.08% 0.59 3.13 0.12% 0.25% 0.60%
Value-Weighted Index 0.40% 2.42% 0.02 20.92 -20.68% 0.27% 15.78% 0.28*
Risk-free 0.28% 0.07% 0.19 2.54 0.17% 0.29% 0.53%
Value-Weighted Index 0.41% 3.01% -0.15 18.61 -20.68% 0.21% 15.78% 0.30*
Risk-free 0.24% 0.08% 0.99 4.26 0.12% 0.23% 0.60%
Value-Weighted Index 0.40% 1.86% 0.55 9.28 -8.16% 0.34% 11.99% 0.25*
autocorrelation of the value-weighted index series of the risk-free rate and the value-weighted market index.The Ta-
ble also displays the descriptive statistics for two subperiods.That is the period of the industrial revolution (1832-1867)
and the period of deregulation and expansion (1868-1913). * = significantly different from zero at 5% (two sided).
Jan. 1832-Dec. 1913
Jan. 1868-Dec. 1913
Jan. 1832-Dec. 1867
Note: In this Table, we compute the descriptive statistics of the risk-free rate and the value-weighted market index.
We calculate the time series mean, standard deviation, skewness, minimum, median, maximum and the first order
14

kurtosis in the first period are not surprising, as the period is characterized with (i) a large
market drop in 1848, a year of revolution across Europe (although not in Belgium) and
(ii) strong restrictions on joint stock holdings that may affect the volatility of prices
(Annaert et al. (2004)). The standard deviation recorded on the 19
th
century BSE is low
compared to the high values recorded on the USA market. However, it is close to the
values recorded on the UK market in the period 1870-1913 (see Grossman and Shore
(2006) page 281, Table 2).
The average difference between the value weighted index and the risk-free rate is
between 0.13% and 0.16% for the overall period and the two sub-periods. This implies a
risk premium of 1.92% per annum. This might be too low compared to the annual risk
premium of 4.61% recorded on the USA markets. However, the low risk premium
recorded on the 19
th
century BSE is not surprising as Grossman and Shore (2006) record
far too low annual risk premium on the UK market between 1870 and 1913 ( risk
premium of 0.53% and 1.93% for value weighted and equally weighted respectively).
1.2.2 Methodology
As shown in the diagram below, basically, there are two methods for testing the cross-
sectional predictability of stock returns. These are the sorting and the regression method.
The sorting method is sub-divided into the univariate, independent and conditional
double sorts. The univariate sort method ranks stocks on beta, or their characteristic, and
groups them to form portfolios. The average returns of the portfolios should be a
monotonic function of the characteristic. The method tests the significance of the
difference in average returns between the high and low ranked portfolios. A univariate
sort has the disadvantage of not being able to disentangle the effect of two (or more)
15

characteristics. To distinguish between the effects of two characteristics on the cross-
section of stock returns, the conditional double sort is applied. In the conditional double
sort method, stocks are first ranked on a single characteristic and split into groups. The
stocks in each group are then sorted on the second characteristic. The conditional double
sorting method keeps one characteristic constant and tests the effect of the other
characteristic on the cross-section of expected returns.

Fama and French (1992) adopted the conditional double sort to disentangle the effects of
beta and size on average returns. The independent double sort method simultaneously
tests the effect of two characteristics on the cross-section of stock returns. In such a case,
one characteristic is expected to predict stock returns better, holding the other constant
and vice versa. Empirically, stocks are separately ranked on each characteristic and
grouped. Portfolios are formed from the stocks that exist at the intersections of the
groups. Asness (1997) used the independent double sort method to test the interactive
effect of value and momentum on the cross-section of stock returns. The advantage of the
Cross-Sectional
Predictability of
Stock Returns
Sorting
Regression
Independent
Double Sort
Conditional
Double Sort
Time Series
Cross-Sectional
Univariate
sort
16

sorting methods is that they do not impose any functional relationship between
characteristics and expected returns. The method answers the question of whether the
high-ranked characteristic portfolio outperforms the low-ranked characteristic portfolio.
As stated before, the method test is based on whether the difference in the average
returns of the high and low- ranked portfolios are significantly different from zero. The
sign and the significance of the difference in average returns determine whether the
relationship is negative or positive. However, it is weak to use the information on the
high and low ranked portfolios to represent the cross-sectional effect of a characteristic
on expected returns. In addition, it is difficult to disentangle the marginal effect of
several characteristics because the sorting method cannot sort on more than three
characteristics. It also ignores the noisy nature of the stock or portfolio returns.
The regression method addresses some of the shortcomings of the sorting methods. The
two types of regressions used to test the cross-sectional predictability of stock returns are
the time series and the cross-sectional regression method. Black et al. (1972) initiated the
time series method, which was later developed by Gibbons, Ross and Shanken (1989).
The method is based on time series regression, for which the intercepts are tested to
determine whether they are significantly different from zero. The method runs the
regression of each asset base on the equation
,
jt j j t jt
r b F o c = + + (1)
where r
jt
is the time series excess returns of asset j (usually portfolio of assets). F
t
is the
factor portfolio, usually comprised of the market excess returns and portfolio returns,
formed from the difference in the high- and low-ranked characteristic portfolios.
j
, b
j

and
jt
are the regression coefficients and the regression error. The estimated intercepts
17

are grouped together to form a vector. The intercepts represent pricing errors. If the
model can explain the cross-sectional variation in returns, the intercepts should not be
jointly significantly different from zero. Gibbons et al. (1989) suggest computing the
statistic
( )
1
' 1 2
1
1 ~ ( , 1) ,
T N
SR F N T N
N
o o

+
where T is the length of the time series, N is the number of assets or portfolio of assets,
o is vector of intercepts from the regression equation above, is the covariance matrix
of the residuals from the regression above and SR is the Sharpe ratio of the factor. Under
the null hypothesis that the intercept equals zero, the method compares the statistic to the
F- distribution with degrees of freedom N and T-N-1 (F (N, T-N-1)).
Unlike the time series method, which requires regressors that are also returns, the two-
pass cross-sectional regression method can take on factors that are not returns (see Fama
and MacBeth (1973)). The method often involves two regression stages (hence two-pass
regression) because of the test of CAPM. The first pass regression (time series) estimates
betas, which then serve as input for the second pass cross-sectional regression. In the
second pass regression, several characteristics can be combined as independent variables
or regressors. That is, for each time t, the cross-sectional regression method runs a
regression of the form

1 1 , 1
2
,
C
jt ot t jt vt v jt jt
v
r | c

=
= + + I +
(2)
where r
jt
is the cross-sectional excess returns of all assets at the time t, and
jt-1
are the
time-varying beta estimates.
v,jt-1
can be any characteristic determined prior to t.
jt
is the
18

error term. ( )
0
, ,
t Ct

is a vector of regression coefficients. Using the time series of the
regression coefficients, we estimate their expected values and test if they are significantly
different from zero. If the CAPM holds, the expected value of
0t
should be zero and the
expected value of
1t
, which is the risk premium on the market should be significantly
positive. Any characteristic that can explain the cross-sectional variation in stock returns
has an average coefficient significantly different from zero. We now form the t-test
( )
( )
v
v
v
t

o
= (3)
where
1
1
T
v vt
t
T

=
=

and
( )
v
o is the standard error normally adjusted for the Newey
West heteroskedastic and autocorrelation consistency (HAC). The Newey-West standard
error is estimated as
( ) ( )
'
0
1
1
1
q
v w w
w
w
q
o
=
| | (
= u + u +u
|
(
+
\ .
where

( )( )
w vt vt vt w vt
E
u = (

and q is the number of lags.
Betas can be estimated with a rolling window or the full window regressions. The rolling
window regressions estimate betas with information in the period prior to t. The betas are
then used in the cross-sectional regression for the subsequent period. There is a tendency
to measure betas with error in the first regression. Using estimated betas in the second
pass regressions therefore induces an error-in-variable problem, biasing the slope
coefficient in small samples. Estimating betas over the full sample period mitigates the
measurement error, as the time series sample size T increases. In this case, the second
cross-sectional regression estimate is T-consistent. In the full window regression method,
portfolios are formed by sorting stocks on betas estimated in the period prior to t.
19

Portfolios are then rebalanced by repeating the process of beta estimation and portfolio
formation for each time t in the period studied. Finally, portfolio returns are computed.
Full sample betas are estimated using these time series of portfolio returns. Based on the
assumption that betas are constant throughout the sample period, the full sample betas
can be used in the cross-sectional regression (see Ibbotson, Kaplan and Peterson (1997)).
On the other hand, asset beta varies with time, so each time, the full sample betas can be
assigned to individual stocks in the portfolios. The main assumption underlying this
approach is that stocks in a portfolio have the same exposure to long-term systematic
risk. This approach will yield N and T consistent estimates, since it uses the full
information in the sample period and the cross-section. Fama and French (1992) utilized
the full-sample beta estimate in the test of the CAPM.
A second approach to minimize the error-in-variable problem induced by beta or any
other estimated variable is to multiply the denominator in equation 3 by a correction
factor. Shanken (1992) suggests an adjustment factor
( )
2
0
2
1
m
m
r
r
o
| |
+ |
|
\ .
. Where
2
m
r
o is the
standard deviation of the market returns in excess of the risk-free rate,
m
r is the average
excess market returns and
0
is the average intercept series from the cross-sectional
regressions.
In this dissertation, we will use all the possible sorting methods depending on the task.
For instance, we will use the double sorting methods to disentangle the marginal effect of
each characteristic on average returns. The sorting method provides the advantage of not
placing any linear restrictions on the return/characteristic relationship. To test the effect
20

of several characteristics on returns, we will adopt the cross-sectional regression method.
This will serve as a confirmation of the result in the sorting method.
For all portfolio formations we adopt the Fama and MacBeth (1973) breakpoint method.
That is, if N is the number of stocks in year t and nis the number of portfolios required,
stocks are allocated to int( / ) N n portfolios, where ( ) int / N n is the nearest integer less or
equal to / N n. The middle portfolios have ( ) int / N n stocks each. If N is even,
( )
( )
int / 1/ 2 int / N n N n N n ( +

stocks will be allocated to the first and the last
portfolio. If N is odd, one stock will be added to the last portfolio. The method
sometimes allocates more stocks to the extreme portfolios, which are of a great deal of
interest because of the formation of the hedge portfolio (top-ranked portfolio returns,
minus the bottom-ranked portfolio returns). It also ensures that no stock is lost in the
portfolio formation process.
Summary of the chapters 1.3
In this doctoral research, we conduct four empirical studies with the BSE data from the
19
th
and the beginning of the 20
th
century. We cover the assessment of beta in chapter 2.
In chapter 3, we test the validity of the CAPM and the size effect. The fourth chapter
tests the momentum effect. The combined effects of Size, Momentum, Total Risk and
Dividend yield on the cross-section of stock returns is studied in chapter 5. The last
chapter concludes the dissertation.
1.3.1 Chapter 2
In the second chapter, we focus on the assessment of beta. Beta serves as the main input
of the CAPM. It is an estimated variable that can be measured with error, thereby biasing
21

the test results of the CAPM. Beta instability, bias and non-robustness to outliers have
become the center stage of research since the development of the CAPM. The chapter
studies the relative performance of different methods of estimating beta, based on their
ability to predict the subsequent beta. Specifically, we compare the market model betas
to the betas estimated with the Blume (1971) and Vasicek (1973) auto-regressive
techniques.
We show that the individual stock market model betas are not stable. The predictability
of the market model betas can be improved by forming portfolios with at least ten or
more stocks. Most strikingly, there is no significant difference in the predictive accuracy
of Blume and Vasiceks adjusted betas. Using Dimsons method of estimating betas,
small numbers of stocks prove to have returns that lead or lag the market returns. To
account for outliers, we use iterative reweighted least square techniques (IRLS) to
estimate betas. The betas from the IRLS are small in magnitude compared to the market
model betas, but they have the same predictive accuracy as the market model. Based on
the study in chapter two, we determine how to estimate beta for the study of the CAPM.
1.3.2 Chapter 3
In the third chapter, we use the sorting and the Fama and MacBeth (1973) (from here on
FM) cross-sectional regression methods to investigate whether the CAPM is valid before
World War I. We will also test if the size effect (the propensity of small size to have
higher returns than large size stocks) exists in the 19
th
century BSE.
With the sorting and the FM cross-sectional regression methods, we find no empirical
validity for the CAPM. We also show that the relationship between beta and returns
varies with time. If anything, the null hypothesis of the equality of the estimated slopes
22

from the cross-sectional regression and the market excess returns is not rejected in the
period 1868 through to 1893. On the contrary, using size-sorted portfolios (equally
weighted) in the cross-sectional regression establishes the beta-excess return relationship
for beta as the lonely regressor (the CAPM is valid). However, the conditional double
sorting method adopted by Fama and French (1992) separates the effect of beta and size
on expected returns. The method makes the average cross-sectional slope of the beta
insignificant, whether placed alone or together with size in the regression.
We find size to be negatively related to excess returns. Size is negatively significantly
related to excess returns (size effect), but beta does not relate to excess returns when
placed together in the cross sectional regressions. Detailed analysis of the data reveals
that the size effect is mainly due to small size stocks, which accounts for about 0.35% of
the total market size. Eliminating these small stocks destroys the relationship between
excess returns and beta or size. Both sorting and cross-sectional regression reveals that
the size effect disappears when stocks are value weighted to form portfolios. In
summary, the CAPM did not work in the 19
th
century BSE. Estimating betas with the
market model, the Dimson and Vasicek method will not establish the model. The size
effect exists, but it is mainly due to a small group of stocks that represent a very small
portion of the total market capitalization.
1.3.3 Chapter 4
In chapter 4, we investigate whether the momentum strategy can generate abnormal
profit in the 19
th
century BSE. The momentum strategy buys stocks that have performed
well in the past 3-12 months, and sells stocks that have performed poorly in the past 3-12
months. There is convincing evidence that the strategy is profitable on the 19
th
century
23

BSE. Finding momentum in this era confirms the assertion that the momentum profit
found on the post-World War II markets is not mainly due to data-snooping biases.
Detailed analysis reveals that the momentum effect does not exist in the small size group
of stocks. Additional investigation into the momentum profit in each calendar month
shows that the profit was positive for all months. The January reversal effect, found in
the post-World War II USA markets, cannot be found in the 19
th
century BSE. In fact,
January records the fourth highest momentum profit relative to the other months of the
year. We find that momentum profit is not strong in the first twenty years of our study
period.
In order to investigate the source of the momentum profit, we use the approach of
Jegadeesh and Titman (2001) to study the returns of the momentum portfolios in the
post-holding period. The momentum returns reverse in the second to fifth years after the
holding period. Further study reveals the reversal is mainly due to small size stocks. We
also test whether the momentum profit and the long run reversal in the cross-section of
stock returns depends on the state of the market. The 6-month formation and the 6- to 12-
month holding period strategies are profitable solely in periods of market gains.
1.3.4 Chapter 5
In this chapter, we investigate whether size, momentum, total risk and dividend yield can
explain the cross-section of stock returns between the years 1868 and 1913. We repeat
size and momentum characteristics in the analysis to test the pervasiveness and the
combined effect of these characteristics on the cross-section of excess returns. The size
sort also tests whether the effect of the other characteristics are not confined to the small,
illiquid group of stocks. Total risk for each stock is measured as the standard deviation of
24

the past 24 to 60 months excess returns. Each year, dividend yield for a stock is the sum
of all dividends paid within the last 12 months divided by the current month price. We
investigate whether the relationships are pervasive across all dividend yield, size, total
risk and momentum groups. Sorting and FM cross-sectional regression methods are
adopted in the analysis. We confirm that size has a negative significant relationship with
excess returns. However, repeating the analysis on different size groups reveals that the
negative relationship is completely driven by micro-size stocks accounting for less than
3.67% of the market capital. This confirms the finding in Chapter 3 that the size effect is
mainly due to the first decile size portfolio, which forms part of the micro size group. We
did not find a consistent relationship for total risk. Momentum shows a consistent
positive relationship with excess returns in stocks, which accounts for more than 96% of
the market capital. We find a negative relationship between dividend yield and
momentum among dividend-paying stocks, but each of them is positively related to
average excess returns. Dividend yield and momentum are positively related to the
average excess returns in our large stocks, which account for about 96% of the market
capital.

25

CHAPTER 2
2 ASSESSMENT OF BETA IN THE 19
th
CENTURY BSE
3

Beta risk estimation and the testing of asset pricing models have a long tradition in
financial literature. Beta serves as the main input of the CAPM, and it is assumed stable
in empirical applications. However, the instability, bias and non-robustness (to outliers)
of the beta have raised concerns in the literature. The notion of its estimation is of
fundamental importance to the testing of CAPM. The objective of this chapter is to
introduce the new data set from the 19
th
century BSE to test the performance of the
alternative techniques of estimating beta. Studying the various techniques of estimating
beta, we may determine relatively unbiased and stable beta for the test of CAPM, which
posits a positive relationship between the beta and expected returns.
Introduction and Literature Review 2.1
Beta stability, bias and robustness evaluations have become the center stage of research
in finance since the development of the CAPM by Sharpe (1964), Lintner (1965) and
Mossin (1966). Beta, one of the parameters of a time series' regression, plays a key role
in CAPM applications. Beta is usually estimated with the standard market model (MM).
The MM is a statistical model that relates the return of any given stock to the return of
the market index. Therefore, for returns of a stock j in period t we write,
,
jt j j mt jt
R R o | c = + + (4)

3
This chapter was presented as a paper at the First World Finance Conference on the 27
th
May 2010 at
Viana do Castelo, Portugal (blind reviewed)
26

where
( )
2
jt j
Var
c
c o =
4
,
jt
R and
mt
R are the period t returns on the stock j and the
market index return, respectively. The parameters
j j
| o , and
2
j c
o are the parameters to
be estimated from the MM peculiar to stock j . The MM assumes that
j
| (beta) is
constant over the estimation period. However, substantial evidence in the financial
literature has established that security betas are not stable over time. Blume (1971) and
(1975) pioneered the research on beta stability based on the cross-sectional correlation
between beta estimates for successive periods. He finds portfolio betas having stronger
correlation across successive periods than individual security betas. This indicates that
portfolio betas are more stable than individual stock betas. Blume (1971) and Vasicek
(1973) show that betas are not stable, but they revert towards their cross-sectional mean.
They propose autoregressive models to capture this variation in beta. The autoregressive
method adopted by Blume is the cross-sectional regression of betas in period t on the
betas in period t-1. That is,

1
,
jt jt jt
a b | | q
= + + for j=1,,N (5)

N is the number of stocks in the cross-section. a and b are the intercept and the slope of
the cross-sectional regression of betas in the period t on betas in the period t-1. The
jt
q is
the error of the cross-sectional regression of betas in successive periods. The objective of
Blumes autoregressive model is based on the tendency of betas from successive periods
to revert to the mean. In effect, the method adjusts historical betas towards their cross-
sectional mean. The Vasicek autoregressive method not only adjusts betas toward their
cross-sectional mean, but the adjustment also depends on the uncertainty (standard error)

4
( ) x E is the expected value of x , ) (x Var is the variance of x.
27

about beta. Betas estimated with high standard errors have a greater tendency to deviate
from the cross-sectional mean of the betas. Therefore, high uncertainties call for greater
adjustment. Vasicek (1973) applied a Bayesian correction method to capture the
differences in standard errors. The Bayesian correction method places weights on the
cross-sectional average beta and the assets beta estimate. The weights sum up to one,
and the more the uncertainty surrounding either estimate of beta, the lower the weight
placed on it. He used cross-sectional information from the previous period betas. That is,
beta is estimated with the model

( )
( )
( )
2
1
1 1
2 2
1 1
var
for 1, 2, , ,
var var
jt
jt jt jt
jt jt
j N
|
| |
|
o
| | |
| o | o

= + =
+ +
(6)
where
jt
| is the mean of the posterior distribution of beta for stock , j which serves as
the beta forecast.
2
|
o is the standard error from the estimation of the market model
regression coefficient,
1 jt
|

.
1 jt
|

is the cross-sectional mean of betas in period 1 t , and
( )
1
var
jt
|

is the variance of the cross-section of betas. Faff and John (1992) use
Australian data to confirm the evidence in beta nonstationarity. They show that there is a
strong connection between beta instability and the market index used to estimate beta. On
the contrary, Gregory-Allen, Impson and Karafiath (1994) use daily data to document
that portfolio betas are not more stable than individual security betas on the US market.
They show that, accounting for the variance around beta estimation will not make
portfolio betas more stable than individual stock betas.
Also, the literature has shown that nonsynchronous trading biases betas estimated with
the MM. When stocks are ranked based on their market capital (size), beta is biased
28

downward for small size stocks and upward for large size stocks. Scholes and Williams
(1977), Dimson (1979) and Fowler and Rorke (1983) developed methods to correct the
bias. Dimson (1979) adopted the aggregate coefficient method, which computes the lead
and the lag betas in a multiple regression of the stock return on a number of lead/lag
market returns. The sum of the estimated beta coefficients from the multiple regressions
is Dimsons estimate of beta. Dimsons technique involves estimating a multiple
regression of the form
,
l
jt j j mt i jt
i l
R R o | c
+
=
= + +
(7)
Dimsons beta estimate is then given by
dim
l
i
i l
| |
=
=
. The error term

jt
c follows the
assumptions of the classical linear regression model. On the other hand, Scholes and
Williams (1977) compute the lead and lag beta coefficients by univariate regressions.
The literature has found that the bias in beta estimates is due to autocorrelation in returns
caused by infrequent trading. Ibbotson et al. (1997) found that beta estimates for small
stocks are severely biased downwards. This implies that the returns of small stocks are
more capable of exhibiting autocorrelation than large stocks. They recommend the
inclusion of the lagged market returns in the estimation of beta when securities are traded
infrequently. On the other hand, Bartholdy and Riding (1994) found that the MM betas
are less biased, more efficient and consistent on the New Zealand Market than the
Dimson, Scholes and Williams biased correction model betas.
The literature also highlights the impact of outliers (extreme observations) on the beta
estimates. Outlier observations in returns, depending on their location in the stock return/
29

market return plane, can have an extreme influence on beta when the MM is used to
estimate beta. Under idealized conditions, the MM beta produces the best estimate
because of the ordinary least squares method. The idealized conditions are that the paired
returns of the stock and the market conform to the linear model relationship with zero
mean error not correlated with the market returns. Unfortunately, these conditions often
fail in empirical settings. One explanation of the failure of the condition is the occurrence
of outlier observations in returns. Nevertheless, the MM does not consider the effect of
outlying observations on beta. Chatterjee and Jacques (1994) document the effect of
outlying observations on the beta parameter. Their detailed study compares the betas
estimated from the MM and the outlier resistance methods. They indicate that the MM
cannot detect a large number of outlying returns in their data. Identifying the outlying
observations and estimating betas with the outlying resistance method yield beta
estimates that are lower than the MM betas. Chan and Lakonishok (1992) said the robust
method (outlier detecting method) of estimating beta is good for stocks susceptible to
stock splits, dividend cuttings, and initial public offerings (IPO). Stocks that exhibit this
behavior are mostly small stocks. Martin and Simin (2003) confirmed that outlier
resistance beta is a better predictor of future beta than the market model beta. They also
reveal that small stocks betas are most vulnerable to outliers.
The literature proposes various techniques of minimizing the impact of outliers on beta
estimates. The weighted least square estimation technique is one of the methods used to
minimize the impact of outliers on beta estimates. Based on the assumption that outliers
have a lower probability of occurring in the future, less weight is placed on them (see
Martin and Simin (2003)). An alternative technique would be to remove the extreme
30

observations and perform the regression with fewer observations. The least squares
approach adopted by the MM minimizes the sum square residuals with respect to the
model parameters
j
o and
j
| :
( )
2
,
min ,
j j
jt j j mt
R R
o |
o |
The squaring of the residuals magnifies the effect of the outliers on the estimated
parameters. For instance, expected return is likely to shift toward the outliers while the
covariance matrix will be inflated toward the outliers. To reduce the influence of outliers,
the statistics literature emphasizes the use of iterative reweighted least-squares (IRLS)
method. This method estimates beta by iteratively minimizing a weighted function that
depends on standardized return residuals:

,
min ,
j j
j
jt j j mt
R R
W
o |
c
o |
o
| |

|
|
\ .
(8)
with weight function W and
j
c
o the standard deviation of the return residual. We
estimate the regression parameters from the least squares regressions and use these
parameters as initial input for iteration. A weight function is applied to the standardized
residuals. We apply Huber or the bisquare weight function, defined as:
( )
2
2
1
,
0
jt
jt
jt
jt
s
for s T
T
W s
for s T
(
| |
s
(
|
= \ .
(

>

31

with the standardized residual
j
jt j j mt
jt
R R
s
c
o |
o

= and T , a tuning constant. Huber
(2004) sets the value of T to 4.685. Small values of T

introduce more resistance to
outliers, but at the detriment of efficiency when returns are outlier free. He chose the
value of T so the method will have less influence on beta when there is no extreme
observation and it still provides protection against outliers. For the iterations, the initial
input is the coefficients from the MM. The residual from the MM is standardized and the
weight function defined above is used to transform them. Estimate the parameters by the
weighted least squares:
| |
j
R W W b
0 '
1
0 ' 1
9 9 9 =

where | |
,
mt
R i 9=
with i a vector of ones with the same size as
mt
R (market index returns).
0
W is an initial
standardized weight diagonal matrix and
1
b the estimated beta parameter. The new set
of parameters serves as input for the next iteration. The procedure above is repeated until
the parameter of interest (beta) converges. A standardized residual observation that
exceeds the tuning constant is assigned zero weight (outliers).
The literature also suggests that estimated betas vary with time and macroeconomic
conditions. Therefore, static beta CAPM should replace with time varying beta
conditional CAPM (see Jagannathan and Wang (1996)). In this dissertation, we did not
pursue the time-varying beta because, Ghysels (1998) used monthly NYSE data to
document that pricing errors with static beta models are smaller than those with time-
varying beta models. He indicated that time-varying beta model might be mispecified.

32

To the best of my knowledge, the literature has not traced the stability, bias and
robustness of beta using data before 1926 in any country. Luoma, Martikainen, Perttunen
and Pynnnen (1994) simulated returns data to generate artificial markets. They found
that various beta estimation techniques behave uniquely in different markets.
In this chapter, individual and portfolio betas from the various estimation techniques will
be compared based on their predictive accuracy. The root mean square error (RMSE)
criterion adopted by Blume (1971) and Klemkosky and Martin (1975) is used to
determine the predictive accuracy of the betas estimated from the various techniques.
The 19
th
century data provide a very good platform to determine whether betas before
1914 exhibit a similar pattern as betas after 1914.
This chapter reveals that for individual stocks, the market model betas are weak in
predicting their future. Predictability can be improved when a portfolio of 10 or more
stocks components is formed. The study also shows no significant difference between the
Blume and Vasicek adjusted betas in terms of their predictive accuracy. Estimating betas
with the Dimson method reveals a very small number of stocks showing a lead and lag
relationship with the market returns. The iterative reweighted least squares method
produces betas that have the same predictive power as the MM betas, but are lower in
their magnitude on average.
The subsequent sections of the chapter are organized as follows: Section 2.2 provides
descriptive statistics of beta when estimated with the least squares regressions of the
MM. Section 2.3 examine the Blume correlation technique of detecting the stability of
beta. The autoregressive technique of Blume and Vasicek is used to adjust betas in
33

Subsection 2.3.1. The predictive accuracy of the adjustment techniques based on the
RMSE criterion is tested in the same section. The modified Diebold and Mariano test
proposed by Harvey, Leybourne and Newbold (1997) is used to test for equal predictive
accuracy of the models. In Section 2.4, we determine how the stocks in each period lead
or lag the market index using the Dimsons model. In Section 2.5, we compare outlier
resistant betas to the benchmark MM betas. The final section presents the conclusion.
Beta Coefficient Descriptive Statistics 2.2
The data sample is divided into 15 five-year non-overlapping periods
i
t for 1,...,15. i =
Stocks with complete returns data for five years are considered in each sub-period. For
the proxies of the market portfolio, we consider the value-weighted market index. Table
2.1 presents the number of stocks in each period under study, the statistical
characteristics of the betas estimated in each period and displays the 15 periods studied,
starting from January 1837 to December 1911. The number of stocks with full returns
data in each period ranges from 21 to 424 as shown in Column 2. It can be seen in Table
2.1 that the number of stocks that have five years of data does not exceed 100 before
1877. The beta coefficient of each stock in each period is estimated by simply regressing
the monthly returns of the stock in each period on the corresponding monthly value-
weighted index of the market by using the MM above. Cross-sectional statistics of the
betas in each period are computed and the result is displayed. The equally weighted
average beta and the value-weighted average beta are displayed in Columns 3 and 4,
respectively. We compute the value-weighted mean beta by considering individual stock
relative to market capitalization at the beginning of the period. From the high values of
equally weighted mean beta and low values of value-weighted mean beta in the periods
34

Table 2.1: Beta coefficient descriptive statistics for the 15 estimated periods
Note: This table displays cross-sectional descriptive statistics of betas estimated by market model (1) of returns on their market counterparts over the 15 five-year
sub-periods between January 1832 and December 1911. The number of stocks in the various periods ranges from 21 to 424 for stocks with full return data within
a period. The table also reports the percentage of beta sample in a period that is less than zero. The maximum and minimum beta of each period is also recorded.
The last column reports the average coefficient of determination ( )
2
R . Equally weighted (EW) and value-weighted (VW) mean betas are also displayed in the
table. The value-weighted mean is the average of the betas weighted by their market capital at the beginning of the period.
Percentage
Number of Average Average Standard of BETAS Mean
Period Stocks beta(EW) beta(VW) Deviation <Zero Minimun 0.10 0.25 0.50 0.75 0.90 Maximum R
2
(%)
1/1837-12/1841 21 1.37 1.24 0.67 0 0.38 0.50 0.77 1.35 1.83 2.29 2.90 29
1/1842-12/1846 33 1.10 1.01 1.84 12 -1.18 -0.04 0.27 0.79 1.39 2.21 10.19 18
1/1847-12/1851 32 0.69 1.05 0.72 0 0.04 0.15 0.33 0.48 0.83 1.24 3.92 25
1/1852-12/1856 33 0.76 0.98 0.55 3 -0.43 0.11 0.35 0.75 1.09 1.56 1.98 19
1/1857-12/1861 54 0.98 1.15 0.55 2 -0.19 0.28 0.71 0.99 1.25 1.55 2.74 27
1/1862-12/1866 72 0.83 1.01 0.73 4 -0.86 0.20 0.40 0.70 1.07 1.57 3.70 12
1/1867-12/1871 76 0.74 0.76 0.58 8 -0.91 0.09 0.42 0.67 1.02 1.35 2.36 11
1/1872-12/1877 82 1.05 0.96 1.01 10 -0.71 0.01 0.30 0.85 1.70 2.59 3.55 15
1/1877-12/1881 113 1.29 0.96 1.50 12 -3.54 -0.03 0.32 0.93 1.99 3.14 7.80 17
1/1882-12/1886 154 1.03 1.03 1.44 18 -1.33 -0.35 0.14 0.66 1.66 2.90 7.94 6
1/1887-12/1891 161 1.52 0.74 1.73 14 -0.76 -0.05 0.10 0.94 2.61 4.03 7.81 22
1/1892-12/1896 196 1.24 0.95 1.32 12 -1.74 -0.02 0.34 1.04 1.71 2.87 7.23 8
1/1897-12/1901 252 1.13 0.93 1.11 8 -3.03 0.02 0.27 1.09 1.71 2.38 5.47 14
1/1902-12/1906 374 1.24 1.03 1.36 11 -1.33 -0.08 0.33 1.15 1.74 2.63 8.92 10
1/1907-12/1911 424 1.16 1.05 1.18 9 -2.14 0.01 0.34 1.06 1.64 2.45 7.33 12
Fractiles
35

after 1873, we can conclude that small stocks have high betas in these periods. The
values in the last column show the average coefficient of determination ( )
2
R in
percentages, which is a measure of explanatory power of the MM. It is worthy to note the
percentage of negative betas in nearly all the sub-periods. Surprisingly, the literature
reports negative betas on different markets after World War I. For example, Altman,
Jacquillat and Levasseur (1974), Dimson and Marsh (1983) recorded a large number of
negative betas in their weekly returns interval estimation of beta on the French and the
American markets. It is also important to note the very high and low betas in the 19
th

century BSE compared to the post-1926 market betas recorded in the literature. The
possible explanation for these extreme values of the beta and low coefficients of
determination might be linked to the infrequent trading effect and the influence of
extreme observations (outliers) in the returns series of the stocks. The removal of stocks
that did not trade fully within the five-year estimation periods does not introduce
survivorship bias since there is no significant difference in the beta coefficient
descriptive statistics when stocks with at least 24 observations are added to the sample
(not reported).
In order to capture the effect of size (market capitalization) on the average beta and
2
R
values, we computed the same cross-sectional statistics for betas in each period.
However, the stocks in each period are sub-divided into three mutually exclusive size-
based sub-samples. The composition of the sub-samples is as follows: the sample of
stocks in each period is sorted in descending order based on their market capitalization
(size) at the beginning of the period.

36

Table 2.2: Average beta and average coefficient of determination of the size based sub-samples

Note: In this table, stocks in each period were sorted in descending order based on their market capital.
The first 30 percent of stocks in the periods are classified as large stocks, the next 40 percent as medium
stocks and last 30 percent as small stocks. Cross-sectional statistics of the betas of the sub-samples were
computed and their equally weighted averages of betas and coefficient of determination ( )
2
R are
displayed. The L, M and S subscripts on betas indicate large, medium and small stocks, respectively.

Figure 2.1: The graph of the average beta of each period for large stocks and small stocks

Note: The average beta estimates for large and small stocks are plotted against their corresponding periods
of estimation. The diamond points indicate the averages of large stocks and the square points for small
stocks. This depicts the extreme values of beta estimate recorded for the small stocks relative to the
benchmark beta of one.
Period mean(
iL
) mean R
2
(%) mean(
iM
) mean R
2
(%) mean(
iS
) meanR
2
(%)
1/1837-12/1841 1.29 30.1 1.73 32.8 1.07 23.0
1/1842-12/1846 1.19 36.5 0.79 10.5 1.40 10.1
1/1847-12/1851 0.94 43.2 0.49 21.6 0.67 10.1
1/1852-12/1856 1.05 35.4 0.66 14.2 0.61 10.2
1/1857-12/1861 1.15 40.5 1.08 27.0 0.69 12.9
1/1862-12/1866 1.04 22.6 0.76 8.0 0.70 5.5
1/1867-12/1871 0.77 17.2 0.59 9.4 0.90 7.2
1/1872-12/1877 1.06 19.2 1.02 16.4 1.10 10.0
1/1877-12/1881 1.03 20.2 1.51 18.0 1.30 12.3
1/1882-12/1886 1.01 9.4 1.16 6.1 0.86 3.4
1/1887-12/1891 0.84 18.2 1.73 27.6 1.95 17.7
1/1892-12/1896 1.00 9.5 1.29 7.4 1.40 5.7
1/1897-12/1901 0.98 17.6 1.22 14.6 1.16 9.3
1/1902-12/1906 1.06 15.5 1.27 9.1 1.39 4.3
1/1907-12/1911 1.18 20.1 1.17 11.0 1.13 5.5
Large Medium Small
37

The first 30 percent of stocks are classified as the large stocks portfolio, next 40 percent
as medium stocks portfolio, and the last 30 percent as the small stocks portfolio. Table
2.2 reports the average beta and the
2
R values for each sub-sample in each period.
Generally, the high values of
2
R recorded for large stocks are striking because on the
UK market, Dimson (1979) recorded a similar pattern with 15 years of monthly data.
This might be attributed to the value-weighted index used in the computation, since the
index will have more explanatory power for large-sized stocks. Figure 2.1 depicts the
cross-sectional average beta of each period for large stocks and small stocks. It is clear
from Figure 2.1 that small stocks record comparatively low betas in the periods before
1867 and high betas thereafter. The result in the periods before 1867 corroborates the
assertion of Dimson (1979), Scholes and Williams (1977) and Beer (1997). They said
that if trading frequency is highly correlated with the market capitalization of the stock,
betas of small stocks (infrequently traded stocks) are lower when estimated with the
market model. On the contrary, Ibbotson et al. (1997), based on NYSE data between
1926 and 1994, found that small stock returns due to infrequent trading show a high
degree of autocorrelation and that they are capable of recording high betas. On the 19
th

century BSE, we confirm their result in the periods after 1867. The only anomaly is the
period between 1882 and 1886, where the small stock average beta lies slightly below the
average beta of large stocks.
Beta Stability 2.3
Here, we adopt the Blume (1971) and Altman et al. (1974) correlation method of
investigating the stability of beta estimates. Blume (1975) shows that the beta coefficient
38

between two successive periods is stationary if
5

( ) ( )
1 i i
t t
E E | |
+
= ,
( ) ( )
1 i i
t t
Var Var | |
+
= ,
( )
1
, 1
i i
t t
corr | |
+
= where
i
t
| are the betas in period
i
t . Betas in period
i
t are used to rank
stocks existing in periods
i
t and
1 i
t
+
in ascending order. In period
i
t equally weighted
portfolios of ,... 3 , 2 , 1 = s stocks are formed as follows: the first portfolio consists of stocks
with s smallest beta estimates. The next portfolio consists of stocks with the next
smallest beta estimates. This process of portfolio formation is repeated until the number
of stocks left is less than s. For each s, the betas of all portfolios in period
1 i
t
+
are also
computed. We compute the correlation and Spearman rank order correlation of the betas
between each two adjacent periods. Table 2.3 reports the weighted average correlation
across all the sub-periods studied. The weighted average correlation takes into account
the number of stocks or portfolios in each adjacent ten-year period.
Table 2.3: Weighted average of correlation and Spearman rank order correlation across successive periods

Note: This table shows the weighted average correlation and Spearmans rank correlation of betas of
individual stocks and portfolios in successive periods across the 15 periods studied. For a stock to be
included in this analysis, it has data for two complete consecutive periods. Betas in a period (estimation
period) are used to rank the betas in that period and the next adjacent period (prediction period) in
ascending order. Portfolios are formed with their constituency as follows: the first portfolio is the first s
stocks for s = (1,2,4,7,10,20). The second portfolio contains the following s stocks and so on until
available stocks is less than s . Assuming equal amounts are invested in each stock, then the portfolio beta
will be the mean of the betas of stocks included in the portfolio. We computed the weighted average
correlations by considering the number of portfolios in each ten-year period.

5
) , ( y x corr is the correlation between x and y .
per Portfolio Correlation
Spearman's
rank Correlation
No. of Stocks
0.92
0.95
0.56
0.64
0.72
0.77
0.91
0.96
0.54
0.62
0.72
0.76
1
2
4
7
10
20
39

The weighted average correlation across the successive periods ranges from 0.54 for
individual stocks to 0.95 for portfolios of 20 stocks. These values indicate that the beta of
individual stocks have, on average, less information about their future beta than the
portfolio beta. Blume (1971) found a similar result on the US market. Due to the limited
number of stocks, we were not able form up to 50-sized portfolios on the BSE. Blume
finds 0.62(0.67) and 0.91(0.93) mean correlation (rank correlation) for 1- and 10-sized
portfolios using 84-month estimation periods respectively. Furthermore, based on 52-
week estimation periods, Levy ( 1971) records 0.44 and 0.82 mean correlation for 1- and
10-sized portfolios on the same exchange. On the UK market between 1955 and 1979,
Dimson and Marsh (1983) used value-weighted market index, monthly returns interval
measurement and sixty-month estimation periods to obtain an average correlation of 0.56
and 0.91 for 1- or 10-sized portfolios respectively. The correlations in Table 2.3 show
that the betas from the 19
th
century BSE are stable compared to the betas in post-1926
US and UK markets.
2.3.1 Blume and Vasicek stability adjustment techniques
Individual stock betas estimated from the MM are noted as unstable in the previous
section (also in Blume (1971), Collins, Ledolter and Rayburn (1987), Faff and John
(1992), Gregory-Allen et al. (1994), Eisenbeiss, Kauermann and Semmler (2007)). There
is a tendency for a high beta estimate to overstate its true value and vice versa. Therefore,
we use the Blume (1971) autoregressive adjustment model to improve the stability of
beta estimates for both individual stocks and portfolios. Table 2.4 presents regression
tendencies implied between adjacent periods, where a and b are the constant term and
slope coefficients, respectively. The values of the coefficients in the periods between
40

1837 and 1867 are striking. It is not consistent with the Blume assertion that all the
coefficients lie between zero and one. The last two Columns show the t-statistics of the
test of a hypothesis of the slope coefficient equal to zero or one. The t-statistics of the
slope coefficients show that the null hypothesis of the slope coefficient equal to zero is
rejected in the adjacent periods after 1872. In addition, the null hypothesis of the slope
coefficient equal to one is rejected in all the adjacent periods (except the first two
periods). The R
2
values also show that the betas in the periods after 1872 have more
explanations for their prior betas than those before 1872. As can be seen in Table 2.4, the
coefficients change over time, but there are extreme coefficients outside the interval
between zero and one in the first two periods.
Table 2.4: Measurement of regression tendency of estimated beta coefficient for individual stocks

The extreme values may be attributed to the number of stocks in a period as we record
less than 50 stocks in our first two five-year periods. A result not reported shows that
increasing the length of the estimation period (such as seven years in Blume (1971))
Regression Tendency
Implied Between Periods a b
R
2
(%) H
0
:b=0 H
0
:b=1
1/1842- 12/1846 and 1/1837- 12/1841 1.40 0.00 0 0.00 -1.31
1/1847- 12/1851 and 1/1842- 12/1846 0.74 -0.07 3 -0.96 -14.11
1/1852- 12/1856 and 1/1847- 12/1851 0.48 0.33 20 2.58 -7.60
1/1857- 12/1861 and 1/1852- 12/1856 0.53 0.54 26 3.22 -5.41
1/1862- 12/1866 and 1/1857- 12/1861 0.27 0.55 18 3.32 -5.43
1/1867- 12/1871 and 1/1862- 12/1866 0.63 0.13 3 1.38 -10.34
1/1872- 12/1876 and 1/1867- 12/1871 0.77 0.43 5 1.94 -4.04
1/1877- 12/1881 and 1/1872- 12/1876 0.65 0.70 20 3.99 -5.03
1/1882- 12/1886 and 1/1877- 12/1881 0.44 0.67 39 7.80 -10.91
1/1887- 12/1891 and 1/1882- 12/1886 0.86 0.69 31 7.66 -10.49
1/1892- 12/1896 and 1/1887- 12/1891 0.58 0.39 27 7.42 -18.75
1/1897- 12/1901 and 1/1892- 12/1896 0.60 0.38 27 7.81 -19.99
1/1902- 12/1906 and 1/1897- 12/1902 0.49 0.70 28 8.94 -11.98
1/1907- 12/1912 and 1/1902- 12/1906 0.78 0.39 15 7.67 -19.21
0.67 0.46 20 18.75 -41.46
null hypothesis of the slope coefficient (b ) equal to zero or one is also reported in the last two columns.
All periods

jt
=a+b
jt- 1
+
jt
Note: In this table beta of stock existing in a period are regressed on the betas of the same stocks in a prior adjacent peri-
od. R
2
(%) is the percentage of the variance of betas in the period t explained by betas in the period t -1. The t-statistic of the
41

improves the
2
R values and the t-statistics but at the expense of losing more stocks, since
fewer stocks have complete returns' data for longer periods.
With the regression tendencies, suppose we want to forecast the beta for any stock or
portfolio in the period 1842-1846. We compute its beta in 1837-1841. The forecast of the
beta is obtained by substituting it for
t-1
in equation (5) with the coefficients in the first
row of Table 2.4.
t
is then computed from the equation and used as the forecast. The
adjustment process is repeated for stocks in the subsequent 13 adjacent periods using
their respective coefficients. We also introduce the Vasicek adjustment model (6) to
adjust betas in successive adjacent periods. We test the predictive performance of the
various adjusted betas by using the root mean square error
6
(RMSE) criterion. The
RMSE tests the performance of the autoregressive methods based on variation and
unbiasedness of their beta forecast. The adjustment method is repeated for all adjacent
periods on equally weighted portfolios of size 2, 4, 7, 10 and 20. In order to compare the
predictive performance of the MM betas and the autoregressive-adjusted betas, we
compute the RMSE of betas estimated with MM in adjacent periods. Table 2.5 displays
the average RMSE of the adjusted betas and MM betas across the 15 periods studied. It is
clear from the table that the average RMSE of the adjusted betas is lower than that of the
market model betas. Table 2.5 also shows that the predictive performance improves as
the number of stocks in a portfolio increases for both adjusted and the MM betas. For
individual stocks, the Bayesian adjustment technique proposed by Vasicek is superior to
Blumes adjustment as reflected in the small average RMSE. Blume (1971) and

6
The root mean square error was calculated by
( )
1
2
i i
jt jt
N
| |

for N j ,..., 1 =

42

Klemkosky and Martin (1975) recorded similar patterns of the predictive performance on
the NYSE market. Their adjusted betas mean square errors were smaller than their MM
betas. For portfolios of size 7 or more, one cannot see much difference between the
Blume adjustment method and the Bayesian approach.

Table 2.5: Predictive performance of Blume and Vasicek (Bayesian) procedures of estimating beta

Note: Average RMSE across the 15 periods studied were used to compare the predictive performance of
the various adjusted betas and the market model (MM) betas in successive periods. The average RMSE
across the successive adjacent periods for the various equally weighted portfolio formations are displayed.
The conclusion is that on the 19
th
century BSE, the predictive accuracy of betas estimated
by the MM can be improved by adjusting betas using either the Blume or Bayesian
adjustment methods and a portfolio with a sizeable number of stocks.
The reliability of the conclusion above can be confirmed by performing an additional test
on the root mean squared error values. The possible method is to test whether the
differences in the values of the RMSEs are statistically significant. Harvey et al. (1997)
presented a modified Diebold and Mariano test statistic that will be used for this purpose.
Therefore, suppose we want to compare the forecasts of Blume (BL) and Vasicek (VA)
models.
{ }
BL
c and
{ }
VA
c are the forecasting errors from the Blume and Vasicek models,
respectively. In our case, we consider the root mean square error function,
( )
BL
f c = root
mean square error of the Blume adjusted betas. The test is based on the loss differential
Bayesian
0.48
0.38
0.31
0.27
0.27
0.23
No.of Stock
per Portfolio
1
2
0.40
1.08
4
7
10
0.57
MM Blume
Average RMSE
0.23
0.19 20
0.45
0.57
0.44
0.35
0.29
0.76
0.25
43

function
( ) ( )
BL VA
j j j
d f f c c = for 1,..., j H = . The null hypothesis of expected equal
predictive performance is H
0
:
( )
0
j
E d = and the alternative hypothesis of the Blume
model predicting worse than the Vasicek model is
a
H :
( )
0.
j
E d > The Modified
Diebold-Mariano (MDM) statistic is:
( )
( )
1
1 2
, 1 1
2
0
1
1 2 1
. ~ 0,1 ,
2
H
i
i
H h H h h
d
MDM t
H
v

=
( + +
=
(
| |
+
|
\ .

where
1
1
H
j
j
d H d
=
=

,
( )
cov ,
i j j i
d d

= , h is the horizon of forecast and
, 1
(0,1)
H
t
v
a
students t distribution with 1 H degrees of freedom and v is the significant level
usually set at 5%. The test compares the Diebold-Mariano test statistics to critical values
from the students t distribution. We reject the null hypothesis of equal predictive
accuracy when the test statistic is greater than the critical value at v level. In order to
apply this test, betas estimated with the Market, Blume and Vasicek models in all periods
are pooled together to form three series of length H. Then we perform the test on the
three series.
Table 2.6 reports the modified Diebold-Mariano test statistics between the various
models under study. Betas estimated from the various models are considered across the
entire period studied. The period studied is divided into two sub-periods based on the
environment in which the BSE operated, a period of strict regulation and a period of
deregulation and expansion. The values in the first row of the table reveal that we can
confidently reject the null hypothesis of one-step ahead equal predictive accuracy of the
44

Blume and MM betas. For instance, in the overall period of our sample the null
hypothesis can be rejected at the 5 percent level. During the strictly regulated period, we
find that Blume-adjusted beta significantly outperforms the market model beta. Between
the Vasicek adjusted betas and the MM betas, we can reject the null hypothesis of equal
predictive accuracy for the deregulated and expansion period.
Table 2.6: Modified Diebold-Mariano test statistics (p-value in parentheses)

The significance level of the rejection becomes weak in the strictly regulated period. The
equality in the predictive performance of the Vasicek and the MM betas is strongly
rejected in the entire period. The values from the bottom row of Table 2.6 show that there
is no significant difference between the Vasicek betas and the Blume betas in terms of
their one-step ahead forecast. Therefore, we cannot reject the null hypothesis of equal
predictive accuracy between the two models.
Beta Bias 2.4
Considering the period of the study and the trading frequency of the market, we might
expect that some stocks may not trade every month for economic reasons or because of
Models
MM
vs.
BL
MM
vs.
VA
BL
vs.
VA
Overall period
1/1837-12/1911
2.18
period
Strict Regulatory
1/1837-12/1871
Deregulation and
expansion period
1/1872-12/1911
(0.01)
1.99
(0.02)
2.87
(0.00)
3.75
(0.00)
4.11
(0.00)
1.44
(0.08)
H
a
= E(d
j
) >0
Note: This table reports the modified Diebold-Mariano test statistics for one step ahead equal forecast
accuracy between the market model, Blume and the Vasicek's adjusted betas.
MM
=market model betas,
BL
= Blume's adjusted betas,
VA
= Vasicek's adjusted betas. The hypothesis test H
0
: E(d
j
) = 0 and
0.56
(0.29)
0.78
(0.22)
0.19
(0.42)
45

regulatory conditions. These stocks may systematically lead or lag the market movement,
producing biased betas when beta is estimated with the MM. In order to expose the
presence of possible lead or lag effects, we test the significance of the coefficient of the
returns on the lagged or lead market index.
Table 2.7: Dimson Aggregate Coefficient (AC) beta Adjustment

Number of AC
Period stocks Beta

-3

-2

-1

0

1

2

3
1/1837-12/1841 21 1.37 0.03 0.05 -0.12 1.42 -0.07 -0.02 0.07
(10) (5) (5) (95) (0) (0) (5)
1/1842-12/1846 33 1.47 -0.11 0.37 0.10 1.06 0.12 -0.10 0.03
(3) (12) (18) (58) (15) (0) (6)
1/1847-12/1851 32 0.69 0.00 0.07 -0.16 0.76 -0.03 0.07 -0.02
(6) (13) (3) (72) (13) (13) (3)
1/1852-12/1856 33 0.80 0.06 0.03 0.00 0.76 0.01 -0.07 0.01
(3) (6) (0) (73) (6) (6) (0)
1/1857-12/1861 54 0.97 0.00 -0.02 0.07 0.99 -0.10 0.00 0.04
(2) (7) (6) (78) (4) (2) (11)
1/1862-12/1866 72 0.56 -0.21 0.19 -0.09 0.90 -0.16 0.05 -0.12
(4) (10) (8) (42) (0) (3) (0)
1/1867-12/1871 76 0.49 -0.09 -0.15 0.05 0.73 0.04 0.00 -0.08
(1) (3) (5) (67) (3) (3) (3)
1/1872-12/1877 82 1.17 0.07 -0.08 0.24 1.02 -0.17 0.12 -0.02
(10) (5) (11) (60) (7) (7) (7)
1/1877-12/1881 112 1.67 0.12 -0.05 0.10 1.19 0.07 0.09 0.15
(6) (7) (13) (55) (6) (11) (11)
1/1882-12/1886 154 0.74 -0.03 -0.03 -0.01 0.97 0.11 -0.10 -0.18
(6) (3) (6) (38) (9) (5) (1)
1/1887-12/1891 160 1.68 0.03 0.12 -0.09 1.58 -0.04 0.00 0.08
(4) (7) (5) (52) (7) (4) (6)
1/1892-12/1896 196 1.42 0.07 -0.07 0.12 1.19 0.08 0.08 -0.04
(6) (6) (6) (42) (8) (5) (4)
1/1897-12/1901 252 0.96 -0.07 -0.05 0.05 1.17 -0.18 0.06 -0.02
(9) (5) (6) (59) (4) (8) (7)
1/1902-12/1906 374 1.11 -0.04 0.09 0.22 1.18 -0.03 -0.16 -0.15
(4) (7) (10) (47) (3) (3) (4)
1/1907-12/1911 424 1.23 -0.03 0.04 0.12 1.13 0.04 -0.05 -0.01
(2) (3) (10) (56) (5) (6) (5)
Overall periods 2075 1.09 -0.01 0.03 0.04 1.07 -0.02 0.00 -0.02
(5) (6) (8) (53) (5) (5) (5)
Mean Lag,Match and the Lead beta estimates
Note: This table reports Dimson's aggregate coefficient adjusted beta in each five year period. The numbers in parenthesis are
the percentage of stocks in a period that reject the null hypothesis of the coefficient been zero at 5% significant level.
46

The Dimson bias adjustment equation with maximum lag or lead of three months is
considered, that is 3 ,..., 3 = i . We use only three months lag and lead because beta bias
has been documented as not prevalent in monthly returns data (see Cohen, Hawawini,
Maier, Schwartz and Whitcomb (1983)). For each stock, the estimates of the parameters
i
| indicate the lagged, matched and lead beta coefficients. We test the hypothesis
0 :
0
=
i
H | against the alternative 0 :
1
=
i
H | for each stock.
Table 2.7 reports the cross-sectional average of the lag and lead betas in each period. The
numbers in parentheses are the percentage of stocks that reject the null hypothesis.
Evidence from this table indicates that beta coefficients
i
| for 0 = i are not significantly
different from zero for the majority of the stocks. This shows that the explanatory power
of the model for 0 = i

is approximately zero for most of the stocks. Unsurprisingly, there
are some stocks with lead and lag coefficients that are statistically significant, but their
numbers does not exceed the coefficients corresponding to the match.

This indicates that
there is no severe timing problem in the 19
th
century data.

As most of the lead and lagged coefficients are significantly equal to zero, we can
interpret this as evidence of the market model (MM) producing statistically reliable beta
estimates in relation to the other models, which incorporate the lagged and lead market
indexes. These results can be compared to the results from the post-World War I markets.
For example, Hawawini and Michel (1979) found a similar pattern of results on the
Belgium stock exchange by using weekly interval returns data between 1963 and 1976.
The result also follows Cohen et al. (1983) hypothesis that there is a strong relationship
between beta estimates and the length of the interval over which returns are measured.
47

They established that beta bias mostly shows up in the short length interval (daily) of
returns, and the bias disappears when the difference of the interval is lengthened
(monthly). Similarly, on the New Zealand market, Bartholdy and Riding (1994) used
monthly data to establish that betas estimated from MM are less biased. On the contrary,
Ibbotson et al. (1997) reports that lagged coefficients should be considered when
estimating beta.
Impact of outlying observations on Beta 2.5
The extreme (maximum/minimum) beta estimates recorded by some stocks in Table 2.1
for the five-year periods studied might be due to the influence of outliers or unexpected
movement by the stock or the market returns. The literature shows that outliers have a
tendency to reduce or increase the magnitude of the beta when it is estimated with the
MM (Chatterjee and Jacques (1994)). In such cases, reducing the impact of outliers in the
estimation of the beta can significantly change the value of beta. We apply the IRLS
(outlier resistant), which minimizes a weighted sum of squares of residuals. The weights
given to each return pair observation depends on the distance between the observation
and the fitted line (Martin and Simin (2003)).
Table 2.8 reports how the presence of outliers affects the beta value. The number of
stocks in each five-year period is grouped into two. As explained in Section 2.2, stocks
with identified outliers less than or equal to 4 are grouped into Category A and those with
identified outliers greater than 4 are grouped in category B. We compute the average beta
of each category. In each period, we compare the cross-sectional average betas of the
MM and IRLS for each category. For example, in the first period out of the 21 stocks,
three fall in Category A with an average market model (MM) beta estimate of 1.29 and
48

IRLS beta of 0.46. In Category A, the difference between the average MM beta and the
average IRLS beta is 0.83. Looking across periods, except for periods 1 to 3, the rest of
the periods have more stocks in Category A than Category B.
Table 2.8: Comparison of the market model betas and the iterative reweighted least square betas

Total Number of Number of detected Number of Average Average
Period stocks outlier observations Stocks MM IRLS
MM
-
IRLS
1/1837-12/1841 21 A 3 1.29 0.46 0.83
B (18) 1.38 0.19 1.19
1/1842-12/1846 33 A 10 0.72 0.49 0.23
B (23) 1.26 0.03 1.24
1/1847-12/1851 32 A 9 0.60 0.52 0.08
B (23) 0.72 0.26 0.46
1/1852-12/1856 33 A 23 0.97 0.81 0.16
B (10) 0.28 0.07 0.21
1/1857-12/1861 54 A 41 1.10 1.00 0.10
B (13) 0.60 0.10 0.51
1/1862-12/1866 72 A 48 0.88 0.60 0.28
B 24 0.72 0.15 0.57
1/1867-12/1871 76 A 42 0.79 0.56 0.23
B (34) 0.67 0.06 0.61
1/1872-12/1877 82 A 54 1.21 0.80 0.42
B (28) 0.75 0.14 0.61
1/1877-12/1881 112 A 78 1.47 1.05 0.42
B (34) 0.91 0.20 0.71
1/1882-12/1886 154 A 110 1.19 1.13 0.06
B (44) 0.62 0.05 0.57
1/1887-12/1891 160 A 120 1.77 1.41 0.36
B (40) 0.80 0.16 0.63
1/1892-12/1896 196 A 146 1.45 1.06 0.39
B (50) 0.62 0.12 0.50
1/1897-12/1901 252 A 198 1.33 1.07 0.26
B (54) 0.40 0.07 0.34
1/1902-12/1906 374 A 290 1.36 1.07 0.29
B (84) 0.83 0.19 0.64
1/1907-12/1911 424 A 346 1.32 1.12 0.20
B (78) 0.47 0.07 0.40
Note: In this Table, stocks with outlier observations less or equal to 4 are grouped in Category A and those with outlying observ-
ations greater than 4 are grouped in category B. In each period, we estimate betas in each category by using the Market model and
iterative reweighted least square (IRLS) model. We also show the number of stocks that falls in each category. We compute the cross
-sectional average of the betas estimated with the market model and the IRLS model.
49

In each period, the difference between the MM betas and the IRLS betas for Category B
is greater than the difference in Category A (last column). This implies that the more the
outlier observations in the return series, the higher the market model overestimates beta.
Figure 2.2: Plot of average market model betas and IRLS betas for stocks with outlier observation less than 4

Notes: This figure depicts the average market model and iterative reweighted least square betas for stocks with at most
4 detected outliers across the 15 periods of study.
The MM beta is always greater than the IRLS beta in Category A (across periods in
Figure 2.2). It confirms the result by Chatterjee and Jacques (1994) that the weighted
least squares estimation reduces the MM betas by a certain percentage. As in Subsection
2.3.1, we apply the modified Diebold-Mariano test to compare one-step ahead predictive
accuracy of the MM and IRLS estimated betas. The modified Diebold Mariano test
statistics proposed by Harvey et al. (1997) are employed to test the null hypothesis of
equal predictive accuracy against the alternative of IRLS betas forecasting better than the
MM betas. A pooled sample of betas within the period studied is considered. The
modified Diebold- Mariano's test statistic between the two models is 1.37 with p-value of
0.09 for one-step forecasts in the period of our study.
50

Table 2.9: Test of equal predictive accuracy between MM and IRLS models

This shows that we cannot reject the null hypothesis of equal predictive accuracy at the 5
percent significance level in the overall period. The null hypothesis can be rejected only
at the 10 percent level. From the deregulation and expansion period, the null hypothesis
of equal predictive accuracy is not rejected. From Table 2.9, we conclude that on the 19
th

century BSE, the IRLS method can help to curb the influence of outliers on estimated
betas, but it does not significantly outperform the standard MM in terms of their ability to
predict one-step ahead in the period of deregulation and expansion.

Models
MM
vs.
IRLS
betas,
IRLS
= iterative resistive least squares method betas. The hypothesis test H
0
: E(d
j
) = 0 and
H
a
= E(d
j
) >0. P-values are in parenthesis.
(0.01) (0.08) (0.09)
Note: This table reports the modified Diebold-Mariano test statistics for one step ahead equal forecast
accuracy between the market model and the iterative resistive least squares model.
MM
=market model
1/1837-12/1871 1/1872-12/1911 1/1837-12/1911
2.29 1.43 1.37
Strict Regulatory Deregulation and
period expansion period Overall period
51

Conclusion 2.6
This chapter evaluates the relative performance of different methods of estimating beta
based on their ability to predict subsequent beta on the 19
th
century BSE. The analysis of
the different beta techniques reveals that beta estimated with the market model is not
stable. Specifically, the study reveals that for individual stocks, the market model beta is
weak in its ability to predict the future beta. The predictability can be improved by
grouping 10 or more stocks to form a portfolio or adjusting betas with the Vasicek and
Blume autoregressive techniques. The study also shows no significant difference
between the Blume and Vasicek adjusted betas in terms of their predictive accuracy.
Applying the Dimson method, correcting nonsynchronous trading effect reveals that
returns of few stocks have a significant relationship with the lead and lag market returns.
There is no significant difference in the predictive accuracy of the betas estimated with
the IRLS method and the market model in the deregulation and expansion period.
In the next chapter, we study the ability of beta to explain returns in the cross-section of
stocks, which is the primary implication of the CAPM.

52

CHAPTER 3
3 THE TEST OF CAPITAL ASSET PRICING MODEL (CAPM)
AND THE SIZE EFFECT IN 19
th
CENTURY BSE

The CAPM posits a positive relationship between the systematic risk (beta) and the
expected return of an asset. Despite the dominance of the CAPM in empirical finance, a
number of researchers have found evidence against the model. The purpose of this
chapter is to investigate if CAPM is valid before World War I, using the Brussels Stock
Exchange (BSE) data. In addition, we will investigate whether a companys market
capitalization (size) is related to its average excess returns (size effect). The use of the
pre-World War I (1868-1914) data will provide very good grounds for an out-of-sample
test of the CAPM and the size effect. This will minimize the data mining critique if the
CAPM is valid, and the size effect exists.
This section investigates the cross-sectional relationship between stock returns and beta.
We use pre-World War I Belgium data. Since the development of the CAPM in the
1960s by Sharpe (1964), Lintner (1965), and Mossin (1966), the literature has questioned
the validity of the model and suggest other characteristics than beta to explain expected
returns. The empirical study that supports the CAPM model in the 1970s is from Fama
and MacBeth (1973). It investigates whether there is a positive linear relationship
between expected returns and beta. In addition, it also examines whether other
parameters such as beta square and idiosyncratic risk can explain expected returns.
However, Banz (1981) finds a size effect in stock returns. The effect implies the
propensity for stocks with low market capitalization to outperform those with high
53

market capitalization. In addition, Lakonishok and Shapiro (1986) and Ritter and Chopra
(1989) do not detect any significant relationship between beta and expected returns. With
the debate on the validity of the CAPM still ongoing, Fama and French (1992) find no
association between betas and average returns, even when beta is the only explanatory
variable in their cross-sectional regressions. Instead, they conclude that size and the
book-to-market value ratio can explain the variation in expected returns when placed
together in a cross-sectional regression. In contrast, Kothari, Shanken and Sloan (1995)
use annual portfolio returns and equally weighted market index to document evidence in
support of the CAPM. The above literature on beta and size focuses on post-World War I
return data and sometimes mainly on data from the US. Another view is that the
characteristic may have been discovered out of luck through data snooping bias (see Lo
and MacKinlay (1990)). In this case, the effect should not be found in other periods.
Dimson and Marsh (1999) test the presence of the size effect by using FTSE all share
monthly returns data from the period 1955 to 1998 and document that the effect has
disappeared after 1979. In addition, Schwert (2003) documents that the size effect
disappears after 1981 in the US market using monthly data for the period 1962 to 2002.
Grossman and Shore (2006) use pre-World War I UK data to present evidence against
the size effect. They find a size effect among extremely small stocks, which account for
about 0.2% of market capitalization, but the size effect disappears when these stocks are
eliminated.
To distinguish between data snooping and the persistence of size effect, we investigate
the effect on another dataset. The anomaly is initially discovered outside the period of
our study. To add to the existing literature on asset pricing, this chapter introduces the
54

19
th
century BSE data to test the validity of CAPM and the presence of the size effect. To
this end, we resort to the sorting and the FM cross-sectional regression methods
discussed in Chapter one.
We find no relationship between beta and expected returns. We also find a size effect on
the 19
th
century BSE, but it disappears when stocks are value weighted to form
portfolios. Detailed investigation reveals that the size effect in our data is confined to
small stocks, which represent on average 0.35% of the total market capitalization.
The remainder of the chapter is organized as follows: In Section 3.2, we show the
expected returns of portfolios sorted on MM betas ( )
MM
| , Dimson betas ( )
dim
| , and
Vasicek betas ( )
V
| . The FM cross-sectional regressions are used to test the relationship
between beta and expected returns (CAPM) in Subsection 3.2.1. In Section 3.3, we
investigate the effect of size and beta on excess returns by using the sorting method. In
Section 3.4, we use FM cross-sectional regression analysis to confirm the above sorting
results. Section 3.4 concludes the chapter.
Expected returns of portfolios sorted on betas 3.2
In the sorting method, we rank stocks based on beta and group them to form portfolios.
As stated before, the question answered by this method is whether high-beta stocks
outperform low-beta stocks. As the aim of this chapter is to test the validity of the
CAPM, the method needed to estimate its input is worth considering. In testing the
CAPM, one needs to form portfolios in order to improve on the precisions of individual
betas. In the previous chapter, Table 2.3 shows that a beta-sorted portfolio should contain
at least seven stocks in order to have a reliably stable portfolio beta estimate. Figure 3.1
55

shows the number of stocks included in our portfolio formation every year. Evidence
from this figure shows that until 1868 decile portfolios will not have the minimum of
seven stocks. The changes in legislation in 1867 ease the establishment of a company,
which is reflected in the number of stocks listed on the BSE. Furthermore, Van
Nieuwerburgh et al. (2006) indicate the importance of the long-term relationship between
the development of the BSE and economic growth in Belgium after legal liberalization.
In addition, as shown in Chapter 2 (Table 2.4), individual betas before 1868 do not
predict well their subsequent five-year beta. Based on these reasons, this chapter and the
subsequent ones will focus on the data between 1868 and 1914.
Figure 3.1: Number of stocks in our selection criteria for the entire period of the pre-world war I SCOB data

For a stock to be included in the portfolio formation, it must have a minimum of 24
months of observations out of the 60 months required to estimate beta before the
56

portfolio formation year. In this chapter, we do not restrict our analysis to stocks with a
complete five-year returns data as in chapter two. This enables us to capture more stocks
in the cross-section. Including stocks with at least 24 months of returns does not change
the descriptive statistics of the prior betas (from here on pre-ranking betas).
For comparison purposes, we study three beta estimates in this chapter. The first is the
MM( )
MM
|

beta, which is the traditional beta. It is the slope coefficient from the
regression equation
( )
,
jt ft MM mt ft jt
R R R R o | c = + + (9)
where
jt
R is the return on a portfolio or stock for period t ,
ft
R is the risk-free rate for
period t , and
mt
R is the market portfolio for period t . We use the value-weighted market
portfolio constructed by Annaert et al. (2004) as a proxy for the market portfolio. The
annualized money market rate, converted to a monthly rate, is used as a proxy for the
risk-free rate. We compute the second beta estimate using the Vasicek model introduced
in Chapter 2. The model is used to compute the posterior Vasicek betas ( )
V
| using the
cross-sectional information on betas estimated by equation (9).
We recall from Chapter 2 (Table 2.7) that some stocks systematically lead or lag behind
the market movement, which may produce biased betas, when we estimate beta by the
MM. Possible explanation for the significant lead (lagged) relationship is because large
(small) firm prices adjust quickly (slowly) to market wide information. Thus, since the
market index used in this analysis is heavily weighted towards large stocks, small stock
returns have the tendency to lead or lag behind in relation to the market wide returns. We
57

adjust for the lag effect by using the Dimson model in Chapter 2 to obtain a third beta
estimate. That is, we run the regression
( ) ( )
,0 , 1 1 1
,
jt ft j j mt ft j mt ft jt
R R R R R R o | | o

= + + + (10)
where
,0 j
| captures the contemporaneous co-variation between the returns of a stock
(portfolio) and the market returns.
, 1 j
|

captures the correlation between stocks current
period returns and the lagged market returns. Dimson one-month lagged beta is estimated
as
,0 , 1 dim j j
| | |

= + , which captures the correlation between the current period returns of
a stock and current and lagged market returns. For our monthly data, we use only one-
month lag because Chapter 2 (Section 2.6) reveals that infrequent trading effect is not a
severe problem. In addition, Dimson (1979) with UK data, documents that the infrequent
trading effect is not severe when monthly returns are used to estimate betas.
Stocks are assigned to decile portfolios using the FM breakpoint method. We estimate
post-ranking portfolio betas for the sample period (1868-1914) by using value-weighted
and equally weighted portfolio returns. Specifically, beginning in January 1868, we
compute betas (pre-ranking) for all stocks using the past 24 to 60 months of returns data.
We sort stocks into decile portfolios based on the pre-ranking betas (univariate sort).
Portfolio 1 contains stocks with the lowest betas, while portfolio 10 contains stocks with
the highest betas. The post-ranking value-weighted and equally weighted return for each
month is calculated for each portfolio. New estimates of pre-ranking betas are calculated
in December each year, and the portfolio formation is repeated. We account for the
possible time-variation in betas by rebalancing stocks in each year. Monthly portfolio
58

formation for each year yields 552 monthly returns for each decile portfolio. This process
is followed for all three beta estimates ( ) , and
MM V dim
| | | .
Table 3.1: Time Series Mean (%), Standard Deviation (%), and Post-ranking Betas of Decile portfolios formed from
pre-ranking betas in Jan. 1868-Dec. 1913

Table 3.1 reports the average excess return (time series), standard deviation, and the post-
ranking betas of the ten portfolios. From Panel A, when both pre-ranking and post-
ranking betas are estimated with the market model, beta does not exhibit any relationship
with average returns. The average returns do not show any pattern as beta progressively
increases from low to high beta portfolios. The result does not change when we consider
Low 1 2 3 4 5 6 7 8 9 High10
0.34 0.13 0.18 0.45 0.45 0.25 0.32 0.23 0.09 0.26
3.91 2.14 1.93 2.88 3.14 3.22 3.68 3.82 4.71 6.67
0.68 0.45 0.53 0.82 1.10 1.19 1.49 1.60 1.94 2.68
0.17 0.18 0.09 0.12 0.40 0.36 0.27 0.16 -0.06 0.05
3.17 1.22 1.62 2.02 2.82 2.90 3.06 3.34 4.07 5.66
0.62 0.33 0.41 0.67 1.01 1.21 1.33 1.45 1.80 2.48
0.18 0.30 0.28 0.26 0.20 0.30 0.35 0.23 0.26 0.32
3.02 3.12 2.33 2.64 3.00 3.07 3.61 4.12 5.13 6.38
0.71 0.64 0.61 0.95 1.26 1.23 1.48 1.71 2.14 2.49
0.06 0.33 0.17 0.32 0.18 0.26 0.30 0.03 0.14 -0.05
1.63 3.35 1.50 1.95 2.79 2.45 3.08 3.56 4.52 5.48
0.39 0.58 0.45 0.75 1.13 0.98 1.32 1.56 1.92 2.32
0.25 0.23 0.26 0.43 0.37 0.25 0.30 0.23 0.20 0.17
3.59 2.32 2.04 3.08 3.09 3.22 3.68 4.70 5.09 5.58
0.92 0.93 0.92 0.97 1.10 1.19 1.45 1.58 1.57 1.55
0.20 0.13 0.12 0.19 0.36 0.36 0.20 0.17 0.06 0.01
2.79 1.24 1.47 2.34 2.71 2.90 3.36 3.39 4.35 4.66
0.80 0.81 0.81 0.86 0.99 1.21 1.36 1.37 1.41 1.39
series portfolio excess returns and the corresponding excess market returns.
Market Model (Value Weighted)
Dimson Betas(Equally Weighted)
Vasicek Betas(Value Weighted)
Mean (%)
Standard Deviation(%)
Beta
Mean (%)
Beta
At the beginning of each year, stocks are sorted based on pre-ranking betas. The pre-ranking betas are estimated with
Beta
Panel A
Mean (%)
Beta
Market Model (Equally Weighted)
Mean (%)
Beta
Beta
Panel B
Mean (%)
Dimson Betas(Value Weighted)
Vasicek Betas(Equally Weighted)
market model (
MM
), Vasicek's adjustment (
V
) model and the Dimson's model with one month lag (
dim
). The Fama-
MacBeth breakpoint technique is used to assign stocks to decile portfolios. Portfolio 1 contains the lowest betas and Portfolio
10 contains the highest betas. Mean (%) is the time series average of the portfolio excess returns for the entire period.We
compute time series Standard Deviation(%) of the post-ranking excess returns. Betas are estimated by using the long time
Mean (%)
Panel C
59

the value-weighted portfolio excess returns. Estimating betas with the Dimson and
Vasicek methods in Panels B and C does not establish the relationship between beta and
expected returns. The most striking of all is that the post-ranking betas almost surely
follow the ordering of the pre-ranking betas (except the first, second, and the sixth decile
portfolios). The univariate beta sorting results confirm Fama and French (1992) findings.
They use Dimson adjusted betas to establish a flat relationship between beta and average
return. We can also compare our result to the evidence of Reinganum (1981) who finds
no relationship between beta and average return in the period 1964-1979.
3.2.1 The Cross-Sectional Regressions
As stated in chapter one, the standard approach to test the validity of the CAPM is
sorting and FM (1973) cross-sectional regression. In this subsection, we use FM cross-
sectional regression to test the robustness of the above sorting result. The FM approach
also provides a straightforward procedure to test whether the reward for bearing beta risk
(risk premium) is equal to the excess market returns (the return of the market less the risk
free rate) as implied by Sharpe, Lintner, and Mossins version of the CAPM. The method
also considers the noisy nature of portfolio or stock returns by running monthly cross-
sectional regressions of beta-sorted portfolio returns on betas. That is,

0 1 jt ft t t jt t
R R | q = + + (11)
where
0t

and
1t
are the regression intercept and slope for month t respectively.
jt
| is
the beta estimated from the full-sample portfolio returns. The slope coefficient from each
regression is treated as the reward per unit of the beta risk in that month (risk premium).
The time series average of the monthly coefficient is the average reward for bearing the
beta risk. The standard deviation of the monthly time series of slopes is used to perform a
60

t-test, whether the average slope is statistically significant from zero, in other words,
whether the beta risk is priced on average. To mitigate a possible error-in-variable
problem, the result in Chapter 2 (Section 2.4) shows that we can rely on the portfolio
betas. Fama and French (1992) also rely on full window portfolio betas to mitigate the
error-in-variable problem. Moreover, it is common to rely on large sample size statistics
to draw inferences. This curbs the argument that the test can be incorrect if the size of the
sample is not large enough for the asymptotic results to provide a good approximation.
We adopt the method by Fama and French (1992) to estimate full window portfolio
betas. The only difference is that we replicate Ibbotson et al. (1997) method and use the
portfolio betas for the cross-sectional regression instead of assigning the portfolio beta to
individual stocks in the portfolio each year. As in the previous section, we sort stocks
based on their estimated pre-ranking betas (Market model betas, Vasicek betas, and
Dimson betas) and form portfolios each year. Portfolio 1 contains the lowest beta stocks
while portfolio 10 contains the highest beta stocks. We form equally weighted and value-
weighted portfolios from the beta-sorted group of stocks each month. We repeat the
process each year to account for time variations in betas. This will produce 552 monthly
returns of decile portfolios (post-ranking returns). The post-ranking betas are estimated
by using the post-ranking long time series' returns of the decile portfolios. We repeat the
process for the various estimates of betas ( ) , ,
MM V dim
| | | .
The post-ranking beta serves as the input for equation (11) to perform the cross-sectional
regressions. Each month, we regress the post-ranking excess returns of the decile
portfolios on their corresponding beta (post-ranking) estimates.
61

Table 3.2: Average time series slopes from the Fama-MacBeth Cross-Sectional Regressions in Jan. 1868-Dec. 1913

Eventually, we obtained 552 cross-sectional regressions for each estimate of beta. After
performing the monthly cross-sectional regressions, the time series mean of the slope
coefficients is tested for statistical significance. The significance of the average slope is
tested by using heteroskedastic and autocorrelation consistent standard errors (Newey
t-test
Intercept
MM

dim
V
H
0
:Slope=(R
m
-R
f
)
Panel A: Equally Weighted Portfolio
0.30% -0.02% 1.41
(2.49) (-0.17)
0.24% 0.02% 0.93
(1.87) (0.14)
0.42% -0.12% 1.05
(1.57) (-0.45)
Panel B: Value Weighted Portfolios
0.24% -0.06% 2.16
(2.92) (-0.48)
0.30% -0.11% 2.63
(3.41) (-0.81)
0.28% -0.10% 0.94
(1.19) (-0.34)
Panel C: Individual Stocks
0.29% -0.01% 1.47
(2.37) (-0.16)
0.26% 0.02% 1.16
(2.10) (0.29)
0.36% -0.06% 1.55
(2.70) (-0.41)
monthly cross-sectional regression of post-ranking portfolio
excess returns on post-ranking beta estimates. It also shows
the hypothesis test of mean slope (risk premium) equal to the
average excess market returns as implied by the Sharpe-Lintner

MM
=Market Model beta,
V
=Vasicek beta and
dim
=Dimson's
CAPM. Newey West adjusted t-statistics are in parentheses.
This table reports average time series slopes and intercepts from
beta with one month lag.
62

and West (1987) correction with default lag of int (
1
4
T ), where T

is 552). We also adopt
the Shankens correction factor discussed in Chapter one for the computation of t-
statistics. This is to eliminate the possible error-in-variable biased induced by the
estimated betas.
Table 3.2 reports the average intercepts, slopes, and their corresponding t-statistics in
parentheses. As shown by the sorting method, Panel A indicates that the market model
post-ranking beta estimated with equally weighted portfolio returns does not provide a
significant relationship with returns. Estimating pre-ranking and post-ranking betas with
the Vasicek and Dimson methods does not establish the beta-return relationship.
Specifically, in Panel A, the mean estimated slope for the market model beta is negative,
and it is only 0.17 standard errors from zero. The negative slope is quite surprising as it
goes against the notion of positive risk premium (CAPM). Fama and French (1992) had a
negative slope for beta when placed together with size in the cross-sectional regression.
The average slope using the Dimson beta is 0.02% with a t-statistic of 0.14. The
estimated mean slope with the Vasicek beta is also not significant.
The values in the last column show the t-statistics from the hypothesis test of average
slope (risk premium) equals the average excess market return as implied by the CAPM.
In Panel A, the hypothesis cannot be rejected at the 5% level, regardless of how beta is
estimated. However, it may be possible that the result is influenced by small stocks, since
equally weighted portfolios give undue weight to small stocks. Therefore, in Panel B, we
use value-weighted portfolios for the estimation of post-ranking betas and in the cross-
sectional regression. The average slope of all the beta estimates in the cross-sectional
regression is not significantly different from zero. Most strikingly, the hypothesis of
63

equality between the average slope and the average excess market return is rejected at the
5% level for the market model and the Dimson betas. In Panel C, we follow the
traditional FM (1973) rolling window approach by using individual pre-ranking betas in
the cross-sectional regression. This is a predictive test since the pre-ranking betas are
estimated over a period prior to the period over which the cross-sectional regression is
performed. The results do not support the CAPM for the three beta estimates.
Although, portfolio betas are used for cross-sectional regression, others believe that
portfolios may conceal important information contained in the individual stock betas. For
example, Ang, Liu and Schwarz (2008) show that the slope coefficient (risk premium) of
cross-sectional regression can be estimated more precisely using individual stocks
instead of portfolios because creating portfolios reduces the cross-sectional variation in
betas. As a result, we apply the Fama and French (1992) approach of estimating full
window portfolio beta and assigning the portfolio beta to the individual constituent
stocks of the portfolio in the cross-sectional regression. This serves as a robustness check
of the results in Table 3.2.
Table 3.3 reports the average cross-sectional regression slopes for both equally weighted
and value-weighted portfolio betas assigned to individual stocks. The market model beta
and the Vasicek beta estimate still maintains the negative non-significant relationship
with average returns. A detailed look at Panel A shows that the Dimson beta is weak in
explaining average returns (average slope of 0.02% but with a t-statistic of only 0.17).
Using value-weighted portfolios (Panel B) to estimate post-ranking betas does not
establish the beta return relationship. This confirms Fama and French (1992) result; they
assert that beta is flat in relationship with average returns for post-1960s USA data.
64

Surprisingly, in all cases the hypothesis that the mean slope is equal to the mean excess
market return is not rejected. The positive average slope of the Dimson beta cross-
sectional regressions (Table 3.3, Panel A) calls for a detailed look into its time series'
behavior with the excess market returns. In addition, the average intercept is marginally
significant, and it is close to the average risk-free rate as postulated by CAPM. To
investigate the evolution of the slope coefficient and the excess market return through
time,
Table 3.3: Average Time Series Slopes from Fama-French Cross-Sectional Regression in Jan. 1868-Dec. 1913

t-test
Intercept
MM

dim
V
H
0
:Slope=(R
m
-R
f
)
Panel A: Fama-French approach (eq)
0.29% -0.02% 1.40
(2.45) (-0.17)
0.24% 0.02% 0.89
(1.81) (0.17)
0.42% -0.12% 1.03
(1.54) (-0.44)
Panel B: Fama-French approach (vw)
0.29% -0.02% 1.39
(2.54) (-0.18)
0.24% 0.03% 0.89
(2.04) (0.19)
0.41% -0.12% 0.97
(1.56) (-0.42)
enthesis.
vw=value weighted. Newey- West adjusted t-statistics are par-
excess market returns.
MM
=Market Model beta,
V
=Vasicek
beta and
dim
= Dimson beta with one lag. eq=equally weighted
In this table, we assign the post-ranking portfolio beta to the
individual stocks in the portfolio. Portfolios are rebalanced
annually. Mean slope and their corresponding t-tstatistic is
reported in parenthesis. We also report the tstatistic for the
test of hypothesis of the mean slope equal to the average
65

Figure 3.2: Sixty months moving average of the cross-sectional slopes and excess market returns using Dimson beta
estimates

Surprisingly, there seems to be a close correlation between the slopes and the excess
market returns for much of the period except between the years (1880, 1885) and (1907,
1913).
Table 3.4: Sub-period look into estimated slopes and excess market returns

Figure 3.2 presents a five-year moving average of the estimated slopes and excess market
returns. The graph shows that the relationship between beta and expected returns varies
with time. In Table 3.4, we report sub-period average slope and intercept from the Fama
1870 1875 1880 1885 1890 1895 1900 1905 1910 1915
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
year
s
l
o
p
e
s
/
e
x
c
e
s
s

m
a
r
k
e
t

r
e
t
u
r
n
s

slope
excess market returns
Intercept Slope
0.15% 0.02%
(0.75) (0.09)
0.32% 0.03%
(3.15) (0.21)
H
0
:Slope=Avg.(R
m
-R
f
)
t-test
sub-periods
0.09
(Avg. R
m
-R
f
=0.04%)
Jan. 1868-Dec. 1893
Jan. 1894-Dec. 1913
(Avg. R
m
-R
f
=0.20%)
2.04
In this table, Dimson's beta estimated from equally weighted
portfolios is used in the cross-sectional regressions for the two
sub-periods. Avg. =Average. Newey West t-statistic in parenthesis.
66

and French (1992) cross-sectional regressions using the Dimson beta. The last column
shows the t-statistics for the test of equality of the average slope and average excess
market return. For the first sub-period, the average excess market return (0.04%) is very
close to the average slope (0.02%). The null hypothesis of the equal average cannot be
rejected. In contrast, the null hypothesis that the average slope is equal to the average
excess market returns is rejected (t-statistic of 2.04) in the second sub-period as the
difference in magnitude confirms (0.03% average slope and 0.20% average excess
market returns). Chan and Lakonishok (1993) document similar results with post-1920
Amex and NYSE data and caution researchers and practitioners not to rush into
discarding beta. The average slope is significantly less than the average excess market
return (a difference of about 0.17%).
Expected Returns, Beta, and the Size Effect 3.3
This section examines the well-known size effect on the 19
th
century BSE. That is, the
propensity for large stocks to have consequent lower returns than small stocks. Early
works of Banz (1981), Reinganum (1981), (1983), Chan, Chen and Hsieh (1985), and
Chan and Chen (1988) first documented the size effect in modern data. Fama and French
(1992) present evidence that size and book-to-market combine to capture the cross-
sectional variation in average stock returns in the period 1963-1990. Subsequently, Fama
and French (1993) build a three-factor model, which uses the excess market returns, size,
and book-to-market factors. The finance literature uses the three-factor model as a
benchmark model to measure long-run abnormal returns and many other factors. This
shows that researchers and practitioners have accepted size as an important characteristic
to explain the cross-sectional behavior of long-run stock returns. On the contrary, a
67

recent paper by Horowitz, Loughran and Savin (2000) presents evidence against the size
effect in the USA market. It conjectures the magnitude of size effect is not robust when
the transaction costs and very small stocks (the removal of stocks with market
capitalization less than $5 million) are taken into account. Schwert (2003) uses US
monthly returns data between the year 1962 to 2002 to document that the size effect
disappears after 1981. With historical data, Grossman and Shore (2006) do not find any
presence of the size effect on UK data between the years 1870 to 1913. This would imply
size is not a systematic risk factor. We present similar evidence on the 19
th
century BSE
covering the same period.
Each year, we sort (univariate sort) stocks based on their size (or market capitalization) at
December of the prior year and then split them into decile portfolios. The market
capitalization is measured as the price of stock multiplied by shares outstanding. Again,
FM breakpoint method is employed to group the stocks into decile portfolios. As in the
previous sections, the smallest size stocks are put in decile one, and the largest size
stocks are put in decile ten. Portfolios are rebalanced each year to capture changes in
their constituent stock market capital over time. Monthly portfolio returns are calculated
as the value-weighted and equally weighted averages of the individual stock returns
within each of the ten portfolios. We compute the relative percentage size of a portfolio
as the time series average of the cross-sectional sum of the market size of the stocks in
the portfolio divided by the sum of the size of stocks in our sample. That is, if
t
n

is the
number of stocks in a portfolio for the month t ,
t
N

is the number of stocks in the cross-
section of our sample for the month t . T is the number of years. The relative percentage
of markets size is computed as
68

1
1
1
1
% Market Size = 100 ,
t
t
n
it T
i
N
t
jt
j
Size
T
Size
=
=
=

Table 3.5: Beta Estimate and Mean Excess Return for the BSE equally weighted size portfolios, Jan. 1868- Dec. 1913

Two beta estimates of the size portfolios are calculated using equation (9) and (10) in the
previous section. Table 3.5 reports the beta estimates and the average excess returns of
the ten-size portfolios during the period 1868 to 1914. It is well known in empirical
finance that small stocks have both a higher beta and average return than large stocks.
However, this is not the case when size portfolios are value weighted in our sample.
Column 3 reveals that equally weighted portfolio 1 has an extreme average excess return
(1.12%) which is almost three times the next largest excess return (0.38% from portfolio
6). The negative relation between size and returns is concentrated in the first decile
Size Portfolio % Market Size EW(%) VW(%)
MM

dim

MM

dim
EW(%) VW(%)
1 0.35 1.12 0.01 1.61 1.89 1.43 1.56 6.16 5.26
2 0.94 0.29 0.14 1.37 1.45 1.26 1.42 4.45 3.86
3 1.60 0.10 0.10 1.16 1.30 1.16 1.28 3.43 3.37
4 2.43 0.14 0.16 1.43 1.54 1.41 1.52 3.65 3.57
5 3.56 -0.06 0.01 1.09 1.16 1.10 1.16 2.77 2.73
6 5.02 0.38 0.52 1.43 1.46 1.61 1.67 3.93 5.16
7 7.04 0.12 0.10 1.16 1.17 1.16 1.17 2.71 2.67
8 10.06 0.20 0.21 1.35 1.36 1.35 1.38 2.97 2.91
9 15.56 0.16 0.19 1.12 1.13 1.12 1.13 2.45 2.43
10 53.46 0.17 0.15 0.88 0.84 0.79 0.74 1.84 1.68
-0.95 0.14
t-statsitics (-3.74) (0.63)
F-statistics with the first decile 4.04
F-statistics without the first decile 0.79
3 and 4. Relative market size is reported in column 2. We also report the Dimson and market model betas for the
In this table, stocks are ranked each year based on their size at the end of the prior year. They are then grouped
deciles for portfolio formation. Portfolio one contains the smallest size stocks, portfolios ten contains the largest
size stocks. Portfolios are rebalancec each year.Average excess returns of the decile portflios are reported in column
EW V W Standard Deviation R
p
-R
f
mean of hedge portfolio (%)
decile portfolios. We also retport the standard deviation of the portfolio return series. The Fstatistic for the test
of hypothesis of equal mean of the porftolio returns is also reported.We test the hypothesis with and without the 1st
decile portfolio. EW=equallyweighted and VW=value weighted.
69

portfolio as the average excess return sharply drops from portfolio 1 to portfolio 2. As the
data source has been well checked for outliers, the extreme excess returns of 1.12%
recorded for the first size decile is not due to data error. Not able to study the events in
1880, 1890 and 1905, the stocks in the lowest size portfolio exhibit very high returns in
these years. In addition, the relative market capitalization (0.35%) of these stocks shows
that they are likely so illiquid that it would have been difficult to profit from buying
them. Looking at the F-statistics, the null hypothesis of equal average returns is rejected
at the 1% significant level when the first decile is included in the test. Excluding the first
decile portfolio fails to reject the hypothesis. In addition, the effect disappears when
stocks are value weighted in the portfolios. The average excess return of the equally
weighted hedge portfolio (mean excess return of -0.95% and t-statistic -3.74) shows that
the size effect exists in our data. Surprisingly, the value increases to 0.14% (t-statistic of
0.63) for the value-weighted hedge portfolio.
Table 3.6: Equally weighted portfolios excess returns without the first-size decile group

R
p
-R
f
(EW%)
1 0.28
2 0.10
3 0.11
4 -0.05
5 0.39
6 0.12
7 0.17
8 0.21
9 0.19
10 0.14
mean of hedge portfolio -0.14
t-statsitic (-0.80)
Size Portfolio
70

Recently, Fama and French (2008) used USA data from 1963 to 2004 to document that
the size effect owes much of its power to micro caps and that it is marginal for small and
big caps. As mentioned earlier, Grossman and Shore (2006) find similar results for the
UK market in the same period of our study.
For robustness, we eliminate the stocks in the first-size decile each year and perform the
size sorting analysis. As shown in Table 3.6, the size effect disappears when we
eliminate the first decile portfolio (portfolio with relative market size of about 0.35%)
before the size portfolio formation every year. This corroborates Horowitz et al. (2000),
who find no size effect in the period from 1963 through to 1981 when they eliminate
firms with less than $5 million in market value on the USA market.
Figure 3.4 plots the MM ( )
MM
| and the Dimson ( )
dim
| betas with the one-month lag
for each equally weighted size portfolio. Clearly, the difference between the
MM
| and
dim
| progressively gets smaller as stock size gets larger.
Figure 3.3: Size Portfolio betas

0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
1 2 3 4 5 6 7 8 9 10
B
e
t
a
Size Portfolio
Market Model Betas
Dimson Betas
71

This shows that small stock betas are underestimated when estimated with the market
model. This might be due to non-synchronous trading as chapter two reveals that some
stocks show lead or lag relationship with the market returns. Ibbotson et al. (1997) find
similar results on the USA market between the years 1926 and 1994. They recommend
the inclusion of lagged information of market returns in the estimation of beta. We also
recommend the use of the Dimson beta with the one-month lag when estimating betas for
small stocks in our sample. This is to curb the possible underestimation of small-stock
beta. There is a clear negative correlation (-0.79 with a p-value of 0.0065) between size
and portfolio beta (Figure 3.4).
3.3.1 Fama-MacBeth Cross-Sectional Regressions to Test the Size Effect
In order to support the above evidence on size effect, we resort to the FM cross-sectional
regression method adopted by Ibbotson et al. (1997). We regress the cross-section of
excess returns for a given month on the beta estimate (full window beta estimate) and
natural logarithm of size by using an extension of equation (11):

( )
0 1 2
1
ln Size
jt ft t t jt t t
jt
R R | q
= + + +
(12)

where
0t
,
1t
and
2t
are the regression intercept and slopes for month t , respectively.
t
| is the full period estimate of beta for portfolio. In our sample, the previous section
reveals that size is cross-sectionally correlated with beta. In addition, Chan and Chen
(1988) argue that as size serves as a proxy for betas, they expect the betas of size
portfolios to be strongly correlated cross-sectionally with size. However, when both
characteristics are included in a regression, the correlation will increase the standard
errors of the estimates, and this will make the outcomes murky for interpretation. Fama
72

and French (1992) show that when portfolios are formed on size alone, there is evidence
of a positive relationship between average return and beta (CAPM). The correlation
between size and beta makes the test on size portfolios unable to disentangle the effect of
size and betas on average returns.
We show that when equally weighted portfolios are built on size alone, there is support
for CAPM. However, allowing the variations in beta that are unrelated to size, it breaks
the effect of size and beta even on equally weighted portfolio excess returns. We achieve
this by conditional double characteristics sorting. Specifically, we first sort stocks based
on size and then sort within each size group on pre-ranking beta. We find a strong
relation between size and average excess return but no relation between beta and average
return for equally weighted portfolios. The size effect disappears when stocks are value-
weighted in portfolios. The size effect does not exist when we eliminate the first-size
decile portfolio in the analysis each year.
As in the sorting method, we form decile size portfolios. This is to confirm the effect of
the correlation between size and beta on the beta-return relationship. To separate the
effect, we sort stocks into three size groups each year. Each size group is then sorted into
five groups based on their pre-ranking
MM
| or
dim
| beta estimates. The equally and
value-weighted return for each portfolio is computed for each month of the following
year. The conditional double-sorting portfolio formation is repeated at the end of each
year. The procedure generates fifteen size-beta portfolios for each beta estimate. For all
portfolio formations, we use the FM breakpoint technique. Post-ranking betas are
estimated with post-ranking returns over the entire period from 1868 through to 1914.
Each month, we regress portfolio excess returns on beta and the natural log of size by
73

using equation (12). The full period post-ranking betas are used in the cross-sectional
regressions. Size is determined at the end of the year before the portfolio formation year.
Table 3.7: Average time series slopes and intercept from the Fama-MacBeth cross-sectional regression: Jan 1868-
Dec. 1913

Intercept
MM

dim
ln (Size) Intercept
MM

dim
ln (Size)
-1.06% 1.05% -0.23% 0.31%
(-3.62) (3.46) (-0.83) (1.12)
-0.88% 0.86% -0.10% 0.20%
(-3.50) (3.43) (-0.45) (0.86)
2.12% -0.13% -0.34% 0.03%
(2.56) (-2.63) (-0.47) (0.79)
-0.42% 0.85% -0.03% -2.00% 0.62% 0.09%
(-0.40) (2.87) (-0.55) (-1.72) (1.74) (1.71)
-2.05% 1.08% 0.06% -2.45% 0.62% 0.12%
(-1.72) (3.50) (0.97) (-1.81) (1.76) (1.86)
Panel B1: Size-
mm
Portfolios Panel B2: Size-
mm
Portfolios
0.26% 0.01% 0.29% -0.12%
(2.18) (0.09) (2.72) (-0.91)
1.63% -0.09% -0.36% 0.03%
(2.08) (-2.05) (-0.56) (0.91)
1.69% -0.03% -0.09% -0.03% -0.12% 0.02%
(2.35) (-0.20) (-2.18) (-0.05) (-0.89) (0.61)
Panel C1: Size-
dim
Portfolios Panel C2: Size-
dim
Portfolios
0.17% 0.08% 0.27% -0.09%
(1.35) (0.56) (2.51) (-0.67)
1.69% -0.10% -0.22% 0.02%
(2.09) (-2.06) (-0.32) (0.63)
1.64% 0.01% -0.09% 0.10% -0.09% 0.01%
(2.17) (0.07) (-2.10) (0.15) (-0.64) (0.31)
EQUALLY WEIGHTED VALUE WEIGHTED
Panel A1: Size Portfolios Panel A2: Size Portfolios
ing betas are used in the cross-sectional regression. t-statistics are in parenthesis.
This will yield 15 size-beta equally and value weighted portfolios. In all portfolio formations we use the FM
break point. Estimate post-ranking betas by using the full period post-ranking excess returns. Post rank-
and value weighted portfolio returns are computed each month in the year. The joint effect of size and
beta is seperated by first forming three size portfolios and splitting each size group into five beta groups.
Each year, we sort stocks into ten portfolios based on their size at the end of the prior year. Equally
74

Table 3.7 reports the time series averages of the slopes and intercept of the regression.
The time series standard deviations of the slopes and the intercepts are used to test
whether the average is significantly different from zero. We use Newey and West (1987)
heteroskedastic autocorrelation corrected standard errors for the computation of the t-
statistics (reported in parentheses). The Shankens adjustment factor is also adopted in
the computation of the t-statistics.
The values in Panel A1 show that the CAPM is valid for equally weighted univariate
size-sorted portfolios. Both
MM
| and
dim
| are positively related to excess return when
placed alone in the cross-sectional regression. Size is negatively related to excess returns.
When size and any of the beta estimates are placed simultaneously as independent
variables, only the beta estimate is significantly related to excess returns. Interestingly,
size is sometimes positively but insignificantly related to excess returns when placed
simultaneously with beta in the regressions. This is contrary to Ibbotson et al. (1997)
result, where size is significant when placed together with market model betas in the
regression. When equally weighted portfolios are formed on size alone, both the market
model and the Dimson beta with the one-month lag can predict returns at the expense of
size.
The implication of the CAPM requires that the average cross-sectional regression
intercept will not be significantly different from zero. However, from Panel A1, the
average intercept is significantly negative when MM and the Dimsons betas are used in
the cross-sectional regressions. The negative intercept becomes significantly positive
when size is placed alone as a regressor in the cross-sectional regression. The average
75

intercept is not significantly different from zero when the MM and Dimsons betas are
placed together with size in the cross-sectional regression.
Panels B1 and C1 show the cross-sectional regression slope and intercept for conditional
double-sorted size-
MM
and size-
dim
portfolios respectively. Both betas

are no more
significantly related to returns, whether placed alone or with size in the regressions.
There is a statistically significant relationship in size to excess returns, whether placed
alone or with any of the beta estimates. This is in support of Fama and French (1992)
evidence that the conditional double-sort portfolio (size-beta sort) allows variations in
beta that are unrelated to size and would break the combined of size and beta on
expected returns. Therefore, size will be related to average returns, but beta will not. It is
worthy to note the significance of the intercepts. These show that beta and size cannot
combine to capture the cross-sectional variation in stock returns.
Most interestingly, when the value-weighted portfolios are used in the analysis (be it
univariate size sorting or conditional double size-beta sorting), beta or size is not
significantly related to the average excess return (See Table 3.7, Panels A2, B2, and C2).
This suggests that the result from the equally weighted portfolio is due to the influence of
small stocks since it assigns equal weights to all stocks in portfolio formations and in the
cross-sectional regressions. This confirms the sorting result in Table 3.5; size effect does
not exist when stocks are value weighted in portfolios. We repeat the above analysis by
adopting Fama and French (1992) method. At the end of each year, the post-ranking
betas estimated with the full period post-ranking returns will be assigned to each stock in
the portfolio. Assigning full period post-ranking betas to stocks does not mean a stocks
76

beta is constant, as stocks can move across portfolios with yearly rebalancing. The
method uses the information available for individual stocks in the cross section.
Table 3.8: Average Time Series Slopes and Intercepts from the Fama-French Cross-Sectional Regressions: Jan. 1868-
Dec. 1913

Intercept
mm

dim
ln (Size) Intercept
mm

dim
ln (Size)
-1.05% 1.04% -0.54% 0.65%
(-3.58) (3.44) (-2.57) (2.79)
-0.87% 0.85% -0.48% 0.58%
(-3.46) (3.42) (-2.58) (2.76)
2.63% -0.16% 2.63% -0.16%
(3.15) (-3.29) (3.15) (-3.29)
1.51% 0.37% -0.12% 2.43% 0.06% -0.15%
(1.34) (1.22) (-2.07) (2.44) (0.26) (-2.88)
1.16% 0.38% -0.10% 2.86% -0.10% -0.17%
(0.86) (1.20) (-1.43) (2.62) (-0.46) (-2.91)
Panel B1: Size-
mm
Portfolios
0.31% -0.03% 0.26% 0.00%
(2.53) (-0.27) (2.27) (0.03)
2.63% -0.16% 2.63% -0.16%
(3.15) (-3.29) (3.15) (-3.29)
2.70% -0.07% -0.16% 2.77% -0.07% -0.16%
(3.54) (-0.59) (-3.39) (3.57) (-0.50) (-3.48)
Panel C1: Size-
dim
Portfolios
0.21% 0.04% 0.17% 0.08%
(1.58) (0.28) (1.44) (0.54)
2.63% -0.16% 2.63% -0.16%
(3.15) (-3.29) (3.15) (-3.29)
2.69% -0.05% -0.16% 2.78% -0.05% -0.16%
(3.52) (-0.35) (-3.47) (3.55) (-0.37) (-3.55)
break point. Estimate post-ranking betas by using the full period post-ranking excess returns. We assign
post-ranking betas to the constituent stocks in the portfolio. Portfolios are rebalanced each year. t-statis-
tics are in parenthesis.
Panel B2: Size-
mm
Portfolios
Panel C2: Size-
dim
Portfolios
77

Table 3.8 reports the average slopes and intercepts of the cross-sectional regressions
using equally weighted and value-weighted portfolios to estimate the post-ranking betas.
The values in parentheses are the Newey West adjusted t-statistics for the test of a
hypothesis of the average slope or intercept significantly different from zero. From Panel
A1, when the full-period equally weighted portfolio returns used to estimate post-ranking
betas are formed on size alone, both
MM
| and
dim
| have a strong relation with returns
when placed alone in the regression. They lose their relationship when placed together
with size in the regression. This indicates that beta, which is correlated with size, serves
as a proxy for size when placed alone in the regression. From Panels B1 and C1,
conditional double-sorting returns based on size and betas break the hold up between size
and beta. It can be seen that beta has no relationship with excess return when it is placed
alone or together with size. The result is similar when value-weighted post-ranking
returns are used to estimate post-ranking betas (See Panels A2, B2 and C2).
For robustness, we repeat Fama and French (1992) cross-sectional analysis, but exclude
stocks in the first-size decile each year. As in the sorting method, Table 3.9 does not
show the significant relationship between betas and expected returns when placed alone
or combined with size for single-sorted size portfolios in panel A1. In Panels B1 and C1,
double sorting stocks to form portfolios will not establish the relationship between betas,
size, and returns. When value-weighted portfolio returns are used in the analysis, size and
beta have no relationship with return as shown in panels A2, B2, and C2. This shows that
any size effect present in our data is driven by a small group of stocks with an average
relative market size of about 0.35%.

78

Table 3.9: Average Time Series Slopes and Intercepts from the Fama-French Cross-Sectional Regressions without
the First-Size Decile: Jan. 1868-Dec.1913

Intercept
mm

dim
ln (Size) Intercept
mm

dim
ln (Size)
-0.21% 0.31% -0.17% 0.28%
(-0.92) (1.25) (-0.85) (1.37)
-0.10% 0.21% -0.11% 0.22%
(-0.50) (0.98) (-0.65) (1.24)
0.20% 0.00% 0.20% 0.00%
(0.28) (-0.07) (0.28) (-0.07)
-1.39% 0.54% 0.06% -0.88% 0.37% 0.04%
(-1.27) (1.83) (1.09) (-0.89) (1.55) (0.77)
-1.65% 0.51% 0.08% -1.22% 0.37% 0.06%
(-1.30) (1.67) (1.22) (-1.10) (1.61) (1.05)
Panel B1: Size-
mm
Portfolios Panel B2: Size-
mm
Portfolios
0.28% -0.09% 0.28% -0.10%
(2.70) (-0.74) (2.82) (-0.72)
0.20% 0.00% 0.20% 0.00%
(0.27) (-0.06) (0.27) (-0.06)
0.34% -0.10% 0.00% 0.52% -0.10% -0.02%
(0.55) (-0.79) (-0.10) (0.86) (-0.76) (-0.41)
Panel C1: Size-
dim
Portfolios Panel C2: Size-
dim
Portfolios
0.28% -0.08% 0.27% -0.08%
(2.57) (-0.62) (2.73) (-0.60)
0.20% 0.00% 0.20% 0.00%
(0.27) (-0.06) (0.27) (-0.06)
0.41% -0.09% -0.01% 0.55% -0.09% -0.02%
(0.68) (-0.67) (-0.22) (0.91) (-0.67) (-0.49)
break point. Estimate post-ranking betas by using the full period post-ranking excess returns. We assign
post-ranking betas to the constituent stocks in the portfolio. Portfolios are rebalanced each year. t-statis-
tics are in parenthesis.
79

Conclusion 3.4
We used sorting and the cross-sectional regression method to investigate whether the
CAPM model is valid in the period before World War I. We find no support for the
CAPM in the 19
th
century BSE. Estimating beta with the market model, Dimson model,
and Vasicek model does not establish the cross-sectional relationship between expected
returns and the beta.
However, when we use equally weighted size portfolios in the cross-sectional
regressions, we find the relationship between excess returns and size or beta. There is a
negatively significant relationship of size to excess returns (size effect), but beta does not
relate to excess returns, when placed at the same time as regressors in the cross-sectional
regression. This is due to a strong correlation existing between size and beta. We find
that conditional double-sorting portfolios by size and then by beta breaks the effect
between size and beta on excess returns. As a result, the average slope of the cross-
sectional regression of returns on betas becomes insignificant when placed alone in the
regression or combined with size. We recommend researchers estimate betas with the
Dimson method with a one-month lag since small-stock betas are underestimated when
estimated with the market model with this data.
Further investigation reveals that the size effect in our data is mainly due to small stocks
with relative market size of about 0.35% of the total market size. Eliminating these small
stocks destroys the relationship between excess returns, beta, and size. Both the sorting
and the cross-sectional regression methods reveal that the size effect disappears when the
value-weighted portfolios are used in the regression.
80

CHAPTER 4
4 DOES THE MOMENTUM EFFECT EXIST IN THE 19
TH

CENTRURY?

Fama and French (2008) document that the momentum or relative strength strategy
produces an average return that is robust and significant across different size categories
of stocks. The strategy buys stocks that have performed well in the past 6 to 12 months
and sells stocks that have performed poorly in the same period. It produces significant
profit in the following 6 to 12 months. However, most studies testing the returns
momentum effect in 6 to 12-month horizons use post-World War II (WWII) data from
the USA, Europe and Emerging Markets (see Jegadeesh and Titman (1993),
Rouwenhorst (1999), Nijman, Swinkels and Verbeek (2004) and etc.). In this chapter, we
use our pre-World War I BSE data to test the presence and source of momentum effect
between the years 1868 and 1914. This will serve as an out-of-sample and robustness test
of the post-WWII momentum profits documented.
Jegadeesh and Titman (1993) document that portfolios or stocks that have performed
well in the past 3 to 12 months often continue to deliver high returns in the subsequent 3
to 12 months. Their result confirms the pervasive, positively significant, lag-predictable
patterns of stock returns documented by Jegadeesh (1990). Subsequent to this, there has
been extensive literature published confirming the robustness of the momentum effect.
For example, Rouwenhorst (1998) documents the momentum effect in 12 European
equity markets. Though researchers and practitioners have subscribed to the view that the
momentum strategy yields a significant profit, the source of the momentum profit poses a
81

strong challenge in the literature. The literature attributes the source of the profit to cross-
sectional variation in expected returns, data mining bias, the state of the market and the
behavioral biases of investors.
To examine the possibility that the profit can be explained by the cross-sectional
variation in expected returns, Jegadeesh and Titman (1993) compute momentum profit
within size and beta-based sub-samples with lower dispersion in expected returns. They
find that the profit is not necessarily smaller in the sub-samples. Based on this result,
they conclude that the momentum profit is not due to the cross-sectional variation in
expected returns. On the contrary, Conrad and Kaul (1998) and Bulkley and Nawosah
(2009) argue that the momentum profit is due to the cross-sectional variation in the
expected returns, but not to autocorrelation in stock returns. Autocorrelation is the
correlation between two returns observations of the same returns series at different times
(time series pattern in stock returns). The cross-sectional variation of expected returns
explaining momentum implies that the returns of the momentum portfolio will be
positive on average in any post-ranking period. However, Jegadeesh and Titman (2002)
attribute the results found by Conrad and Kaul (1998) to a small sample bias in their
empirical test and bootstrap experiments. The bias is due to the likelihood of drawing
with replacement, the same returns observation in the formation and the holding periods.
Jegadeesh and Titman (2002) adopt sampling without replacement to mitigate the bias
selection of a particular observation and realized that the cross-sectional variations in
expected returns contribute very little of the momentum profits.
While some argue that the momentum profit is due to cross-sectional variations in
expected returns, others believe the profits may be due to data mining. To counter the
82

data mining explanation, Jegadeesh and Titman (2001) extend the data from the period
1965-1989 to 1998 and perform an out-of-sample test. Extending the data enables them
to assess whether the result in the period 1965-1989 is merely due to chance. In addition,
it will allow them to assess whether investors have changed their investment strategies
and if the profit of the momentum strategy does not exist anymore, especially for large
size stocks, which are easier and less expensive to trade. However, they find that the
momentum effect still exists in their extended data, and that the profit found in Jegadeesh
and Titman (1993) is therefore not likely to be a data mining bias. They find momentum
in the first 12 months after portfolio formation. However, the cumulative return in
months 13 to 60 after portfolio formation is negative (return reversal), which is consistent
with the behavioral theories.
On the relationship between the state of the market and momentum, Cooper et al. (2004)
apply the behavioral theory of Daniel et al. (1998) to predict the differences in
momentum profit across different market states, like a bull versus a bear market, as
aggregate overconfidence should be greater following market gains. They find that
momentum profits depend on the state of the market. Likewise, Chabot et al. (2009) use
19
th
century data (Victoria Era) from the UK to document the dependence of momentum
profit on the state of the market, when a three year lag in average return is used to define
market states.
It is imperative that executing the momentum strategy involves frequent trading. Despite
the existence of momentum profits on different markets and in different time periods,
Lesmond, Schill and Zhou (2004) argue that when transaction costs are considered,
momentum profit is just an illusion. This is because the strategy tends to pick stocks that
83

have a high trading cost. Frequent trading in these high cost stocks will prevent gainful
strategy execution.
The literature also shows that momentum is strong in industry portfolios than individual
stocks. For example, Moskowitz and Grinblatt (1999) document a strong and persistent
industry momentum effect that cannot be explained by size, value, individual stock
momentum, the cross-sectional dispersion in expected returns and possible
microstructure effects. In contrast, Nijman et al. (2004) documents that the momentum
profits on the European markets are mainly due to individual stock effects. Since the
primary motive of this chapter is to investigate whether momentum profit exists in the
19
th
BSE, future research can be conducted on this data to determine whether the
momentum profit is due to individual stock or industry effect.
In this chapter, we investigate whether a 3 to 12-months momentum effect existed in the
19
th
century and early part of the 20
th
century BSE using the methodology similar to
Jegadeesh and Titman (1993). The only difference is that Jegadeesh and Titman (1993)
used decile portfolios in their analysis, while we use quintile portfolios in order to have a
sufficient number of stocks in our portfolios. We find investors can earn significantly
positive returns over 3 to 12-month holding periods when they adopt the momentum
strategy. The result is not influenced by forming the portfolios just after the formation
period or by skipping one month after the portfolio formation period. Finding a
momentum effect in the 19
th
century casts additional doubt on the data-mining
explanation for momentum.
84

A detailed look into an extensively researched momentum strategy (6 months formation
period and 6 months holding period) on the 19
th
century BSE reveals that the strategy
tends to pick small size stocks. However, further tests also suggest that momentum exists
in beta and large size subsamples. Unlike Jegadeesh and Titman (1993), (2001) who find
negative momentum profit in January (a reversal) each calendar year, we do not record a
January reversal in the 19
th
century BSE. The momentum strategy yields a positive return
in all months throughout the year. Sub-period analysis also shows that momentum profit
is robust in all ten-year sub-periods, except the years between 1878 and 1887.
When we replicate the post-holding period event analysis documented by Jegadeesh and
Titman (2001), we find a strong short-term momentum profit in the first 12 months and
long run reversal, two years after portfolio formation. There is no return reversal in the
first month after the portfolio formation. There is evidence of a sharp rise in profit in the
first year, but the profit declines and reverses to become negative in the second to fifth
year after the portfolio formation, as in the US. We can conclude that in the 19
th
century,
the profit does not remain positive after 12-month holding periods, as claimed by Conrad
and Kaul (1998). We also investigate the momentum profit across market states, as it
may be consistent with many behavioral explanations. We follow Cooper et al. (2004) to
predict the cyclicality of the momentum profit across different market states. We find
that momentum profit depends on the state of the market when three years lagged value
market weighted returns are used to define market states.
The rest of the chapter is organized as follows: we explain and compute the returns of the
various combinations of momentum strategies in Section 4.2. In Subsection 4.2.1, we
will focus on a six-month formation/ six-month holding period strategy. This strategy has
85

been extensively researched in the literature to represent the other strategies. We
compute average returns (equally weighted and value weighted), and the average market
capitalization of the quintile momentum portfolios. We document the returns of the
strategy in size and beta based sub-samples in Subsection 4.2.2. Section 4.3 investigates
seasonality in the momentum profits across calendar months and ten-year sub-periods. In
section 4.4, we investigate the source of the momentum profit by studying the
characteristics of the momentum profit after the 12-month holding period. We study the
momentum profit across different market states in section 4.5. Section 4.6 concludes our
findings.
Momentum Trading Strategies and their Returns 4.2
As in the previous chapters, this chapter uses the 19
th
century BSE data, consisting of
monthly prices, total returns (returns adjusted for stock splits and dividends), and
outstanding shares listed in Brussels between the years 1868 and 1913. As can be seen
from Figure 4.1, the number of stocks in our sample varies from 71 in 1868 to 513 in
1913. The number of stocks increases sharply from 1890 to 1913. The momentum
strategy and its profitability are constructed in this section as follows: At every given
month , we rank stocks in ascending order based on their previous
(formation period) compound returns. Based on these rankings, five equally weighted
and value weighted portfolios are formed. In order to obtain sufficiently diversified
portfolios, we form quintile portfolios instead of the decile portfolios formed by
Jegadeesh and Titman (1993) and Bulkley and Nawosah (2009). The weights are
determined by the market capitalization of each stock in the portfolio at the end of the
portfolio formation.
t months P
86

Figure 4.1: Number of common stocks in our sample for the momentum studies

Figure 4.2: Time line of sample periods
Formation Period Holding Period

t-P t t+Q
The bottom quintile portfolio is called Loser," and the top quintile portfolio is called
Winner. The momentum strategy is to buy the winner portfolio and sell the loser
portfolio in each month . The strategy holds the position for

(holding
period). Either the holding period is immediately after the formation period, or we skip
one month after the formation period. Figure 4.2 shows the various periods we
considered in the strategy. We update our portfolio formation every month. Various
combinations of P and Q are considered, where P and Q equal three, six, nine and twelve
months. This yields 16 strategies each for equally weighted and value-weighted portfolio
formation. We calculate holding period returns from the same month in which the stocks
were ranked in order to form portfolios. Specifically, the strategy that selects stocks
0
100
200
300
400
500
600
1860 1870 1880 1890 1900 1910 1920
N
u
m
b
e
r

o
f

S
t
o
c
k
s
year
t months Q
87

based on the past P-months return and holds it for Q-months is known as a P/Q strategy.
For a comparative purpose, we also calculate returns for skipping one month between the
formation period and the holding period. This will be referred to as P/Q/1 strategy.
Specifically, we construct 6/6 strategy as follows: at the end of each month we sort
stocks in our sample based on their past six month returns (month -6 to month -1) and
group the stocks to form five portfolios based on the ranks. Portfolios are held for six
months (month 0 to month 5) following the ranking months. In order to improve the
power of our test, in all the portfolio formations, we replicate the overlapping portfolio
method adopted by Jegadeesh and Titman (2001). That is, for the 6/6 strategy, the
winning portfolio in month t will contain the top quintile of stocks ranked over the
previous t-5 to t, t-6 to t-1, t-7 to t-2, and it will continue until t-11 to t-6. To illustrate
this on the calendar months, the return in December of the winning portfolio will contain
the top quintile of stocks ranked over the previous June to November, the previous May
to October, and it will continue over the period from January to June. We form equally
weighted and value-weighted quintile portfolios from the returns of stocks that coexist in
the same ranking. We resort to the Fama-MacBeth method for the portfolio break points.
Table 4.1 presents the monthly average returns of the equally weighted and value-
weighted portfolios of the various strategies, over the period from January 1868 to
December 1913 on the BSE. In Panel A, we report both the equally weighted and value-
weighted average returns for strategies with holding periods starting immediately after
portfolio formation. In Panel B, we report equally weighted and value-weighted average
portfolio returns for strategies with holding periods of one month after the portfolio
formation.
88

For each cell in Table 4.1, we report the average returns of the winner, the loser and the
zero cost (winner minus loser) portfolios of the various strategies. The t-statistics based
on Newey-West heteroskedasticity and autocorrelation-adjusted standard errors are
shown in parentheses. For all strategies, the average return of the value weighted zero-
cost portfolio is higher than the equally weighted zero-cost average return. This suggests
that momentum on the 19
th
century BSE is mainly due to large size (market capital)
stocks. In Panel A, the average return of the equally and value-weighted zero-cost (W-L)
portfolios for all strategies are positive and significant, except the equally weighted
average return of the 3/3 strategy, which is marginally significant. In Panel B, the
average returns of the equally and value weighted zero-cost portfolio for all strategies are
positive and significant, except the equally weighted 12/12/1 strategy, which is
marginally significant.
The most profitable zero-cost strategy selects stocks based on their previous 6 month
returns and holds the portfolio for 6 months. It yields 0.71% and 0.75% per month for the
equally weighted zero-cost portfolio, when portfolios are formed immediately and one
month after the formation period respectively. For the value-weighted portfolio
formations, it yields 1.06% each for both 6/6 and 6/6/1 strategies. This is contrary to the
results found by Jegadeesh and Titman (1993), where the most profitable strategy was a
12/3 strategy. The 9/3 strategy, with an equally weighted portfolio formation, yields
almost the same profit as the 6/6 equally weighted strategy (0.70% per month). However,
the profit for the latter strategy declines (0.69 % per month) when there is a time lag
between the portfolio formation period and the holding period.

89

Table 4.1: Profitability of momentum Strategies on BSE (Jan.1868-Dec. 1913)
Formation Period Portfolio
EQ VW EQ VW EQ VW EQ VW EQ VW EQ VW EQ VW EQ VW
3 Winner 0.81% 0.75% 0.87% 0.79% 0.89% 0.82% 0.90% 0.80% 0.87% 0.79% 0.90% 0.81% 0.92% 0.83% 0.88% 0.78%
(4.32) (4.76) (5.44) (6.10) (6.41) (7.31) (7.10) (7.72) (4.69) (5.12) (5.82) (6.60) (6.83) (7.67) (7.09) (7.57)
Loser
0.53% 0.15% 0.44% 0.11% 0.39% 0.06% 0.37% 0.06% 0.48% 0.12% 0.40% 0.08% 0.35% 0.03% 0.36% 0.06%
(2.54) (0.95) (2.31) (0.71) (2.13) (0.42) (2.20) (0.48) (2.29) (0.71) (2.04) (0.51) (1.88) (0.22) (2.12) (0.48)
W-L 0.29% 0.60% 0.44% 0.68% 0.51% 0.76% 0.52% 0.74% 0.39% 0.67% 0.51% 0.73% 0.57% 0.80% 0.52% 0.72%
(1.73) (4.08) (3.40) (5.40) (4.58) (7.56) (5.56) (8.69) (2.37) (4.32) (4.09) (5.93) (5.22) (8.16) (5.78) (8.53)
6 Winner 1.04% 0.98% 1.04% 0.98% 1.00% 0.94% 0.91% 0.85% 1.07% 0.99% 1.04% 0.98% 0.97% 0.91% 0.88% 0.81%
(6.61) (6.86) (7.50) (7.95) (7.72) (8.27) (7.43) (7.79) (6.99) (7.13) (7.58) (7.98) (7.49) (7.92) (7.12) (7.33)
Loser 0.40% -0.05% 0.33% -0.07% 0.33% -0.04% 0.38% 0.04% 0.33% -0.08% 0.30% -0.08% 0.33% -0.02% 0.40% 0.07%
(1.68) (-0.26) (1.52) (-0.41) (1.69) (-0.26) (2.16) (0.25) (1.35) -0.38 (1.35) (-0.44) (1.68) (-0.16) (2.29) (0.50)
W-L 0.64% 1.04% 0.71% 1.06% 0.67% 0.98% 0.53% 0.82% 0.74% 1.06% 0.75% 1.06% 0.64% 0.93% 0.48% 0.74%
(3.75) (5.99) (4.84) (7.44) (5.73) (8.89) (5.37) (8.63) (4.28) (6.44) (5.32) (8.02) (5.78) (9.05) (5.10) (8.29)
9 Winner 1.04% 0.99% 0.99% 0.94% 0.92% 0.87% 0.86% 0.80% 1.03% 0.97% 0.95% 0.90% 0.88% 0.83% 0.83% 0.77%
(7.02) (7.18) (7.25) (7.51) (7.20) (7.46) (7.05) (7.23) (6.89) (7.12) (6.92) (7.22) (6.81) (7.04) (6.72) (6.86)
Loser 0.35% -0.12% 0.37% -0.02% 0.42% 0.07% 0.46% 0.14% 0.33% -0.08% 0.38% 0.03% 0.44% 0.12% 0.48% 0.17%
(1.49) (-0.62) (1.77) -0.135 (2.25) (0.44) (2.76) (0.97) (1.43) -0.414 (1.85) (0.19) (2.42) (0.77) (2.91) (1.23)
W-L 0.70% 1.11% 0.62% 0.96% 0.50% 0.80% 0.40% 0.67% 0.69% 1.05% 0.57% 0.87% 0.44% 0.71% 0.35% 0.60%
(4.36) (7.25) (4.66) (7.72) (4.63) (7.92) (4.31) (7.41) (4.46) (7.12) (4.43) (7.29) (4.19) (7.25) (3.88) (6.72)
12 Winner 0.94% 0.90% 0.89% 0.84% 0.84% 0.79% 0.80% 0.75% 0.91% 0.87% 0.86% 0.81% 0.81% 0.76% 0.78% 0.72%
(6.17) (6.49) (6.35) (6.57) (6.44) (6.64) (6.44) (6.59) (5.95) (6.26) (6.06) (6.23) (6.16) (6.32) (6.17) (6.28)
Loser 0.47% 0.06% 0.52% 0.15% 0.55% 0.20% 0.58% 0.23% 0.49% 0.12% 0.55% 0.19% 0.58% 0.23% 0.60% 0.25%
(2.05) (0.33) (2.56) (0.87) (3.07) (1.31) (3.49) (1.67) (2.17) (0.63) (2.72) (1.15) (3.23) (1.57) (3.63) (1.88)
W-L 0.47% 0.84% 0.37% 0.70% 0.29% 0.60% 0.23% 0.52% 0.42% 0.75% 0.31% 0.62% 0.24% 0.53% 0.18% 0.47%
(2.92) (5.42) (2.81) (5.50) (2.66) (5.65) (2.35) (5.46) (2.71) (5.02) (2.46) (5.04) (2.25) (5.14) (1.89) (4.97)
weighed by their market capital, one month before the holding period. EW= Equally Weighted and VW= Value Weighted.
Panel A: Holding Period Panel B: Holding Period
3 6 9 12 3 6 9 12
This table reports the returns from momentum strategy based on BSE returns data. Each month t, stocks a ranked based on P-months compound returns. Q-months return are
calculated based on the ranking. Equally weighted and value weighted quintile portfolios are formed. The top quintile is called "Winner" portfolio and the bottom quintile
is called "Loser" portfolio. The zero-cost (Winner-Loser) portfolio is the portfolio formed by going long on the winner and short on loser portfolios. In Panel A, portfolios are
formed immediately after the portfolio formation and in Panel B they are formed one month after the portfolio formation. Different values of P and Q are shown
in the second row and first column respectively. The monthly average of the three portfolios are reported with their Newey-West heteroskedastic and autocorrelation adjusted
standard error t-statistics. We adopt the Fama- MacBeth break point method in the portfolio formation. W=Winner and L=Loser. For the value weighted portfolios, stocks are
90

All equally weighted strategies yield profits around 0.50%, regardless of the holding
period and of the one-month lag between the formation period and holding period. It is
out of the ordinary to note that the positive average returns of the equally weighted loser
portfolios are usually significant (except the 6/3, 6/6, 6/9, 9/3, 9/6, 6/3/1, 6/6/1, 6/9/1,
9/3/1 and 9/6/1 strategies, which are marginally significant). This shows that the positive
returns of the zero cost portfolios are mostly due to the positive significant returns of the
winner portfolios and positive significant returns of the loser momentum portfolios. In
effect, 19
th
century investors who would find it difficult to go short on loser portfolios
could have profited from the momentum effect by going long on the winner portfolios. In
contrast, the average return of the value weighted loser portfolio for all the strategies is
not significant, may even be negative sometimes.
The strategy that has been extensively researched in the literature is the 6/6 strategy.
After confirming that momentum strategies of all combinations would yield significantly
positive returns on the 19
th
century BSE, the rest of the chapter will focus on the 6/6
strategy (this strategy will be analyzed, and the result will be used to represent the other
strategies that comprise formation and holding periods ranging from three to twelve
months).
4.2.1 Expected Returns and Average Size of Quintile Portfolios
Table 4.2 shows the portfolios average returns (equally weighted and value weighted)
and average market capitalization (price multiplied by the shares outstanding) of the
quintile portfolios. As before, we construct portfolios by adopting the overlapping
method. For equally weighted portfolios, the average returns increase from the losing
91

portfolio to the winning portfolio, but there is a marginal drop in average returns from the
loser portfolio to the second portfolio.
Table 4.2: Average Returns and Average Size of Quintile Momentum Portfolios

In contrast, the average return of the value-weighted portfolios monotonically increases
from the losing portfolio to the winning portfolio. The values in the last column of the
Table indicate that past losing and winning portfolios pick stocks that have low market
capital on the average. This result is consistent with the findings of Jegadeesh and
Titman (1993) and Chabot et al. (2009).
4.2.2 Momentum Profit within Size and Beta-based Subsamples
In order to investigate whether the profit from the momentum strategy is not confined to
any particular group of stocks, we follow Jegadeesh and Titman (1993) to examine the
profitability of the strategy on subsamples. We group stocks into subsamples based on
the beta and size. Size may serve as a proxy for liquidity and beta for volatility. The
subsample analysis of the momentum profit also provides evidence about the source of
the momentum profit. It tests whether the momentum profit is due to the cross-sectional
Average Market
Capitalization
Portfolio EQ VW (Million Belgium Franc)
Loser 0.33% -0.07% 3.80
2 0.26% 0.11% 9.23
3 0.51% 0.43% 11.76
4 0.76% 0.70% 11.28
Winner 1.04% 0.98% 7.50
W-L 0.71% 1.06% -
Average returns of the values weighted quintile portfolios.
weighted. The sample period is January1868 to December 1913. Market capital is in millions.
W=Winner , L=loser, EQ=Average returns of the equally weighted portfolios and VW is the
This table reports the average returns and average market capital of the quintile portfolios
of 6 /6 months momentum strategy. The market capital is the price times the number of
shares outstanding. Average market capital is the average of the holding period market
capitals of the stocks in each portfolio. Quintile portfolios are equally weighted and value
92

variations in expected returns or predictable patterns in the time series of stock or
portfolio returns. If the profit is due to the cross-sectional variations in expected returns,
the profit will be reduced in the subsamples, as the dispersion of the cross-section of
expected returns is lower in the subsamples than in the full sample. If the momentum
profit is due to predictable patterns in individual stock returns, then the subsamples will
yield positive profits, as will the entire sample.
Table 4.3: Portfolio Returns of the Momentum Strategies with Size and Beta Subsamples

Table 4.3 presents the average return of momentum portfolios, momentum profits and the
difference in the mean of the full sample profit and each sub-sample profit. Portfolios are
Portfolio All Micro Small Big
1
3
Loser 0.33% 1.01% -0.04% 0.09% 0.31% 0.42% 0.53%
(1.52) (3.19) (-0.19) (0.69) (1.72) (2.14) (1.80)
2 0.26% 0.47% 0.35% 0.41% 0.32% 0.41% 0.39%
(2.03) (2.17) (3.01) (3.96) (3.89) (3.37) (1.89)
3 0.51% 0.58% 0.60% 0.54% 0.48% 0.64% 0.58%
(4.58) (3.61) (5.22) (5.48) (8.63) (5.56) (3.16)
4 0.76% 0.92% 0.75% 0.70% 0.74% 0.80% 0.77%
(6.85) (5.92) (6.34) (7.61) (8.05) (7.23) (4.36)
Wi nner 1.04% 1.30% 0.98% 0.82% 0.98% 1.06% 1.06%
(7.50) (7.12) (7.07) (7.02) (7.93) (7.69) (5.69)
W-L 0.71% 0.29% 1.01% 0.73% 0.67% 0.64% 0.53%
(4.84) (1.17) (7.55) (6.26) (5.09) (4.16) (2.84)
Difference - 0.42% -0.30% -0.02% 0.04% 0.07% 0.18%
- (2.86) (-3.35) (-0.18) (0.32) (0.68) (1.77)
months. The overlapping method is employed in the portfolio formation. Loser/winner
portfolio is the equally weighted portfolio of stocks in the lowest/top quintile when
stocks are ranked over the previous six months returns. We adopt Fama Macbeth method
portfolio formation to estimate betas. The last two rows show the difference in mean
This table reports the average monthly returns of portfolios formed based on size and
beta subsamples. Portfolios are formed based on six month returns and held for six
Average Monthly Returns
sample period is January 1968 to December 1913. W=Winner and L=Loser
subsamples. Size is determine at the beginning of each formation period. We estimate
betas for stock returns prior to the portfolio formation period. Dimson's method one
month lag is applied to stocks with at least 24 month returns within the five years prior to
and their Newey West tstatistics between full sample profit and each subsample profit.The
for all portfolio breakpoints. "Micro" contains the smallest stocks, "Small" contains the next
smallest and "Big" contains the largest stocks.
1
,
2
and
3
the lowest to the highest beta
93

formed based on the 6/6 strategy. The momentum strategy is applied on beta and size
subsamples. Size is determined at the end of each formation period. We estimate beta by
using the Dimson method with a one-month lag. The analysis includes stocks with at
least 24-month returns within five years before the formation month.
Clearly, the momentum strategy is profitable in all subsamples except the Micro size
subsamples. This indicates that the profit does not exist for the Micro size subsample
of stocks. The values in the last two rows show the average of the difference in profit
between the full sample momentum profits and the sub-sample momentum profits. The t-
statistics for the test of a hypothesis that the difference is equal to zero is reported in
parentheses. The nonexistence of momentum in the Micro is mainly due to the extreme
average return recorded by its loser portfolio (average return of 1.01% with a t-statistic of
3.19). Not finding momentum in our micro size stocks support Hong, Lim and Stein
(2000) results. They used NYSE and AMEX data between 1980 and 1996 to document
that momentum does not exist in their smallest size decile stocks. On the contrary, the
result does not support Jegadeesh and Titman (1993) and Chabot et al. (2009), who find
momentum profit in all the small size subsamples. The result also deviates from Fama
and French (2008) who find significantly positive momentum profit for small stocks. As
in our sample, it is positive but not significant. Like Chabot et al. (2009) we do not find a
strong relationship between size and momentum profit. However, it is worthy to note the
high profit (the profit of 1.01% with t-statistics of 7.55) for the Small size sub-sample.
Chabot et al. (2009) also find the largest momentum profit for the middle size group.
Consistent with the previous chapter, the beta subsamples show no significant difference
in profit from the full sample. The Big size sub-sample also shows no significant
94

difference in the mean with the full sample. The difference in momentum profits between
the full sample and the subsamples (except the Micro and Small size subsamples)
indicates that the cross-sectional difference in the expected returns of the stocks may not
determine the momentum profits. If anything, the predictable patterns in individual stock
returns may contribute to the momentum profits in the beta sub-samples, as the profit is
not reduced significantly in these groups. It is important to note that the profit of the
Big size sub-sample is mainly due to the buy side of the transaction rather than the sell
side. The expected return of the past winner portfolio is significantly positive, and the
average return of the past loser portfolio is not statistically significant. This implies that
even when investors are not allowed to short stocks, they can still earn a momentum
profit by buying winners.
Seasonality and Sub-period Analysis of the Momentum Profit 4.3
We are motivated to look into the seasonal patterns in momentum profits, as Jegadeesh
and Titman (1993) document positive momentum profits in all calendar months except
January. Jegadeesh and Titman (2001) also confirm the January reversal (negative)
returns in later data. However, the reversal becomes marginal when they exclude stocks
with a price lower than $5 per share and stocks in the smallest decile in their sample.
Therefore, they infer that most of the negative profits recorded in the month of January
by Jegadeesh and Titman (1993) are due to small and low priced stocks, which are likely
to be difficult to trade. Table 4.4 reports the momentum profits in all calendar months on
the Brussels stock exchange between January 1868 and December 1913. This will test
the possible seasonal effect of the momentum profits in the 19
th
century.

95

Table 4.4: Seasonality in momentum profits

Figure 4.3: Average returns of the momentum profit in all calendar months

It also shows the percentage of positive profits in all the calendar months. Again, we
form momentum portfolios based on the 6/6 strategy. The zero-cost portfolio is the
winning portfolio minus the losing portfolio return in each month. We also report the
Months Jan. Feb. Mar. Apr. May June July Aug. Sep. Oct. Nov. Dec. Feb-Dec Fstats
All 0.87% 0.94% 0.97% 0.70% 0.76% 0.72% 0.52% 0.49% 0.50% 0.63% 0.65% 0.76% 0.70% 0.41
(4.02) (4.61) (3.95) (2.72) (2.95) (2.73) (1.39) (1.46) (1.73) (2.80) (2.69) (3.96) (8.84)
76 78 74 74 78 80 78 76 72 67 65 67 74
Micro 0.66% 0.39% 0.47% 0.37% 0.52% 0.35% 0.21% 0.17% 0.38% 0.34% 0.16% 0.34% 0.34% 0.11
(2.08) (1.15) (1.18) (0.85) (1.36) (0.84) (0.46) (0.42) (0.96) (0.96) (0.42) (1.03) (3.11)
65 65 63 59 61 63 67 61 59 54 52 50 59
Small 1.07% 1.29% 1.30% 0.89% 0.96% 0.96% 0.87% 0.88% 0.66% 0.75% 0.93% 0.96% 0.95% 0.65
(3.91) (5.45) (4.89) (3.04) (3.42) (3.51) (3.14) (2.97) (2.25) (4.10) (5.37) (4.31) (11.78)
85 85 83 72 85 80 83 80 78 72 87 76 80
Big 0.92% 0.93% 0.90% 0.69% 0.56% 0.56% 0.56% 0.59% 0.62% 0.89% 1.00% 0.94% 0.75% 0.81
(3.81) (4.32) (4.23) (3.28) (2.68) (3.08) (2.68) (3.49) (3.41) (6.40) (5.44) (5.26) (12.18)
85 83 83 76 76 89 85 80 85 89 87 85 83
For all months in the period 1868-1913, this table reports the average momentum profit, their related t-statistics and the percentage
of momentum profits that are positive in calendar month. The momentum portfolios are formed based on past 6 months returns
and held for 6 months. The equally weighted portfolio is formed from the lowest past return decile and is called Loser. The
equally weighted portfolio returns formed from the past return top rank quintile stocks is called Winner. The zero cost portfolio is
the Winner minus Loser portfolio. The average return, related t-statistic and the percentage of positive profits is also reported for
each size subsample."Micro" contains the smallest stocks, "Small" contains the next smallest stocks and "Big" contains the largest
size stocks. Size is determined at the end of each portfolio formation period. Newey West standard error adjusted tstatistics are in
parenthesis.The last column report the F-statistics for the test of hypothesis of equal average profits in the various calendar months.
This figure report s t he average ret urns of t he moment um profit by calender mont h.
0.00%
0.20%
0.40%
0.60%
0.80%
1.00%
1.20%
Jan. Feb. Mar. Apr. May June July Aug. Sep. Oct. Nov. Dec.
M
o
m
e
n
t
u
m

P
r
o
f
i
t
s
96

average return, corresponding t-statistics and the percentage of positive momentum
profits in each calendar month for size subsamples. Size is determined at the end of each
portfolio formation period. Micro is the lowest tercile group for size, Small is the
middle tercile group for size, and Big is the highest tercile group for size. We compute
the F-statistics in the last column under the null hypothesis that the average returns on the
zero cost portfolios are equal in all calendar months.
The results from the 19
th
century BSE do not support the negative momentum profit
found on the USA market by Jegadeesh and Titman (1993) and Jegadeesh and Titman
(2001). The winner portfolio returns exceed the loser portfolio returns in all months. This
is supported by a significant percentage of positive momentum profits in all calendar
months. Specifically, January records the third highest momentum profit (0.87% average
return, with t-statistic of 4.02 and 76% positive profits) after February and March. The
absence of the January effect on 19
th
century BSE is not surprising, as Chabot et al.
(2009) record similar results in almost the same period (Victorian Era) in the UK. They
argue that investors, as of that time, were not taxed on capital gains and that the tax year
does not end in December. The same reason applies to the BSE. The F-statistics in the
last column indicate that the null hypothesis of equal profit in all calendar months cannot
be rejected for the entire sample and the size sub-samples. This shows that there is no
significant evidence of difference in average momentum profit for the calendar months.
As shown in Figure 4.3, the profit decreases gradually from January to August but picks
up again from September to December. Jegadeesh and Titman (1993) also recorded the
lowest profit in August. Comparing the profits in all calendar months for the various size
subsamples, the Micro size group records the lowest profits in all months. Table 4.5
97

documents the zero-cost portfolio for the 6/6 month strategy in ten-year sub-periods and
the last five years before World War I. This table shows that the momentum strategy
yields positive significant profit in all sub-periods except in the period 1878 to 1887,
although still positive.
Table 4.5: Sub-period Analysis of Momentum Profit

This table reports the average monthly returns of the zero-cost portfolios in ten year subpe-
Sample Month 1868-1877 1878-1887 1888-1897 1898-1908 1908-1913
All 0.55% 0.19% 1.05% 1.13% 0.59%
(1.91) (0.48) (5.60) (4.89) (4.03)
All Jan. 0.96% 0.29% 0.60% 1.96% 0.33%
(1.88) (0.84) (3.60) (4.55) (3.99)
Feb.-Dec. 0.51% 0.18% 1.09% 1.05% 0.62%
(3.96) (0.81) (11.21) (6.54) (6.87)
All 0.17% -0.16% 1.22% 0.30% 0.26%
(0.29) (-0.32) (3.48) (0.68) (1.00)
Micro Jan. 0.50% 0.06% 0.87% 1.68% -0.11%
(0.52) (0.08) (2.18) (2.18) (-0.57)
Feb.-Dec. 0.14% -0.18% 1.25% 0.17% 0.30%
(0.58) (-0.72) (7.02) (0.66) (1.73)
All 0.54% 0.43% 1.17% 1.76% 0.87%
(2.58) (1.29) (6.17) (7.88) (9.08)
Small Jan. 0.67% 0.52% 0.76% 2.42% 0.90%
(1.55) (3.16) (2.46) (3.94) (4.65)
Feb.-Dec. 0.53% 0.42% 1.21% 1.70% 0.86%
(4.82) (2.09) (9.56) (11.43) (9.86)
All 0.75% 0.31% 0.72% 1.23% 0.83%
(2.57) (1.26) (6.73) (6.13) (7.10)
Big Jan. 1.28% 0.36% 0.34% 1.86% 0.69%
(3.25) (1.73) (0.95) (3.27) (2.66)
Feb.-Dec. 0.70% 0.30% 0.76% 1.17% 0.85%
(6.36) (2.09) (11.75) (8.33) (11.45)
riods and the last five years before World War I. The zero-cost portfolios are formed based
on six month past returns and held sixm months. We sorts stocks in ascending order based
on past six months returns and equally weighted portfolios are formed from lowest qui-
ntile stocks group. This group is called the sell portfolio and equally weighted portfolios
formed from the top quintile group of stocks is called Winner portfolio. The zero cost portfo-
lio is the winner minus loser portfolios.The average return of the zero-cost portfolio formed
using size-based subsamples of stocks within sub-periods is also reported."Micro" is the sub-
sample which contains the smallest stocks and "Big" contains the largest size stocks. Newey
west standard error adjusted tstatistics is reported in parenthesis. The sample period is
Jan. 1968- Dec. 1913.
98

When the strategy is applied to the middle and largest size subsamples, it produces
significantly positive profits in all sub-periods except in the periods 1878 and 1887. It is
important to note the significant average returns in January and outside January for
Micro size subsamples in 1888 and 1897 period. This shows that there is a momentum
profit for small stocks in this ten-year sub-period. This may counter the assertion that
momentum profit does not exist in small stocks for the entire period (the assertion is not
robust over time). If we base our research on this ten-year period, we may find
momentum profit for all size subsamples, and small stocks will contribute to the greater
part of the momentum profit across the entire sample period. Except for the periods
1868-1877, 1878-1887 and 1888-1897, there is a significant positive profit for January
and outside January in all the periods for the Small and the Big size subsamples.
Post holding Period Momentum Profits 4.4
The support for short-term reversal, intermediate term continuation and long run post-
holding period reversal of momentum profits has been extensively documented in the
literature (Jegadeesh and Titman (1993), Conrad and Kaul (1998), Jegadeesh and Titman
(2001) and Chabot et al. (2009)). In effect, there seems to be strong evidence of past
losers exceeding past winners in the first month after portfolio formation (short-term
reversal). In addition, past winners continue to exceed the past losers for two to twelve
months after portfolio formation (intermediate term continuation), and past losers will
reverse to exceed past winners over three to five years. The behavioral explanation of
momentum indicates that the time series' variation in individual stock or portfolio returns
contributes to momentum profit. That is, the time series abnormal holding period return
(momentum) is due to investor delay in overreaction to information, which pushes the
99

prices of winners (losers) above (below) their fundamental values. Upon returning to
their senses, investors force the prices to revert to their fundamental values. However,
Conrad and Kaul (1998) argue that the cross-sectional variation in expected returns
generates momentum profit. Empirical evidence presented by Bulkley and Nawosah
(2009) support the assertion. Conrad and Kaul (1998) hypothesize that stock prices
follow a random path with various drifts, and that every stock has a unique drift. They
show that the difference in unconditional drifts across stocks explains the momentum
profits. Since the predictability under this hypothesis is based on the difference in
unconditional drifts across stocks, but not on the variation in the time series of prices for
individual stocks in any particular period, the profits from the momentum strategy will
continue to remain positive in any post formation period. To differentiate between the
behavioral bias hypothesis and the Conrad and Kaul (1998) hypothesis, we examine the
returns of the momentum portfolios, in the periods following the holding periods
considered in the previous sections. That is, if momentum profits are completely due to
behavioral biases, we expect profit to be reduced to zero overtime, and if possible,
reverse its sign. On the other hand, if the profit is due to cross-sectional variation in
expected returns, the profit should continue to increase after the formation period with
time.
Table 4.6 presents the average monthly returns for the first five years after portfolio
formation. From Panel A, across all stocks, the winner portfolio returns drop from 0.91%
in the first year to 0.46% in the fifth year. In contrast, the loser portfolio increases from
0.38% in the first to 0.86% in the fifth year, more than double its value in the first year.
The profit in the second year is approximately the same for the winner and the loser
100

portfolios. From the third year to the fifth year, the zero-cost (winner minus loser)
portfolio profit becomes significantly negative, as the loser portfolio returns exceed the
winner portfolio in these periods. The strongly significant average returns of -0.31%
from the second through the fifth year confirms the return reversal.
Table 4.6: Long Horizon Momentum Profits

Month Months Months Months Months Months Months
Portfolios 1 1-12 13-24 25-36 37-48 49-60 13-60
Panel A: All
0.50% 0.53% -0.02% -0.27% -0.46% -0.40% -0.31%
(2.66) (5.37) (-0.23) (-2.90) (-4.95) (-4.29) (-6.83)
0.99% 0.91% 0.64% 0.53% 0.46% 0.46% 0.59%
(5.65) (7.43) (4.64) (4.34) (3.72) (3.97) (8.76)
0.50% 0.38% 0.66% 0.80% 0.92% 0.86% 0.90%
(2.03) (2.16) (4.07) (4.54) (5.35) (5.11) (10.09)
-0.02% 0.30% -0.23% -0.52% -0.73% -0.59% -0.53%
(-0.06) (1.94) (-1.70) (-3.62) (-5.37) (-4.61) (-7.83)
1.24% 1.13% 0.81% 0.64% 0.54% 0.52% 0.73%
(5.43) (7.15) (4.18) (4.08) (3.50) (3.25) (8.41)
1.26% 0.83% 1.03% 1.15% 1.26% 1.11% 1.27%
(3.41) (3.46) (4.85) (4.95) (5.88) (5.43) (10.79)
called "loser portfolio". Zero-cost portfolio is the winner minus loser portfolios. We
repeat the portfolio formation for size subsamples. "Micro" is the lowest size group,
"Small" is the middle size group and "Big", the largest size group. Size is the market
capitalization of stock and is determined at the end of portfolio formation. Newey
This table reports monthly average momemtum profit for zero-cost, winner and
loser portfolios, one to five years after portfolio formation. Portfolios are formed
based on past six months returns. Stock are sorted based based on the pased six
months returns and equally weighted portfolio is formed from the lower and the
upper quintile group of stocks. The portflio formed from the upper group of stocks
is called the "winner portfolio" and the portfolio formed from the lower group is
West standard error adjusted tstatistics are reported in parenthesis. The sample
period is Jan. 1868 to Dec. 1913.
Zero-cost Portfolio
Winner Portfolio
Loser Portfolio
Winner Portfolio
Loser Portfolio
Panel B: Micro Zero-cost Portfolio
101

Figure 4.4: Cumulative Returns for Five years after portfolio formation

It is important to note that there is no short-term return reversal in the 19
th
century BSE,
as the momentum profit one month after portfolio formation is positively significant
(0.50% with t-statistic of 2.66). Figure 4.4 depicts the cumulative momentum profits for
Table 4.6 Continued
Month Months Months Months Months Months Months
Portfolios 1 1-12 13-24 25-36 37-48 49-60 13-60
0.90% 0.77% 0.27% -0.02% -0.25% -0.18% -0.09%
(4.85) (8.02) (2.76) (-0.23) (-2.46) (-1.59) (-1.81)
0.82% 0.78% 0.58% 0.48% 0.46% 0.46% 0.53%
(4.64) (6.56) (4.81) (4.13) (3.37) (4.16) (7.91)
-0.08% 0.01% 0.31% 0.51% 0.71% 0.63% 0.62%
(-0.36) (0.06) (2.09) (3.11) (4.01) (3.54) (7.42)
0.70% 0.61% 0.17% 0.10% -0.13% -0.21% -0.03%
(4.22) (8.27) (2.31) (1.59) (-2.04) (-2.76) (-0.95)
0.82% 0.76% 0.51% 0.50% 0.43% 0.44% 0.50%
(5.03) (7.29) (4.62) (4.61) (4.30) (4.32) (8.28)
0.12% 0.15% 0.34% 0.40% 0.56% 0.65% 0.54%
(0.83) (1.39) (3.14) (3.74) (4.73) (5.09) (7.74)
Panel D: Large Zero-cost Portfolio
Winner Portfolio
Loser Portfolio
Winner Portfolio
Loser Portfolio
Panel C: Small Zero-cost Portfolio
-0.30%
-0.20%
-0.10%
0.00%
0.10%
0.20%
0.30%
0.40%
0.50%
0.60%
0.70%
0.80%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
C
u
m
u
l
a
t
i
v
e

R
e
t
u
r
n
s
Event Months
Figure 4.4: Cumulative Returns for five years after portfolio
formation
102

five years after portfolio formation. For our 1868 to 1913 periods, the graph shows that
the momentum profit increases monotonically from the first month to the seventh month.
The profit declines from the seventh month to the twelfth month, when it falls below its
first month value. As shown in Table 4.6, the momentum profits become negative around
the second year after portfolio formation, but the negative profit is not strong enough to
offset the positive profit recorded in the first year. This keeps the cumulative profits
above zero until the fourth year, in which the negative profits exceed the positive profits.
Therefore, the cumulative profit falls below zero on the fourth year after portfolio
formation. In effect, there is a strong return reversal from the second to the fifth year
after portfolio formation on the 19
th
century BSE. This result is consistent with the
behavioral explanation found on the USA market by Jegadeesh and Titman (2001).
However, similar analysis on size subsamples (Panels B, C and D) shows that the strong
long-term return reversal recorded for all stocks is mainly due to the lower long run
return reversal in the Micro size group of stocks. For the Micro size stocks in Panel
B, the first month after portfolio formation records an insignificantly negative
momentum profit. The momentum profit one year after portfolio formation is significant
(0.30% with a t-statistic of 1.94). There is a strong return reversal from the second to the
fifth year (-0.53, t-statistic of -7.83). From Panels C and D, the first year records a very
strong momentum profit, but the profit reverses from the second year to the fifth year,
and it is not significant (-0.09 and -0.03 with t-statistics of -1.81 and -0.95 respectively).
The evidence that the return reversal draws its power from small size stocks cautions us
to interpret the behavioral model explanation of momentum with care. In general, the
result is still not consistent with the argument by Conrad and Kaul (1998) either.
103

The Momentum profit and the Market State 4.5
In this section, we investigate whether the condition of the state of the market can
determine the profitability of the momentum strategy. As indicated in Cooper et al.
(2004), the behavioral theory can be extended to predict differences in momentum profit
across the state of the market. The confidence of a group of investors is expected to be
higher following market gains (Daniel et al. (1998) and Gervais and Odean (2001)). The
upwards adjustment of market prices will tend to be attributed unduly to investor skill.
This will result in greater aggregate overconfidence, as aggregate investors hold the long
positions in the equity market. The high overconfidence in following up markets will
trigger strong overreactions, and it will eventually lead to short-run momentum. We
follow Cooper et al. (2004) and Chabot et al. (2009) in defining two market states: for
each month t, the UP (DOWN) market is when the lagged three-year value weighted
market return is positive (negative). We compute average momentum profits across
market states by taking the momentum profit (winner minus loser quintiles) for each
formation month, and taking the average across all formation months that qualify for a
particular market state. Our sample period is still January 1868 to December 1913. For
each month t, we sort stocks into a quintile portfolio based on the past six-month return.
We adopt the portfolio-overlapping method in all portfolio formations. We do not skip
one month between the formation period and the holding period, since the previous
section shows no sign of return reversal in the first month after the formation period. The
Fama and MacBeth (1973) breakpoint method is used to form portfolios. The holding
period profits are computed for three horizons: t to t +5, t to t +11 and t +12 to t +59.

104

Table 4.7: The Momentum profit and the Market State

Table 4.7 reports the mean of the momentum profit for both UP and DOWN markets in
the period January 1868 to December 1913. In order to verify whether the number of
months used to define the market state has any influence on the results, we investigate
the profit of the 6/6 strategy on three (one year, two years, three years, etc.) different
definitions of the market state.
7
For brevity, we report the UP and DOWN market
momentum profit for the three-year definition of market state. In Panel A, for the 6 and
12 months holding periods, the UP market momentum profits are significant, and are all
above 0.60%. Cooper et al. (2004) and Chabot et al. (2009) find similar results. The
DOWN market in panel B shows that momentum profit for 6 and 12 month holding
periods are not significantly different from zero, when the three year lagged value
weighted returns index is used to define the market state. To corroborate the

7
The result is not robust when less than 3 years return is used to define the market states.
Market States Definition
N
Mean Momentum profit
t-statistics
market. The sample period is January 1868 to December 1913.
424
0.88%
(7.42)
424
0.65%
(7.61)
markets. Newey-West t-statistics are in parenthesis. N is the number of observation for each state of the
Months t to t +5 Months t to t +11 Month t +12 to t +59
Market States Definition
Mean Momentum profit
N
Holding Periods
This table reports momentum profits for the UP and DOWN market states. The momentum portfolio is formed
based on past six month returns and held for six months. Each month, stocks are sorted in ascending order
based on past six months returns. Equally weighted quintile portfolios are formed. The First quintile portf-
olio is called the loser portfolio and the top quintile is called the Winner portfolio. The profit of the mo-
mentum portfolio (winner minus loser quintiles) is cumulated across the holding periods : months t to
t +5, t to t +11 and t +12 to t +59 if t -1 is the end month of the formation period. Market is classified as
UP (DOWN) if the value weighted market index over the period t -36 to t -1 is positive (negative).We
report the mean monthly profits, Panel A and B reports the momentum profit following UP and DOWN
Panel A: UP market states
3-years
4.28 4.59 3.65
128
0.15%
(0.37) (0.62) (-4.85)
Panel B: Down market states
Panel C: Test of Equality of the Mean Momentum profit (UP=DOWN)
128
3-years
0.15%
376
-0.27%
(-6.07)
128
-0.43%
105

unconditional results from Lee and Swaminathan (2000), Jegadeesh and Titman (2001)
and Cooper et al. (2004), we find that UP market momentum profit significantly reverses
over the long run. The average momentum profits are significantly below -0.20% over
the holding period of 13 to 60 months. We also find significant long-run reversal for
DOWN markets. Cooper et al. (2004) find similar results, and they assert that long run
reversals are not solely due to corrections of prior momentum.
In Panel C, we test the hypothesis of equals in the mean of the momentum profit between
the UP and DOWN markets. The hypothesis is rejected at 5% for the five-year holding
period horizons considered after portfolio formation. Chabot et al. (2009) did not find a
significant difference in the mean between the UP and DOWN momentum profits.

106

Conclusion 4.6
We investigate the trading strategy that buys past return winners and sells past return
losers (momentum trading strategy) over the period of January 1868 to December 1913.
There is convincing evidence that the momentum strategy is profitable. For instance, the
6/6 strategy that we study in detail here yields an average profit of about 8.52% and
12.7% per annum, for equally and value-weighted portfolios, respectively. Finding
momentum profits in the 19
th
century provide some evidence that the profits found for
the post-World War II US market are not mainly due to data snooping biases. Detailed
analysis of the 6/6 strategy on beta and size subsamples shows that momentum profit is
not confined to particular beta subsamples. We find no momentum profit for small size
stocks. Except for the small size sample, the momentum profits for beta and large size
samples are not significantly different from the full sample profit.
Investigating the momentum profit for each calendar month shows that the negative
January momentum profit found on the USA market by Jegadeesh and Titman (1993),
(2001) is not in the 19
th
century BSE. January records the third highest momentum profit
relative to the other months of the year. The momentum profit is not robust across our
sample period, as a ten-year sub-period analysis indicates that the profit is not strong in
the first twenty years of our sample. The profit is marginally significant or not significant
in the first and the second ten years respectively. Unsurprisingly, the marginal profit
recorded in the first ten years is mainly due to the middle and the large size groups of
stocks, since the average momentum profit for the small size group is insignificant.
To investigate the source of momentum profit, we resort to the Jegadeesh and Titman
(2001) approach to examine the returns of the momentum portfolios in the post-holding
107

period. We study post-holding period returns to differentiate between the assertion by
Conrad and Kaul (1998) and behavioral explanations presented by Barberis, Shleifer and
Vishny (1998), Daniel et al. (1998) and Hong and Stein (1999). Conrad and Kaul (1998)
assert that the momentum profit is due to the cross-sectional differences in expected
returns, not to the time series' pattern in asset returns. Therefore, the momentum strategy
will continue to yield positive returns in all periods after the holding period. The
behavioral explanations suggest that, if momentum profit is completely due to under
reaction to information, the returns of the momentum portfolio will decay to zero.
Conversely, if the profit is due to lagged overreaction, the momentum portfolio will
possibly reverse in sign before decaying down to zero overtime. Our evidence supports
the Jegadeesh and Titman (2001) results, and suggests that the profit of the momentum
portfolio in the 13 to 60 months after the portfolio formation is significantly negative (-
0.31 with t-statistics of -6.83). This evidence supports the behavioral explanations and
clearly rejects the Conrad and Kaul (1998) assertion. Further study of the size
subsamples revealed that the negative profit recorded in the months 13 to 60 after
portfolio formation is mainly contributed by the small groups. This suggests that the
support for the behavioral explanations of the momentum profit should be interpreted
with caution, although the result is still not consistent with the Conrad and Kaul (1998)
assertion.
We also test if momentum profit and long-run reversal in the cross-section of stock
returns depend on the state of the market. Cooper et al. (2004) document that the Daniel
et al. (1998) behavioral theory can be extended to predict differences in momentum profit
across states of the market, like UP and DOWN markets, as aggregate overconfidence
108

should be greater following market gains. For each month between 1868 and 1913, we
define the state of the market as UP (DOWN), if the lagged three-year returns of the
value weighted market index are positive (negative). The 6 month formation and 6 to 12
month holding period strategies are solely profitable in the periods of market gains, when
three year lagged value weighted returns are used to define the market state. Specifically,
from 1868 to 1913, the average monthly momentum profit for UP market states is
significantly positive (0.88% and 7.42 t-statistic, 0.65% and 7.62 t-statistics for both 6
and 12 months holding periods respectively). In contrast, the DOWN markets in the same
period record insignificant positive momentum profits of 0.15% (t-statistic of 0.37 and
0.62 for both 6 months and 12 months holding periods). We also find that the momentum
profit in the UP market states is reversed in the long term.

109

CHAPTER 5
5 THE COMBINED EFFECT OF DIVIDEND YIELD, SIZE,
TOTAL RISK AND MOMENTUM (1868-1913 EVIDENCE)
8

In the previous chapters, stock returns do not show a positive relationship with beta
(CAPM) in the 19
th
century BSE data. However, the winners (losers) over the past 3 to
12 months predict the subsequent 3 to 12 month winners (losers) for the 19
th
century BSE
data (momentum). Size seems to have a relationship with average returns in our data.
However, the relationship is completely driven by extremely small stocks, accounting for
about 0.35% of the total market capitalization. In this chapter, we use the 19
th
century
BSE data to study the predictive pattern of stock returns based on size, momentum, total
risk and dividend yield. The next section explains the motivation for repeating size and
momentum and choosing total risk and dividend yield. By studying the predictive power
of these characteristics on stock returns in a new database, we provide out-of-sample
evidence that may address the data-mining critique. The results may also provide
evidence on the combined effect of different characteristics on the cross-section of stock
returns.
The Capital Asset Pricing Model (CAPM) of Sharpe (1964), Lintner (1965) and Mossin
(1966) provides a particularly appealing way to look at the cross-section of expected
returns. The model implies that expected returns are a linear function of the stocks beta
coefficient, i.e. the regression slope of the stocks return on the market return. It

8
Different versions of this chapter have been presented at conferences: the 14
th
Conference for Swiss
Society for Financial Market Research (SGF), 8
th
April, 2011, Zurich Switzerland and (Blind Reviewed).
Eastern Finance Association 2011 Annual Meetings, 15
th
April, 2011, Savannah- Georgia, USA (Blind
Reviewed)
110

measures the securitys systematic risk, which is the only part of its total risk that should
be rewarded in the market. Although initially the model was corroborated empirically
(e.g. Fama and MacBeth (1973)), its prediction is not borne out in more recent empirical
research, as summarized in Fama and French (1992). Not only is there no longer a
positive relation between beta and average stock returns, the returns are also found to be
related to other firm or stock characteristics, such as size (market capitalization) and its
book-to-market ratio. Later research adds even more characteristics that are associated
with average returns. For instance, Fama and French (2008) dissect the return patterns
based on momentum, asset growth, profitability, net stock issues and accruals, as well as
size and book-to-market.
These patterns beg explanation. One possibility is that the characteristics are proxies for
exposure to common risk factors, which then leads to the question of which factors
should be studied and how to measure them. Another explanation is that they reflect
irrational investor behavior that a rational investor could exploit. This then raises the
issue of to what extent these patterns are tradable, as the patterns are often stronger for
smaller and less liquid stocks and exploitation involves high portfolio turnover. For
instance, Hanna and Ready (2005) demonstrate the vulnerability of some patterns to
turnover-induced transaction costs. Finally, the patterns could be illusory or simply found
thanks to collective data dredging (see e.g. Schwert (2003)), in which case the patterns
should not be discernable in new datasets.
In this chapter, we shed some light on this issue by studying cross-sectional patterns in
Belgian stock returns for the period of 1868-1914. This unique and high-quality data set
provides a genuine out-of-sample testing environment. In chapter 1, we tested the
111

predictive power of the beta coefficient for average stock returns. We find that, similar to
the post-1969 U.S. results of Fama and French (1992), the relation in pre-World War I
Belgian data is flat. We pay particular attention to the computation of beta, as 19
th

century stock markets were less liquid than their modern counterparts, but to no avail.
Also, within the largest group of stocks, representing on average about 96% of stock
market capitalization, beta and size are of no importance in explaining average returns.
We now turn our attention to the test of the presence of some repeated and other
characteristic effects, namely size, momentum, total risk and dividend yield. We repeat
size in order to investigate whether the relationship between stock returns and the other
firm characteristics will not be confined to the extremely small stocks. In addition, Fama
and French (1992) show that book-to-market and size are the most important stock
characteristics related in the cross-section to average returns. As we do not have
accounting information, we cannot include the book-to-market ratio, but we do include
the size (market capitalization) of the firm in our analysis. Given the evidence that the
size effect seems to have attenuated since the publication of Banz (1981) (see Schwert
(2003)), additional robustness analysis, using a new and independent dataset may shed
light on the explanation of the size effect. Horowitz et al. (2000) list three potential
explanations: (a) data mining; (b) the increased popularity of passive investments, which
would have driven up prices of large companies; (c) the awareness of investors after
publication of the research results has eliminated the profit opportunities. As index funds
did not exist in the 19
th
century, finding a size effect would favour the awareness
explanation, whereas not finding a size effect is more consistent with the data mining
argument.
112

Since Jegadeesh and Titman (1993) published their article on momentum, or short to
medium term past returns, this type of stock prediction has received a prominent place in
empirical asset pricing. Although it is still not clear why momentum is positively related
to future average returns, it is complementary to the size and book-to-market effects
(Fama and French (1996)). Also, Jegadeesh and Titman (2001) have updated their
previous results and still consistently observe the momentum effect in the years 1965 to
1998, making it less likely that it is due to data mining. Fama and French (2008) use a
different measure of momentum in their dissection of the anomaly test and find
momentum to be positively related to the average return. We follow Fama and French
(2008) to repeat momentum in the analysis. Including this characteristic in our tests
yields additional out-of-sample evidence.
Furthermore, there has recently been a debate on the usefulness of idiosyncratic or total
volatility to predict average returns in the cross-section. Theoretically, Merton (1987)
argues that when investors do not or cannot hold the entire market portfolio due to
various exogenous reasons, idiosyncratic risk and total risk should be priced. On this
note, Malkiel and Xu (2006) find that portfolios of stocks with high idiosyncratic
volatility have higher returns than portfolios with low volatility stocks. In contrast, Ang
et al. (2006) find a negative relation between average returns and idiosyncratic volatility.
In addition, Blitz and Van Vliet (2007) based on Sharpe Ratio and alpha from the Fama-
French three factor model document a negative relationship between excess returns and
total volatility on global markets. However, Bali and Cakici (2008) show that the
relationship between total risk and excess return is induced by methodology and the
predominance of small illiquid stocks in the sample. Screening stocks using liquidity and
113

price filters destroys the relation. In addition, in the 19
th
century, stocks were expensive
relative to the average daily wage (see Scholliers (1997)). The above literature measures
idiosyncratic risk as the standard deviation of the errors in the Fama and French (1993)
three-factor model. However, we cannot compute the value factor in the Fama-French
model as the accounting data has not been digitalized in our data. In addition,
idiosyncratic risk estimated from the standard deviation of the error term in Fama-
Frenchs three-factor model may introduce error-in-variable complications in the cross-
sectional regression.
On this note, we considered total risk to minimize the error inherent in the estimation of
risk. Investors might not be able hold the market portfolio due to the high prices of stocks
relative to the daily wage. Therefore, total risk could also be priced to compensate
rational investors for their inability to hold the market portfolio. Total risk, which is the
main arbitrage risk, may also prevent arbitrageurs from exploiting mispricing
opportunities on the market. Therefore, total risk will have a positive relationship with
mispricing (arbitrage limit theory). Given this debate and the household income situation
in the 19
th
century, we decide to include total risk in our list of characteristics. We follow
Blitz and Van Vliet (2007) to measure total risk as the standard deviation of the past two
to five years excess returns.
Lastly, we would have investigated the relationship between book-to-market value and
the cross-section of stock returns in the 19
th
century BSE, following Fama and French
(1992). However, as we do not have data on the book-to-market value of equity, we
follow Grossman and Shore (2006) and use dividend yield as the best available proxy for
value. In addition, the investigation on the interaction between dividend yield (as a
114

measure of value) and momentum has recently been documented in the literature. For
example, Asness (1997) documents a negative correlation between momentum and
dividend yield. This implies investment in high value stocks, and to some extent entails
investment in poorly performing stocks and vice versa. Similarly, Gwilym, Clare, Seaton
and Thomas (2009) use independent double-sort quintile portfolios to document that the
momentum strategy yields a significant return among the lower value quintile, and the
value strategy is less effective within the highest momentum quintile. They advise value
investors to stay away from high momentum firms until they exhibit at least some
relative strength compared to the general market. On historical data, Grossman and Shore
(2006) also record that a negative relationship exists between dividend yield and
momentum. In view of this, we also investigate the relationship between momentum and
dividend yield in the 19
th
century BSE. Moreover, as it has been shown in the literature
that small firms find it necessary or desirable to pay dividends in the earlier periods, we
also investigate the relationship between size and dividend yield in the 19
th
century BSE.
In addition, we investigate the effect of the relationship between dividend yield and total
risk on the average returns.
The rest of the chapter is organized as follows: The next section describes the
measurement of the characteristics. In section 5.3, descriptive statistics of the
characteristic sort portfolios are studied. Section 5.4 describes the average returns of the
single-sort and independent double-sort portfolios. We use the FM cross-sectional
regression method to confirm the sorting result in section 5.5. We conclude the chapter
with section 5.6.
115

Measures of Characteristics 5.2
We follow two paths to establish the importance of the characteristics: portfolio sorting
(mainly single-sort and independent double-sort) and FM cross-sectional regressions. To
study the pervasiveness of any pattern, we divide our sample into three groups based on a
characteristic and investigate the effect of dividend yield separately in each group.
As in chapter 3, our sample starts in 1868 and Figure 3.1 illustrates our motivation. As
part of our analysis relies on portfolio sorts (sometimes with double sorting), a minimum
of stocks in the cross-section is needed. Thanks to a change in legislation in 1867, it
became much easier to set up a company, and the number of listed firms increased
accordingly. Portfolios are constructed in January of each year based on information
available by the end of December of the previous year. More specifically, to be included
in the analysis, the stock has to comply with the following requirements:
1. In order to obtain some accuracy in the estimation of total risk, a minimum of 24
out of the 60 months return observations are required.
2. Six months return data in the previous year is necessary to compute our
momentum measure (discussed below).
We use excess return, which is computed as the difference between the realised return
and the risk-free rate used in chapter three.
In order to study the relationship between dividend yield, size, total risk and momentum,
we form yearly portfolios. Unlike Asness (1997), we pay more attention to zero-dividend
stocks, as they account for, on-average, more than a quarter of the stocks in the cross-
section each year. Figure 5.1 shows the percentage of stocks in the cross-section that did
not pay dividends each year.
116

Figure 5.5.1: Percentage of Zero-dividend paying stocks and their Relative Market Capital: Jan. 1868- Dec.1913

It is clear from the figure that about 45% of the stocks in the cross-section did not pay
dividends in 1880 and or in 1903. On average, about 28% of stocks did not pay
dividends across the period of our study. This tends to be smaller stocks compared with
those that pay dividends. The relative market capitalization of these firms does not
exceed 20% across the period of our study. The result is not surprising, as Gwilym et al.
(2009) and Grossman and Shore (2006) record that small stocks do not pay dividends on
the contemporary and historical UK markets respectively. We treat the zero-dividend
stocks as a group, rather than incorporating them into the lowest dividend group of
stocks. We do this because both Keim (1985) and Morgan and Thomas (1998) had a
comparable return between the zero-dividend firms and the highest dividend yield firms.
In effect, a U-shaped relationship between dividend yield and return. Each year, in
order for a stock to be part of this analysis in the following year, it must have at least 6
months of return data to facilitate the computation of the momentum characteristic.
As before, each year, we measure the size (market capitalization) as price multiplied by
shares outstanding in December of the year before the portfolio formation. Momentum is
0
5
10
15
20
25
30
35
40
45
50
1860 1870 1880 1890 1900 1910 1920
P
e
r
c
e
n
t
a
g
e

year
Relative Market Capitalization
% Zero-Dividend Paying stocks
117

measured as the compound return from June to November of the previous year. We are
motivated to use the six-month compounded return, as the 6/6 strategy emerges as the
most profitable strategy in chapter 4. Fama and French (2008) rely on the most profitable
strategy in Jegadeesh and Titman (1993) to compute the 12-month compound return as a
momentum measure. Total risk is the standard deviation of the past 24 to 60 months
excess returns. Finally, each year, we measure the dividend yield as the sum of all
dividends paid in the past 12 months, divided by the price in December of the year. As
Annaert et al. (2004) observe seasonality in dividend yield, summing dividend over a
full year removes any seasonal patterns in dividend payments, and the current price level
is used to incorporate the most recent information in the stock prices.
Each year, we separately sort stocks into three groups based on size, momentum or total
risk. We use the FM break-point method in all stock groupings. The lowest size group is
called Micro cap stocks. The second group is called Small cap stocks and the largest
size group is called Big cap stocks. For total risk and momentum, we place the lowest
total risk (momentum) in group 1 and the highest in group 3. In the same year, we
separately sort stocks based on dividend yield. Zero-dividend paying stocks are placed in
one group. The rest is then sorted into three groups based on their dividend yield. The
stocks that pay the lowest dividends are in group 1 and the highest dividend paying
stocks are in group 3. We follow Lakonishok, Shleifer and Vishny (1994) and Fama and
French (1996) to form 12 (for example, Micro/D/P= 0, Micro/1, Micro/2, Micro/3,
Small/D/P=0, Small/1, Small/2, Small/3, Big/D/P=0, Big/1, Big/2, Big/3) sets of
portfolios from the intersections of dividend yield sorts and other characteristic sorts.
These are the independent double-sort portfolios. We repeat the portfolio formations for
118

dividend sorts and the size, momentum or idiosyncratic risk sort. Sorting on two
characteristics will test the marginal effect of each characteristic on average excess
returns.
Descriptive Summary Statistics of the Characteristics 5.3
In this section, we report the summary statistics of the twelve sets of portfolios formed on
size-dividend yield, momentum-dividend yield and idiosyncratic risk-dividend yield. We
also report the univariate sort on the various characteristics.
5.3.1 Size-Dividend yield double sorts
Table 5.1 reports the summary statistics of the characteristics for the 12 portfolios formed
from the intersections of the size sort and dividend yield sort stocks. We also report
separate univariate sort portfolios for size and dividend yield. Between 1868 and 1914,
our sample contains 84 to 518 different firms, which results in an average of 240 stocks
traded every year (see Panel A and Figure 3.1). On average, there are seven or more
stocks in every independent double-sorted portfolio. The number of Micro stocks that
do not pay dividends in the last 12 months exceeds the number of total stocks that pay
dividends. It is not surprising to see fewer numbers of Big stocks that do not pay
dividends (seven stocks on the average).
In panel B, the average relative market capitalization of each portfolio is reported. The
first column indicates that the Big portfolio accounts for the largest percentage of
money invested in the stock market, representing, on average, 81% of total stock market
capitalization. On the contrary, the Micro portfolio accounts for less than 4% of stock
market capitalization. The remaining stocks represent only 15% of market capitalization.
For the double-sorted portfolios, it is surprising to see more money invested in Micro
119

Table 5.1: Summary statistics for Size-Dividend double-sorts
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 72 56 56 56 8.96 43.96 29.67 17.42
Micro 80 46 8 10 16 3.67 1.52 0.47 0.65 1.02
Small 79 19 19 20 22 15.22 3.10 3.86 3.97 4.28
Big 80 7 30 26 18 81.11 7.00 29.50 26.00 17.67
Market 0 3.31 5.00 8.58 12.65 5.90 5.10 6.44
Micro 5.31 0 3.15 5.05 10.34 11.54 14.58 8.86 6.72 7.59
Small 5.05 0 3.25 5.00 7.62 6.96 9.86 6.65 5.27 6.50
Big 4.50 0 3.37 4.99 7.01 5.06 7.58 4.84 4.39 5.24
Market 0.49 10.30 5.09 2.63 226 1006 893 802
Micro 0.63 -2.08 7.75 4.83 2.00 244 134 364 409 364
Small 5.86 5.84 10.22 4.90 2.48 606 311 644 716 725
Big 7.33 7.49 10.66 5.39 3.48 1259 668 1376 1226 1330
Market D/P= 0 1 2 3
Market 2.55 15.69 10.84 6.19
Micro 0.94 0.71 1.29 1.30 1.16
Small 3.92 3.59 4.23 3.96 3.87
Big 20.40 12.29 26.01 19.83 14.25
Panel A: Average Number of stocks Panel B: Relative Market Cap
Panel C: Dividend Yield (%) Panel D : Annual time series average of Total risk(%)
Panel E:Annual time series Momentum(%) Panel F: Annual Average of Price (Belgian Franc)
Panel G:Annual time series average Size (*10^6)
are then used to create a set of 12 portofolios (Micro/D/P= 0, Micro/1, Micro/2, Micro/3, Small/D/P= 0, Small/1, Small/2, Small/3, Big/D/P= 0, Big/1,
Big/2 and Big/3.We compute the annual average number of stocks in each portfolio. The annual average relative market capital of the stocks in each
portfolio is reported. Dividend Yield is measured as the sum of all dividends paid in the year before year t divided by the current price.Total risk is
the standard deviation of the past 24 to 60 months excess returns. Momentum is measured as the compound returns for six months before the
In this table we show summary statistics of some characteristic identified to capture the cross-section of stock returns. At the beginning of January
for each year t, we sort stocks into three groups (Micros, Small and Big) based on their size (price time's shares outstanding) at the end of the
previous year. We seperately sort the stocks in the same year based on their dividend yield in the previous year. No dividend paying stocks are
assigned one group. The rest of the stocks is splitted into three groups in order of magnitude of their dividend yield. The lowest dividend paying
stocks are place in 1 and the highest dividend paying stocks are place in 3. The intersections of the stocks in the size-sort and dividend-sort
t .We also report the annual averages for univariate sorted characteristic on the market. The sample period is January 1868 -December 1913.
120

zero-dividend paying stocks as compared to the various dividend-paying stocks. This
might be due to the high number of zero dividend paying stocks in the micro size group
(see Panel A). Unsurprisingly, less money is invested in zero-dividend paying Big
stocks. Among the dividend paying Big stocks, the average market capital for the
highest dividend paying stocks is lower. For the Small group of stocks, the zero-
dividend paying stocks have slightly lower market capital than the dividend paying
stocks.
Panel C shows the average dividend yield of the univariate sort on size. It also shows the
dividend yield for the 12 size-dividend yield double-sorted portfolios. The first column
shows very small reduction in dividends as size increases. Among the double-sorted
dividend paying stocks, there is a positive relationship between the lowest dividend yield
stocks and size. On the contrary, the relationship is less negative for the middle and the
highest dividend paying stocks.
In panel D, there is a strong correspondence between total risk and size. As the first
column shows, total risk monotonically reduces as size increases. This relationship
persists in all double-sorted size-dividend yield portfolios. It is worthy to note the high
total risk recorded for zero-dividend paying stocks. For the univariate sort on dividend
yield, the relationship between total risk and dividend yield is U-shaped among
dividend paying stocks. The U-shaped relationship persists in all size groups.
Panel E of Table 5.1 shows the average annual past performance (momentum) of the
univariate size and dividend yield sort portfolios. For univariate size sorts, the first
column shows that size increases with momentum. The positive relationship persists in
121

all columns of size-dividend yield double-sorted portfolios. The zero-dividend paying
stocks performed poorly in the past six months, compared to the dividend paying stocks.
The poor performance may be due to Micro zero-dividend stocks, as they record an
average momentum of -2.08%. It is important to note the strong negative relationship
between momentum and dividend yield among dividend paying stocks. This confirms
the Asness (1997) results and is consistent with the phenomena that value stocks are
stocks that performed lower in the past 6 to 12 months, and vice versa.
Panel F reports the annual average price of the size and dividend yield sorted portfolios.
Unsurprisingly, there is a positive relationship between size and price in both single and
double sort portfolios. It is clear that zero-dividend paying stocks have the lowest price.
This persists for all size groups. This confirms the results of Grossman and Shore (2006),
who found that stocks that do not pay dividends tend to have lower market capitalization
and price.
Finally, panel G reports the average size (in millions) of the stocks in each portfolio
across the period of the study. It is obvious to observe the zero-dividend paying stocks
recording the lowest size in the 19
th
century BSE. Size seems to show a negative
relationship with the dividend yield (single sort among dividend paying stocks). Detailed
analysis of the double-sort portfolios reveals that the Small and the Big stocks
contribute to the negative relationship.
5.3.2 Momentum-Dividend yield double sorts
from the intersections of the momentum sort and dividend yield sort stocks. We also
report univariate sort portfolios on momentum, and again on the dividend yield. As in the
122

previous section, between 1868 and 1914, our sample contains 84 to 518 different firms,
which result in an average of 240 stocks traded every year (see Panel A and Figure 3.1).
On average, there are twelve or more stocks in every double-sorted portfolio. The
number of stocks that performed poorly in the last six months (displaying low
momentum) and do not pay dividends in the past 12 months exceed the dividend-paying
stocks in each portfolio. For the middle momentum group, zero-dividend stocks are less
than the numbers that pay dividends.
In panel B, the average relative market capitalization of each portfolio is reported. The
first column indicates that the middle-momentum portfolio accounts for the largest part
of money invested in the stock market, representing, on average, 45% of total stock
market capitalization. On the contrary, the lowest momentum portfolio accounts for less
than 22% of stock market capitalization. The remaining high momentum stocks represent
only 33% of the market capitalization. For the double-sorted portfolios, it is not
surprising to see less market capital for zero-dividend paying stocks in all momentum
categories. For single-sort dividend paying stocks, the amount of money invested in the
various portfolios decreases as dividend yield increases. This is also true for all
categories of double-sort portfolios. Among the dividend paying portfolios, the middle
momentum portfolios show a higher amount of investment than the other portfolios. On
the contrary, the middle momentum records the lowest amount among the zero-dividend
paying portfolios.
Panel C shows the average dividend yield of the univariate sort on momentum. It also
shows the dividend yield for the 12 momentum-dividend yield double-sorted portfolios.
The first column shows a slight increase in dividend yield as momentum increases to the
123

Table 5.2: Summary statistics for Momentum-Dividend yield double sorts
Market 72 56 56 56 8.96 43.96 29.67 17.42
1 80 37 12 12 19 21.62 3.55 7.11 5.97 4.99
2 79 13 20 25 21 45.32 2.35 21.11 14.63 7.23
3 80 22 23 19 16 33.05 3.06 15.73 9.07 5.19
Market 0 3.31 5.00 8.58 11.74 5.31 4.55 5.75
1 4.78 0 3.29 4.99 8.55 9.29 12.96 6.20 5.71 6.98
2 5.45 0 3.33 5.03 7.32 5.95 10.48 5.16 4.50 5.73
3 4.62 0 3.33 4.98 8.56 8.30 13.29 6.56 5.78 6.85
Market 0.49 10.30 5.09 2.63 226 1006 893 802
1 -16.11 -22.63 -9.29 -9.25 -12.63 468 174 829 830 591
2 2.86 2.77 3.21 2.97 2.60 839 308 1064 942 874
3 26.91 37.08 26.76 18.08 21.20 808 261 1061 911 946
Market D/P= 0 1 2 3
Market 2.55 15.69 10.84 6.19
1 5.49 1.94 11.48 8.75 5.25
2 11.55 3.70 19.74 12.63 7.10
3 8.35 2.86 13.97 9.58 5.84
for each year t, we sort stocks into three groups (1, 2 and 3) based on their momentum ( compound returns in prior sixmonth ) at the end of the
previous year. The lowest momentum stocks are assigned group 1 and the highest assigned group 3. We seperately sort the stocks in the same
year based on their dividend yield in the previous year. No dividend paying stocks are assigned one group (D/P=0). The rest of the stocks is spl-
itted into three groups in order of magnitude of their dividend yield. The lowest dividend paying stocks are place in 1 and the highest dividend
paying stocks are place in 3. The intersections of the stocks in the momentum-sort and dividend-sortare then used to create a set of 12 portofolios
(1/D/P= 0, 1/1, 1/2, 1/3, 2/D/P= 0, 2/1, 2/2, 2/3, 3/D/P= 0, 3/1, 3/2 and 3/3).We compute the annual average number of stocks in each portfolio.
The annual average relative market capital of the stocks in each portfolio is reported. Dividend Yield is measured as the sum of all dividends paid in
the year before year t divided by the current price. Total risk is the standard deviation of the past 24 to 60 months excess returns. Momentum
is measured as the compound returns sixmonths before the t .We also report the annual averages for univariate sorted characteristic
on the market. The sample period is January 1868 -December 1913.
Panel C: Dividend Yield (%) Panel D : Annual time series average of Total risk(%)
Panel E:Annual time series Momentum(%) Panel F: Annual Average of Price (Belgian Franc)
Panel G:Annual time series average Size (*10^6)
124

middle portfolio. Dividend yield increases slightly with momentum in the lowest double-
sorted dividend paying stocks. On the contrary, the relationship is cup-shaped and U-
shaped for the middle and the highest-dividend paying portfolios.
In panel D, the second column shows a very high total risk for the zero-dividend paying
portfolios. This is sometimes almost twice the value of the total risk for the dividend
paying portfolios in each momentum category. Surprisingly, the total risk for the middle
momentum portfolios is the lowest. The total risk for high and low momentum portfolios
is almost the same for all dividend yield groups. This is surprising, as one would expect
the stocks that performed well in the past to have lower total risk than stocks that
performed poorly in the past.
Panel E of Table 5.2 shows the average annual past performance (momentum) of the
univariate and double-sort momentum and dividend yield portfolios. As in the previous
section, for the univariate sort, it is important to note the strong negative relationship
between momentum and dividend yield among dividend-paying stocks. The double-sort
portfolios reveal that the strength of the negative relationship is due to the low and the
middle momentum portfolios.
The average price for the univariate sort and double sort portfolios are reported in panel
F. As before, the prices for zero-dividend paying portfolios are low, sometimes one-third
of the price of their dividend paying counter parts (univariate sort). This persists in all
double-sort momentum groups. For the univariate sorts, there is a negative relationship
between average price and dividend yield among dividend paying stocks. The double-
125

sort portfolios indicate that the negative relationship pulls its strength from the middle
portfolios.
Finally, we show the average size of the single and double-sort portfolios in panel G.
Clearly, the zero-dividend paying stocks are the small size stocks. For the univariate sorts
on dividend yield, the zero-dividend paying portfolio has an average size that is less than
that of the dividend paying portfolios. This persists in all double-sort portfolios. The
negative relationship between size and dividend yield in dividend-paying portfolios is
obvious for both single and double-sort portfolios.
5.3.3 Total risk-Dividend yield double sorts
from the intersections of the total risk sort and dividend yield sort stocks. We also report
the univariate sort for total risk and dividend yield. As in the previous section, between
1868 and 1914, our sample contains 84 to 518 different firms, which result in an average
of 240 stocks traded every year (see Panel A and Figure 3.1). On average, there are five
or more stocks in every double-sorted portfolio. Looking at the rows of the double-sorts,
the number of high-risk zero-dividend stocks exceeds (sometimes by 8 times) the
dividend paying stocks. This is not surprising, as the small stocks are more likely the
high-risk stocks (see Table 5.1).
In panel B, we report the relative amount of money invested in each double-sort total
risk-dividend yield portfolio. For the univariate total risk sorts (first column), there is a
negative relationship between the amount of money invested in each portfolio and its
risk. Perhaps smaller firms are simply less diversified and therefore have a higher total
risk. The double-sort portfolios show that the negative relationship between the amount
126

Table 5.3: Summary statistics of total risk- dividend yield double sorts
Market 72 56 56 56 8.96 43.96 29.67 17.42
1 80 5 24 31 20 61.84 1.47 31.85 21.01 7.51
2 79 15 21 18 25 25.31 2.99 7.90 7.00 7.41
3 80 52 10 6 12 12.85 4.50 4.21 1.65 2.49
Market 0 3.31 5.00 8.58 11.74 5.31 4.55 5.75
1 5.55 0 3.52 4.98 7.03 3.13 3.65 2.84 3.05 3.58
2 6.65 0 3.17 5.03 8.00 6.31 6.71 6.31 6.13 6.19
3 2.68 0 2.99 5.04 9.90 14.06 15.35 12.57 11.07 11.46
Market 0.49 10.30 5.09 2.63 226 1006 893 802
1 3.29 -1.86 4.97 3.71 2.25 902 450 1096 917 792
2 4.17 -2.49 8.13 6.17 3.00 808 336 949 914 885
3 6.35 1.77 22.53 9.92 2.90 406 170 857 659 534
Market D/P= 0 1 2 3
Market 2.55 15.69 10.84 6.19
1 15.54 5.36 25.17 14.07 7.99
2 6.53 3.85 8.16 7.65 5.91
3 3.32 1.88 7.42 4.20 3.48
divided by the current price. Total volatility is the standard deviation of the past 24 to 60 months excess returns. Momentum is measured
as the compound returns for sixmonths before the t . We also report the annual averages for univariate sorted characteristic on the market.
The sample period is January 1868 -December 1913
Panel C: DividendYield(%) Panel D: Annual time series average of Total risk(%)
Panel E:Annual time series momentum(%) Panel F: Annual Average of Price (Belgian Franc)
for each year t, we sort stocks into three groups (1,2 ans 3) based on their total risk (standard deviation of excess returns) at the end of the
previous year. The lowest total risk group are assigned group 1 and highest group 3. We seperately sort the stocks in the same year based on their
dividend yield in the previous year. No dividend paying stocks are assigned one group. The rest of the stocks is splitted into three groups in
order of magnitude of their dividend yield. The lowest dividend paying stocks are place in 1 and the highest dividend paying stocks are place
in 3. The intersections of the stocks in the size-sort and dividend-sort are then used to create a set of 12 portofolios (1/D/P= 0, 1/1, 1/2, 1/3, 2
/D/P= 0, 2/1, 2/2, 2/3, 3/D/P= 0, 3/1, 3/2 and 3/3.We compute the annual average number of stocks in each portfolio. The annual average relative
market capital of the stocks in each portfolio is reported. Dividend yield is measured as the sum of all dividends paid in the year before year t
Panel G:Annual time series average size (*10^6)
127

of investments and total risk is completely due to the dividend paying stocks. In fact, the
relationship is positive for zero-dividend paying stocks. A negative relationship is clearly
visible between dividend yield and the amount of investments. This relationship is
mainly due to the low total risk stocks.
In panel C, dividend payment does not show any sign of increasing or decreasing with
risk (column one). This is reflected in all double-sorted portfolios. In Panel D, total risk
does not show any sign of decrease or increase in the univariate sort dividend yield
portfolios. An exception is the total risk of the zero-dividend portfolio. The double-sort
portfolios show that the high risk recorded by the univariate sort zero-dividend portfolio
draws its power from the highest total risk portfolios.
Panel E reports the past performance of the total risk-dividend yield double-sort
portfolios. We also report the univariate sort portfolio characteristics. Momentum
increases as the firms specific risk also increases for univariate sort portfolios. The
positive relationship is reflected in the lowest and the middle dividend yield portfolios.
As in the previous sections, the negative relationship between dividend and momentum
are prevalent in both single-sort and double-sort portfolios (among dividend paying
portfolios). The lowest momentum recorded by the zero-dividend paying portfolios is
also apparent.
As before, the price for zero-dividend portfolios remains smaller (Panel F). There is a
negative relationship between price and total risk (first column). The negative
relationship persists in zero-dividend, low and middle dividend yield portfolios. As
128

before, the negative relationship between price and dividend yield are still present in the
univariate sort. The negative relationship is persistent in all double-sort portfolios.
Finally, size shows a negative relationship with firm-specific risk (Panel G, column one).
The negative relationship continues to exist in all double-sorted portfolios. As in the
previous sections, size decreases as dividend yield increases among the dividend paying
stocks. This relationship persists in all dividends paying double-sort portfolios. The
average size for zero-dividend paying stocks is lower compared to their dividend-paying
counterparts.
From the previous sections, there is enough evidence to conclude that zero-dividend
stocks had low past performance, small market capitalization, low price and high total
risk in the 19
th
century. This is consistent with the notion that dividends play an
important role in communicating to investors in the early capital markets; they can
assume that the small size and high-risk firms that do not pay dividends are distressed
(see Baskin and Miranti (1997), Cheffins (2006) and Cheffins (2008)). Alternatively,
this may be security for smaller, riskier, lower-priced and younger firms that may have
chosen not to pay a dividend because they were newly established and had more growth
prospects. Often, growth firms in their early lives have low or zero payouts ratios, with
the hope of returning more earnings back to investors at maturity.
To investigate the robustness of the relation among the suggested characteristics, we
compute the cross-sectional correlation matrix for each year. In Table 5.4, we report the
time-series averages of the bivariate correlations. We perform a hypothesis test on the
time series average to test whether it is significantly different from zero. We also include
129

realized annual returns in the matrix. We indeed retrieve the negative correlations
between total risk on the one hand and dividend yield and size on the other hand. Both
are significantly different from zero at the 1% level. Likewise, size and momentum are
significantly positively related.
Table 5.4: Annual Time Series Average of the correlation between the entire characteristic and Average return

This is also visible in panel E of Table 5.1. It is not surprising to see the positive
relationship between size and dividend yield, as the relationship would be influenced by
the high number of zero-dividend stocks (see sections 5.3.1-5.3.3). The dummy variable
is correlated with all the other characteristics except the average returns. Investigating the
correlations with realized returns in the last line of Table 5.4, we notice that all are small
in magnitude. Only two are significantly different from zero at the 5% level: the negative
Dummy DY In(Size)
i
Mom AR
Dummy
DY 0.609
***
In(Size) 0.503 0.187
** **
i
-0.509 -0.268 -0.570
*** *** ***
Mom 0.139 0.027 0.157 -0.015
** ***
AR -0.001 -0.005 -0.072 0.037 0.071
** **
log of size,
i
= total risk, DY = dividend yield, Mom = momentum. We also report
stocks.
a dummy variable which is 1 for zero-dividend stocks and 0 for dividend paying
cance at 1 % , ** = signifincance at 5 %. AR=average returns, In(Size) = natural
Note: This table shows the time series averages of the bivariate annual cross-
sectional correlation between the various characteristics define to explain average
returns. We test the hypothesis that, the time series average of the bivariate corre-
lation is equal to zero.The significance of the test are in parenthesis. *** = signifi-
130

correlation with size and the positive correlation with past returns. We now turn to more
formal testing of any predictive relation between stock characteristics and future returns.
Average Excess Returns on Portfolio Sorts 5.4
In this section, we investigate the relationship between stock characteristics and excess
returns using tercile portfolios. As before, we sort the stocks for each year into tercile
groups based on a specific characteristic (size, momentum and total risk) observed the
previous December. We sort stocks separately based on the dividend yield in the
previous year. We assign zero-dividend stocks to one group. The rest of the stocks are
divided into terciles. Twelve sets of portfolios are formed from the intersection of stocks
from the dividend yield sort group and the size, momentum or total sort groups. We then
hold these portfolios for the entire year and compute monthly returns. We compute
portfolio returns either using equal weights or value weights. Both weighting schemas
are complementary: with equal weights, the smaller stocks dominate, whereas the large
caps stand out with value weights. In addition, we form portfolios on univariate sort
characteristics and hold the position for one year. The FM breakpoint is used in all
portfolio formations. Repeating the portfolio formations each month yields 552 portfolio
returns in our sample period. We report the average excess returns and their
corresponding Newey-West autocorrelation, heteroskedastic standard-error adjusted t-
statistics for each portfolio. We also report the difference in average excess returns and
their t-statistics for the highest and the lowest terciles characteristic portfolios (hedge
portfolio). For the dividend yield, we show the difference in average excess returns for
the highest dividend-paying portfolios on the one hand, and the lowest and zero-dividend
paying portfolio on the other hand. The results are reported in Table 5.3.
131

Table 5.5: Equal and Value-weighted portfolios excess returns (%) of double-sorted characteristics
Market D/P= 0 1 2 3 3-0 3-1 Market D/P= 0 1 2 3 3-0 3-1
Panel A: Sorti ng on Si ze and Di vi dend Yi el d
Market 0.27 0.14 0.34 0.26 -0.01 0.12 1.33 1.43 3.59 2.19 -0.05 1.23
Mi cro 0.47 0.53 0.00 0.53 0.17 -0.36 0.11 2.73 2.21 0.02 3.15 1.17 -1.58 0.50
Smal l 0.22 -0.11 0.21 0.27 0.48 0.62 0.26 1.74 -0.48 1.48 2.63 2.00 2.19 1.10
Bi g 0.19 0.00 0.14 0.33 0.22 0.35 0.08 1.91 -0.02 1.43 3.09 1.69 2.06 0.74
Bi g-Mi cro -0.28 -0.28 -0.65 0.07 -0.20 0.05 -2.20 -2.20 -2.86 0.34 -1.36 0.33
Market 0.01 0.19 0.38 0.42 0.42 0.24 0.04 2.50 4.36 3.03 2.70 1.94
Mi cro 0.16 0.03 0.04 0.53 0.06 0.03 0.01 1.16 0.13 0.25 2.98 0.46 0.17 0.08
Smal l 0.35 0.01 0.26 0.33 0.71 0.70 0.46 2.75 0.06 1.89 3.32 2.43 2.20 1.54
Bi g 0.25 0.09 0.19 0.38 0.33 0.24 0.14 3.15 0.49 2.54 4.18 2.64 1.43 1.37
Bi g-Mi cro 0.09 0.06 0.14 -0.15 0.27 0.83 0.31 0.86 -0.93 1.95
EW Mean Monthly Returns(%) t-statistics for EW
VW Mean Monthly Returns(%) t-statistics for VW
dividends in the past one year are assigned one group. The rest of the stocks is split into three groups based on their dividend yield. The stocks in the
intersections of dividend sorts and the Size, Momentum, and Idiosyncratic sorts are grouped to form equally and value weighted portoflios. For the size
and dividend yield sort, we have the following 12 portfolios, Micro/D/P= 0, Micro/1, Micro/2, Micro/3, Small/D/P= 0, Small/1, Small/2, Small/3, Big/D/P= 0,
Big/1, Big/2 and Big/3. Similar portfolio formations are repeated for Momentum and Idiosyncratic risk. Size is measured as the price times the number of
shares outstanding in December of the year before portfolios formations. Momentum is the measured as the compound returns of the stock six month
the portfolios formation year. Total risk is the standard deviation of the past 24 to 60 months excess returns. Dividend yield is the sum of all
dividends paid in the year before the protfolio formation divided by the price at December of the same year. We also report the univariate sort on Divide-
nd Yield, Size, Momentum and Total risk. Portfolios are formed annually. We report the Newey-West Heteroskedastic autocorellation adjsuted
t-statistic for the average excess return for each portfolio. The sample period is Jan. 1868- Dec. 1913.
EW= Equally Weighted, VW= Value Weighted
At the beginning of January for each year t , the 19th centrury BSE stocks are allocated to three groups based on their sorted size (market capitalization),
Momentum and Idiosyncratic Risk in the previous year t-1. The BSE stocks are also sorted on dividend yield and grouped into 4. Stocks which do not pay
132

Table 5.5 continued
Panel B: Sorti ng on Momentum and Di vi dend Yi el d
Market 0.27 0.14 0.34 0.26 -0.01 0.12 1.33 1.43 3.59 2.19 -0.05 1.23
1 0.17 0.30 0.02 0.14 -0.13 -0.43 -0.14 1.16 1.39 0.11 0.86 -0.72 -1.84 -0.70
2 0.26 -0.18 0.13 0.36 0.42 0.59 0.29 2.86 -0.72 1.44 3.67 3.95 2.46 2.93
3 0.60 0.60 0.44 0.62 0.56 -0.01 0.12 4.21 2.20 3.35 4.73 4.44 -0.06 1.11
3-1 0.43 0.25 0.42 0.49 0.69 3.05 0.91 2.38 2.78 3.75
Market 0.01 0.19 0.38 0.42 0.42 0.24 0.04 2.50 4.36 3.03 2.70 1.94
1 -0.09 -0.28 -0.21 0.14 0.11 0.39 0.32 -0.72 -1.53 -1.18 0.95 0.42 1.39 1.09
2 0.21 -0.22 0.16 0.31 0.36 0.58 0.21 2.86 -0.94 2.07 3.31 3.30 2.64 2.03
3 0.56 0.39 0.53 0.60 0.63 0.24 0.10 4.69 1.76 4.13 4.78 4.63 1.25 0.81
3-1 0.66 0.67 0.74 0.46 0.52 4.90 3.05 3.89 2.64 1.85
Panel C: Sorti ng on Total ri sk and Di vi dend Yi el d
Market 0.27 0.14 0.34 0.26 -0.01 0.12 1.33 1.43 3.59 2.19 -0.05 1.23
1 0.29 -0.35 0.17 0.36 0.42 0.77 0.25 4.07 -1.30 2.39 4.74 2.91 2.69 1.78
2 0.19 0.13 0.18 0.23 0.20 0.07 0.02 1.58 0.68 1.27 1.69 1.61 0.43 0.15
3 0.39 0.40 0.07 0.90 -0.06 -0.45 -0.13 1.92 1.59 0.38 3.47 -0.29 -1.94 -0.61
3-1 0.10 0.75 -0.10 0.54 -0.47 0.56 2.17 -0.57 2.29 -2.31
Market 0.01 0.19 0.38 0.42 0.42 0.24 0.04 2.50 4.36 3.03 2.70 1.94
1 0.21 -0.23 0.16 0.31 0.30 0.53 0.15 3.29 -0.81 2.68 3.84 1.67 1.65 0.82
2 0.23 0.11 0.23 0.25 0.29 0.18 0.06 1.86 0.60 1.58 1.78 2.24 1.15 0.53
3 0.08 -0.14 -0.07 0.82 -0.04 0.10 0.03 0.43 -0.61 -0.36 2.99 -0.20 0.48 0.15
3-1 -0.13 0.08 -0.23 0.51 -0.34 -0.86 0.25 -1.22 2.05 -1.44
EW= Equally Weighted, VW= Value Weighted
133

In panel A, average excess returns and t-statistics for the sorting on size and dividend yield
are shown. When sorting on dividend yield across all stocks, there is no obvious pattern in
equally weighted portfolio returns. There is no significant difference in average returns of the
highest and the lowest dividend yield portfolios (12 basis points with t-statistic of 1.23). The
difference in average returns of the zero-dividend yield and highest dividend yield portfolio is
not significant.
We find high returns for zero-dividend stocks, which is consistent with Grossman and Shore
(2006) and Gwilym et al. (2009). For value-weighted portfolios, there is a monotonic
increase in average excess returns for univariate-sort dividend yield portfolios. The difference
in the average excess returns of the highest and the lowest dividend yield portfolios is 24 basis
points, with a t-statistic of 1.94. The profit improves when the investor goes long on the high
dividend yield and short on the zero-dividend yield portfolios. The difference in average
excess return increases to 42 basis points with the t-statistic of 2.70. This shows that
weighting stocks by their market capital in portfolio formations, a strategy that goes long on a
high-dividend yield portfolio and short on a low-dividend yield portfolio, results in a
significantly positive profit. The result is consistent with the findings for the UK market by
Dimson, Marsh and Staunton (2002) and Grossman and Shore (2006).
For equally weighted double-sort portfolios, among the size groups, Small and Big stocks
show a significantly positive difference in the highest and the zero-dividend yield portfolios.
Average returns increase monotonically from zero-dividend yield stocks to the highest
dividend yield stocks (except the middle dividend paying portfolio in Big stocks). The
positive significant difference disappears among the dividend yield stocks. It is obvious that
the size effect exists in the cross-section of the stocks. However, the effect obtains its power
from the Micro and zero-dividend yield stocks. These are the securities for distressed firms,
as they have low prices and negative past performance (see Tables 5.1 to 5.3). For value-
134

weighted portfolios, within the Small size groups, the dividend yield shows a strong positive
relationship with the average excess returns. Surprisingly, the positive relationship between
dividend yield and average returns disappear in the Big size group. It is not surprising to see
the size effect disappear when stocks are value-weighted to form portfolios. This confirms the
result from Chapter 3. The result adds credence to the view that the initial evidence for the
size effect is due to data mining.
In Panel B, double-sort portfolios are formed based on dividend yield and momentum. In
Table 5.1, dividend yield is negatively related to momentum; however, momentum is
positively related to average returns (Chapter 4). In view of this, one would expect
momentum to be an additional predictor of returns holding the dividend yield constant, and
vice versa. Panel B confirms momentum as a predictor of future returns. An exceptional case
is the insignificant average excess return, spread between the high momentum portfolio and
the lower momentum portfolio, for zero-dividend yield portfolios. The spread becomes
significant when stocks are value-weighted. In general, there is a strong momentum effect and
a significant dividend yield effect when stocks are value-weighted.
For the double-sort portfolios, the spread between the average excess returns of the high and
low dividend yield portfolios in the momentum tercile groups is not always significant. The
dividend yield only shows a linear pattern with average excess returns in the middle
momentum group for equally weighted portfolios. The average excess return for the hedge
portfolio is 29 basis points, with a t-statistic of 2.93. This value increases to 59 basis points
when the zero-dividend yield portfolio is used to compute the hedge portfolio. The positive
relationship between dividend yield and average excess return in the middle momentum
portfolios is robust to the weighting scheme used to form portfolios. The relationship still
exists for value-weighted portfolios. The notable exception is the negative spread in average
excess returns for the lowest and the highest momentum groups in equally weighted
135

portfolios, which becomes positive when stocks are value-weighted. Surprisingly, the value-
weighted portfolios average excess return spread for the largest momentum stocks is not
significant (the average excess return spread of 10 basis points with a t-statistic of 0.81). The
results show that among dividend yield stocks, the relationship between dividend yield and
average excess return in the middle momentum group is strong. However, the relationship
becomes stronger when zero-dividend yield stocks are considered. This finding supports those
of Gwilym et al. (2009), who found a strong relationship between dividend yield and average
excess returns among dividend-paying stocks in their quintile momentum portfolios. On zero-
dividend yield stocks, Grossman and Shore (2006) have shown that stocks that do not pay
dividends and have performed poorly in the past have very high future returns, while stocks
that do not pay dividends and have a high past return have lower future returns. We record
opposite results on the 19
th
BSE when stocks are equally weighted, and are even stronger
when stocks are value-weighted. This shows that, for stocks that do not pay dividends, past
performance is most likely to determine future returns.
We now turn to total risk. The evidence presented in panel C does not show clear patterns.
There does not seem to be a monotonous relation between total risk and average returns,
neither in the entire market, nor across the dividend yield quartile, regardless of how returns
are weighted. When focusing on the hedge portfolios, the only difference that is statistically
significant is the middle dividend-yield portfolio, where high total risk translates into high
average excess returns. The difference is 54 and 51 basis points per month, 2.29 and 2.05
standard deviations from zero for equally weighted and value-weighted portfolios,
respectively. Our result does not support Blitz and Van Vliet (2007), who argue based on
Sharpe Ratio and alpha to confirm a negative relationship between total risk and average
excess returns on international markets. However, the results confirm Bali, Cakici, Yan and
Zhang (2005), which show that the positive relationship between total risk and expected
136

returns documented by Goyal and Santa-Clara (2003) is driven by small stocks (traded on
NASDAQ, in their case). Looking across dividend yield groups, there is no statistically
significant result in the hedge portfolios, be it equally weighted or value weighted. However,
we repeat that the absence of a monotonous pattern across the total risk portfolios in the entire
market, as well as the lack of any significant results for the high dividend yield portfolios,
indicates that total risk is not a pervasive characteristic of price in the 19
th
century BSE
market. Generally, there is no total effect.
The Cross-Sectional regressions 5.5
From Table 5.4, it is clear that our characteristics are cross-sectionally correlated. In order to
study the marginal effects of the characteristics, we resort to the FM cross-sectional
regression method.
Every month, we do a regression of the cross-section of individual stock excess returns on the
characteristics proposed to explain the average excess return. We update characteristics
annually in January, based on information available at the end of December. The fact that
characteristics are recomputed every year introduces some time variations. The full usage of
the available individual stocks ensures the maximum utilization of information about the
cross-sectional behavior of individual stocks, which might have been lost when portfolios are
formed. We run the cross-sectional regression equation of the form
(13)
where is the excess return with , the return on the individual stocks in the month
t, and is the short-rate used as a proxy for the risk-free rate. DY
j,t-1
is the dividend yield
estimated in the previous year.

(price times shares outstanding) is the size of a stock
in December of the year before, and is the total risk measured as the standard deviation
( )
0 1 1 2 1 3 1 4 1 5 1
DY Size Mom Dum
c
o ,

= + + + + + +
, , , , ,
ln ,
jt ft t t j t t j t t j t t j t t j t jt
R R
jt ft
R R
jt
R
ft
R
, 1
Size
j t
, 1 j t c
o

137

of the past 24 to 60 months excess returns.

is the momentum estimated as the
compound gross return from June to November of the year before portfolio formation. Dum
j,t-
1
is the dummy variable which takes the value 1 for zero-dividend yield stocks and 0 for
dividend paying stocks. From Figure 5.1, the number of zero-dividend stocks constitutes
about 28% of the stocks in the cross-section each year. To eliminate the effect of zero-
dividend stocks on the dividend yield-excess return relationship, one has to delete these stocks
or add a dummy variable. The dummy variable takes a value of 1 when a stock does not pay
dividends, and 0 for all other stocks. Deleting the zero-dividend stocks has the disadvantage
of reducing the cross-sectional information on stocks. The advantage of using the dummy
variable is to allow full usage of the cross-sectional information on individual stocks. In
addition, using the dummy variable allows a direct measurement and a test of significance of
the difference in behavior of the excess returns in the zero-dividend stocks.

is the
vector of regression coefficients and the regression error. We also estimate the regression
with subsets of the characteristics. The estimation of the cross-sectional regression every
month yields 552 time series for regression coefficients. The time-series average of the slope
coefficients is the estimated premium earned for the different exposures. The averages are
tested for statistical significance using heteroskedastic autocorrelation corrected standard
errors. We use the Newey and West (1987) correction with T
1/4
lags.
9

The average of the coefficients and their corresponding Newey-West standard error adjusted
t-statistic (parenthesis) is reported in Table 5.6. In Panel A, model 1 confirms the positive
(negative) relationship between momentum (size) and average excess returns. The importance
of size is consistent with the post WWII-USA evidence presented by Fama and French
(1992). The significance of the size premium is also consistent with our equally weighted

9
We used , where T is 552, the number of months in our sample.

, 1
Mom
j t
( )
0 5
, ,
t t

jt
,
( )
1 4
int T
138

portfolio sorts. This is not surprising, as the regression observations are unweighted, giving
undue importance to the Micro size stocks. We return to this issue in the next section. Total
risk does not show a significant relationship with average returns (average
Table 5.6: Cross-Sectional Regression of Excess Returns on Dividend Yield, Size, Total Risk and Momentum

Dividend Total Dividend
Model Intercept Yield Size Risk Momentum Dummy
1 1.88% -0.15% 0.51% 0.71%
(2.74) (-4.13) (0.38) (2.21)
2 1.81% 0.51% -0.15% 0.63% 0.69%
(2.66) (0.37) (-4.10) (0.47) (2.14)
3 2.58% -1.64% -0.18% 1.13% 0.54% -0.41%
(3.63) (-0.82) (-4.81) (0.82) (1.63) (-2.60)
Subperiods
Jan.1868-Dec. 1877 -0.08% -6.23% 0.04% 4.35% 0.03% -0.89%
(-0.05) (-1.83) (0.55) (1.06) (0.03) (-2.46)
Jan. 1878-Dec. 1887 6.22% 4.80% -0.34% -0.49% -0.83% -0.27%
(2.85) (0.68) (-3.08) (-0.16) (-1.04) (-0.61)
Jan. 1888-Dec. 1897 1.64% -0.21% -0.18% -1.60% 1.78% 0.16%
(1.31) (-0.06) (-2.66) (-0.61) (3.08) 0.55
Jan. 1898-Dec. 1907 3.19% -4.42% -0.25% -0.22% 1.10% -0.58%
(2.66) (-1.57) (-3.39) (-0.11) (2.26) (-1.94)
Jan. 1908-Dec. 1913 1.68% -2.37% -0.15% 4.65% 0.61% -0.48%
(1.63) (-0.75) (-2.64) (1.47) (1.35) (-1.65)
Panel A
Panel B
is Jan. 1868- Dec. 1913.
This table reports the coefficients of the monthly cross-sectional regressions of the excess return on the characte-
ristic.The characteristics identified in this regression are dividend yield size, momentum and idiosyncratic risk. We
also consider a dummy variable for dividend paying and zero-dividend paying stocks in the last 12 months before
the portfolios formation year. The dummy variable is assigned a value is assigned a value of 1 for zero-dividend
paying stocks and 0 for dividend paying stocks. Each year, the characteristics are measured based on the inform-
ation available prior to the year. Size is measured as price times number of shares outstanding. Momentum is comp-
uted as 6 months compound returns prior to the regression year. Dividend yield is the sum of dividends paid in the
last 12 months dividend by the current price. Total risk is the standard deviation of the past 24 to 60 months excess
returns.The interactions between size, momentum and dividend dummy are considered. Newey-West autocorrelation
and heteroskedastic adjusted t-statistics are in parentheses.We compute characteristics each year.The sample period
139

coefficient is 0.51 with a t-statistic of only 0.38). This confirms the sorting result, as for the
entire market, the average total risk hedge portfolio is 10 basis points with a t-statistic of 0.56.
Model 1 in Panel A confirms the positive relationship between past returns and future returns
in the equal and value weight sorts. Adding dividend yield to the regressors in model 2 shows
that the dividend yield does not relate to excess returns. The average coefficient is 0.51% with
a t-statistic of 0.37. This is not surprising, as the cross-sectional regression gives equal weight
to all stocks, and it confirms the results for equal weight sorts on dividends. The momentum
and size premiums are still significant in model 2. Including the dividend yield and its dummy
in regression model 3 reveals a negative insignificant relationship between dividend and
average excess returns. However, the dummy variable has a negative significant relationship
with average returns. This shows that there is a negative significant difference in average
returns between the zero-dividend and dividend paying stocks. Specifically, zero-dividend
paying stocks have an average excess return less than the dividend paying stock. The value -
0.41 is significant at a 5% level. The intercepts in models 1 to 3 are significantly different
from zero, and it is far greater than the risk-free rate.
As a further check on the form of these relationships, we run the regression in model 3 on five
(four ten-year and a five-year sub-periods) non-overlapping sub-periods between 1868 and
1914. The result in Panel B shows that most periods support the overall results in model 3.
The relationship between size and average excess returns does not persist in all sub-periods.
Size is not related to excess returns in the first ten-year period. Momentum seems to pull its
power from the last twenty-five years of the study. Dividend yield does not show a significant
relationship in any of the sub-periods. The effect of the dummy variable is not always
significant in the sub-periods. Table 5.6 confirms the result from the sorting method for
equally weighted portfolio formations.

140

5.5.1 Pervasiveness of the Cross-Sectional Relationships
From a broad economic point of view, for a characteristic to be accepted and to explain the
cross-section of the returns, its effect should be market-wide. The danger with the cross-
sectional regression approach is that illiquid, small stocks drive the results, as the observations
are not weighted. Recent evidence shows that some characteristics are not pervasive. For
example, Bali and Cakici (2008) documents that the relationship between total volatility and
the expected stock returns are not robust, but depend on the (i) weighting scheme used to
compute average portfolio returns, (ii) the frequency of the return data used to estimate total
volatility and, (iii) the breakpoints used to sort stocks into quintiles. Likewise, Horowitz et al.
(2000) indicate that when firms with less than $5 million in value are excluded, the size effect
is considerably reduced and becomes statistically insignificant. Using equally weighted
portfolios and ignoring transaction costs therefore overstates the size effect. This seems to be
consistent with Schwert (2003). He reports that the abnormal performance of the Dimensional
Fund Advisors (DFA) US 9-10 Small Company Portfolio, which invests in the two lowest
decile of stocks by market value, has been close to zero since 1982. Also, Fama and French
(2008) report that the asset growth anomaly is prevalent in their so-called small and micro
stocks.
We investigate whether our results suffer from this caveat by separately running the Fama-
MacBeth analysis for our Micro and large (Small plus 'Big) stocks. Table 5.7 reports the
time series averages of the monthly cross-sectional regression coefficients. First, the evidence
for size is not consistent across size groups (model 1). It appears that the microcap group,
which represents, on average, only 3.67% of market capitalization (Panel B), drives the
negative relationship between size and average return.
Unsurprisingly, total risk does not have a relationship with average excess returns in the large
stocks (Panel A). In model 2, introducing the dividend yield in the regressions further reduces
141

the importance of total risk. It is still negative for the large stocks, but is statistically
insignificant. These conclusions do not change when the dividend yield dummy is included in
the regression. As an expectation, dividend yield has a consistent significant positive
relationship with average excess returns in the large stocks. The average coefficient of the
dummy variable is no more significant.
Table 5.7: Cross-Sectional regression of Excess Returns on Dividend Yield, Size, Total Risk and Momentum of Size
subsamples

Panel B confirms that the size effect pulls its strength from the Micro stocks. The strength of
the dividend yield effect and the momentum effect is concentrated in the large stocks, which
Dividend Total Dividend
Model intercept Yield Size Risk Momentum Dummy
3 -0.86% -0.06% -2.46% 2.23%
(-1.13) (-1.54) (-1.42) (6.68)
4 -1.15% 5.55% -0.05% -1.41% 2.12%
(-1.63) (2.04) (-1.36) (-0.95) (6.53)
5 -1.42% 8.36% -0.03% -1.41% 1.98% 0.09%
(-2.16) (1.64) (-0.99) (-0.88) (6.06) (0.38)
1 9.02% -0.64% 0.09% 0.05%
(4.66) (-4.68) (0.04) (0.10)
2 8.96% -0.59% -0.63% 0.00% -0.01%
(4.62) (-0.28) (-4.59) (-0.00) (-0.02)
3 10.27% -4.06% -0.69% 0.74% -0.26% -0.57%
(5.06) (-1.35) (-4.84) (0.36) (-0.51) (-2.16)
paying stocks. Each year, the characteristics are measured based on the information available prior to the year.
Size is measured as price times number of shares outstanding. Momentum is computed as 6 months compound retur-
ns prior to the regression year. Dividend yield is the sum of dividends paid in the last 12 months dividend by the
current price. Total risk is the standard deviation of the past 24 to 60 months excess returns.The interactions between
size, momentum and dividend dummy are considered. Newey-West autocorrelation and heteroskedastic adjusted
t-statistics are in parentheses.We compute characteristic each year.The sample period is Jan.1868- Dec.1913.
Panel A (Small and Big)
Panel B (Micro)
This table reports the coefficients of the monthly cross-sectional regressions of the excess return on the characte-
ristic without the Micro size stocks. We seperately perform similar analysis on Micro stock.The characteristics iden-
tified in this regression are dividend yield size, momentum and idiosyncratic risk.We also consider a dummy vari-
able for dividend paying and zero-dividend paying stocks in the last 12 months before the portfolios formation year.
The dummy variable is assigned a value is assigned a value of 1 for zero-dividend paying stocks and 0 for dividend
142

accounts for about 96% of market capital. The dummy variable contributes its effect in the
Micro stocks. Table 5.7 confirms the results from the sorting method, as the value-weighted
dividend sort portfolios seem to have the positive relationship with average excess returns.
The result is not surprising, as Elton and Gruber (1983) obtained similar results on the USA
market between the years 1927 and 1976, using NYSE data. Finally, the results for
momentum are quite consistent across size groups. In all models, we find a positive
relationship between past returns and average realized returns. The premium for momentum
significantly ranges from 1.98% to 2.23% (compared to 0.54% and 0.71% from the full
sample regressions).

143

Conclusion 5.6
Since beta fails to explain the cross-section of stock returns in the 19
th
century BSE, we
document the cross-sectional relationship between average excess returns and size, total risk,
momentum and dividend yield. We investigate these relationships with completely out of
sample data in the 19
th
and first few years of the 20th century from the Brussels Stock
Exchange. We also investigate the pervasiveness of the cross-sectional relationships across
different size groups. We use sorting and cross-sectional regressions in the analyses.
Unsurprisingly, we find a significantly negative relationship between size and expected
returns. However, further investigation reveals that the negative significant relationship
between average return and size is completely driven by our Micro stocks, accounting for
about 3.67% of the market capitalization. There is no consistent pattern to be found for total
risk. The momentum effect, on the other hand, does not exist in zero-dividend yield stocks but
strong in dividend paying stocks for equally weighted sorts (average regression coefficient
with t-statistics not less than two standard errors from zero). Momentum can explain average
excess returns of stocks, which account for about 96% of the market capital (Small and
Big stocks groups).
Dividend yield is negatively related to momentum and each of them is positively related to
excess returns. Further investigation reveals that the relationship between excess return and
dividend yield does not exist among Micro group, which are mostly zero-dividend yield
stocks. The relationship is significant in Small and Big stocks.

144

6 CONCLUSION

The behavior of the aggregate U.K stock markets before World War I is similar in many ways to that
of the modern US markets. The relative size of the market and the relative numbers of large and small
stocks are also similar. However, the cross section of stocks looks quite different in the two samples
Grossman and Shore (2006)
In this doctoral thesis, we answered the following research questions:
- Is beta (systematic risk) stable, unbiased and robust to outliers in the 19
th
century BSE?
- Does the CAPM provide a good description for expected returns in the 19
th
century BSE?
- Does size affect the cross-section of stock returns on the 19
th
century BSE market?
- Does the momentum effect exist in the 19
th
century? If it exists, what is its source?
- Does total risk predict returns on the 19
th
century BSE?
- Does dividend yield predict returns on the 19
th
century BSE?
- What is the marginal effect of the above characteristics on the cross-section of stock returns?
This doctoral dissertation answered the above questions, thereby contributing to the existing
literature on the cross-sectional predictability of stock returns. More specifically, we
examined the robustness of the cross-sectional predictability of stock returns. The study used
the 19
th
and the first few years of the 20
th
century Brussels stock exchange data. To this end,
we conducted four empirical studies to answer the above research questions. These studies
covered the assessment of betas (the main inputs in the CAPM), testing of the CAPM and the
effects of size, the presence and source of momentum and the combined effects of size,
momentum, dividend yield and total risk on the cross-section of stock returns.
We found that market model betas for individual stocks were poor predictors of future betas.
Predictability was improved by adjusting betas with Blume and Vasiceks autoregressive
methods. Grouping stocks to form portfolios also improved beta stability. On the 19
th
century
145

BSE, non-synchronous trading effects were not prevalent. This may have been due to the
measurement interval of returns, as monthly returns were used to compute betas. We also
found that the iterative re-weighted least square method, which accounts for outliers in the
estimation of beta, produced betas that were not significantly different from the market model
betas, in terms of predictive accuracy. By studying the behavior of betas in the 19
th
century,
we produced sufficient evidence of how betas should be adjusted for instability and bias when
testing the CAPM. Overall, our results also suggested that betas on the 19
th
century BSE were
biased and not stable, as in the post-World War II period.
In the second study, we examined the validity of the CAPM for the 19
th
century BSE. We also
investigated whether size effects determined cross-sectional variation in stock returns. First,
sorting and cross-sectional regression methods were used to investigate the relationship
between asset beta and the cross-section of stock returns. We found no relationship between
beta and average excess return for the various estimates of beta. Our results indicated that the
CAPM is not a valid model for capturing cross-sectional variations in stock returns on the 19
th

century BSE. The results also confirmed Fama and French (1992) finding that beta has a flat
relationship with the cross-section of average returns in the US market. Second, we adopted
sorting and cross-sectional regression methods to test the cross-sectional relationship between
size and average excess return. We found a strong relationship between size and average
excess return on the 19
th
century BSE. To this end, size could have been used to capture the
cross-sectional variation of stock returns on the 19
th
century BSE. This would have confirmed
results from Banz (1981), Reinganum (1981), (1983), Chan et al. (1985) and Chan and Chen
(1988) and Fama and French (1992). However, the relationship was not cross-sectionally
robust. Detailed investigation confirmed that the size effect drew its power from a small group
of stocks accounting for about 0.35% of market capital. The size effect disappeared when
these stocks were omitted. In effect, we could not rely on size as a cross-sectional predictor of
146

returns on the 19
th
century BSE, as it was confined to a group of stocks representing a small
fraction of market wealth. Our results implied that size should not be considered a systematic
proxy for risk, which corroborated Horowitz et al. (2000) findings for the US market between
1963 and 1981. The results also confirmed the findings of Fama and French (2008), who used
US data from 1963 to 2004 to document that the size effect owes much of its power to micro
caps and that it is marginal for small and big caps. In our data, the size effect did not exist for
small and large stocks.
The third study investigated the relationship between the short-term past performance of
stocks (momentum) and their future short-term performance. We found that momentum
existed in stocks, which constituted, on average, more than 90% of market wealth. Our
finding that large stocks had momentum showed that momentum could be used as a firm
attribute when predicting returns in the cross-section of the 19
th
century BSE. The presence of
momentum on the 19
th
century BSE also provided evidence that the momentum profit found
in post-WWII US and other markets was not due to data snooping bias. We further
investigated the source of momentum profit in the 19
th
century and found that profit reversed
two to five years after portfolio formation. This evidence supported Jegadeesh and Titman
(2001) results, and it is contrary to that of Conrad and Kaul (1998). Therefore, the cross-
section of expected returns cannot explain the momentum profit. However, investigating the
momentum profit in size sub-samples showed that post-holding period reversal was mainly
due to Micro size stocks. Therefore, the Jegadeesh and Titman (2001) behavioral
explanation of momentum profit should be interpreted with caution. Extensions of behavioral
theory have postulated that investors aggregate overconfidence is high following market
gains. This compelled us to test whether momentum profit in the cross-section of stocks
depended on the state of the market. We found that momentum depended on the state of the
market when a three-year lagged return was used to define the state of the market.
147

Dependence of momentum on market state for the 19
th
century BSE added to the behavioral
explanation of momentum profit. This result was not surprising, as it corroborated Cooper et
al. (2004) and Chabot et al. (2009) results on the contemporary US and Victorian Era UK
markets, respectively.
In the last study, we examined the combined effects of size, momentum, total risk and
dividend yield on average returns. We included size to investigate whether the other effects
were confined to small and illiquid groups of stocks. Given its importance, momentum was
also included, to investigate the marginality of its effect in the presence of other
characteristics. In addition, we did not compute momentum as we did in the third study (i.e.
did not use the portfolio-overlapping method). Here, we measured momentum as the
compound return of individual stocks over six-month periods. We used six months compound
returns, because the six formation and six month holding period strategy has been the most
profitable in previous studies. As stated earlier, following Basu (1983), Rosenberg, Reid and
Lanstein (1985) and Fama and French (1992), we would have added price-earnings ratio or
book-to-market ratio as a measure of value or growth. However, data for the computation of
price-earnings and book-to-market ratios were not available for the 19
th
century BSE. As a
result, dividend and price, which were readily available, were used to compute dividend yield,
which serves as a proxy for value. We considered total risk, as investors might not have been
able to hold the market portfolio, due to transaction cost constraints. Therefore, total risk
could also be priced to compensate rational investors for their inability to hold the market
portfolio. In this case, total risk was useful in explaining the cross-section of stock returns.
We found that the size-effect that existed in the 19
th
century was driven by our so-called
Micro stocks, of which more than 50% had a zero-dividend yield, negative momentum, low
price and very high total risk. Low price, coupled with low momentum, high total risk and no
dividend payment, may point to firms that were distressed. Total risk did not show any
148

consistent relationship with average return. This confirmed Bali and Cakici (2008) finding
that (i) the interval of return measurement used to estimate risk (ii) the portfolio breakpoint
method (iii) the weighting scheme use to compute portfolio returns and (iv) using liquidity
and price filters to screen stocks determines the existence and significance of the cross-
sectional relationship between risk and expected returns.
Among dividend paying stocks, momentum was negatively related to dividend yield.
However, each was positively related to excess stock return for our large stocks. These
results were not different from what Asness (1997) and Gwilym et al. (2009) documented
recently in the US and UK markets, respectively.
Based on the predictability of stock market returns, we found several similarities between
contemporary stock markets and the 19
th
century Brussels Stocks Exchange. Specifically,
CAPM was not valid for the 19
th
century BSE, as it is with current markets. At first sight,
there seem to be evidence for the size effect, but the effect disappears when stocks in the
lowest size decile are eliminated on the 19
th
century BSE. The findings on size effect confirm
Fama and French (2008) results, who conjectured that the effect draws its power from micro
stocks on the recent US market. The momentum effect existed on the 19
th
century BSE, as on
contemporary markets. Total risk did not show any consistent relationship with average
excess returns on the 19
th
century BSE. Similar results are found for contemporary markets,
where positive, negative and no relationships have been found between total risk and average
excess return (see Ang et al. (2006), Bali and Cakici (2008), Ang, Hodrick, Xing and Zhang
(2009), Fu (2009)). Dividend yield showed a positive relationship with average returns on the
historical BSE. A similarly positive relationship also existed in the current UK and US
markets (as in Asness (1997) and Gwilym et al. (2009)). The positive relationship between
dividend yield and expected returns shows the presence of value effect on the 19
th
century
BSE as the dividend yield is used to represent value.
149

Although any inference across historical periods and systems must be interpreted with care,
the similarities that we found with the 19
th
century BSE mostly supported the conclusions of
current research on the cross-sectional predictability of stock returns. Specifically, not finding
the relationship between expected return and beta in our data, we confirm the doubt placed on
the CAPM in the contemporary markets. In this research, size effect is not market wide but
rather found in small size stocks. This indicates that the effect is not due to market
inefficiency as confirmed in the contemporary markets. Therefore, we rule out size as a
predictor of the cross-section of expected return of securities. Total risk should not be
considered as a predictor of the cross-section of stock returns and limit to arbitrage in our
data. This is because the total risk does not show a consistent relationship with average
returns. Furthermore, returns from the dividend yield effect are not positively related to total
risk.
Our findings on momentum effect, dividend yield effect and the interactive effect of these two
on the 19
th
century BSE revealed the robustness of these characteristics as predictors of the
cross-section of stock returns across time. In view of this, we see momentum and value as the
main predictors of the cross-section of stock returns. Even though the environment in which
the historical market operated has changed so much in relation to the contemporary market,
value and momentum turn out to be the common characteristics that can predict returns.
Back to Chapter 1, we differentiate between characteristics that are consistent predictors of
returns and those that are due to data snooping or statistical artifacts. Specifically, in this
dissertation, we reveal that characteristic such as size, total risk, momentum and dividend
yield are not time series and cross-sectionally robust in predicting stock returns.

150

References
Altman, E. I., B. Jacquillat, and M Levasseur, 1974, Comparative analysis of risk measures: France and
the united states, Journal of Finance 1495-1511.
Amihud, Y., 2002, Illiquidity and stocks returns; cross-section and time series effect, Journal of
Financial Markets 5, 31-56.
Amihud, Yakov, and Haim Mendelson, 1986, Asset pricing and the bid-ask spread, Journal of Financial
Economics 17, 223-249.
Ang, Andrew, Robert J. Hodrick, Yuhang Xing, and Xiaoyan Zhang, 2006, The cross-section of volatility
and expected returns, Journal of Finance 61, 259-299.
Ang, Andrew, Robert J. Hodrick, Yuhang Xing, and Xiaoyan Zhang, 2009, High idiosyncratic volatility
and low returns: International and further u.S. Evidence, Journal of Financial Economics 91,
1-23.
Ang, Andrew, Jun Liu, and Krista Schwarz, 2008, Using individual stocks or portfolios in tests of factor
models, working Paper, Columbia University.
Annaert, J., Frans Buelens, and M. De Ceuster, 2004, Equity return estimation for 19th century
belgium, working paper ,University of Antwerp.
Asness, Clifford S., 1997, The interaction of value and momentum strategies, Financial Analysts
Journal 29-36.
Bairoch, P., 1982, International industrialisation levels from 1750 to 1980, Journal of European
Economic History 269-312.
Bali, Turan G., and Nusret Cakici, 2008, Idiosyncratic volatility and the cross section of expected
returns, Journal of Financial and Quantitative Analysis 43, 29-58.
Bali, Turan G., Nusret Cakici, Xuemin (Sterling) Yan, and Zhe Zhang, 2005, Does idiosyncratic risk
really matter?, Journal of Finance 60, 905-929.
Banz, Rolf W., 1981, The relation between return and market value of common stocks, Journal of
Financial Economics 9, 3-18.
Barberis, N., A. Shleifer, and R. Vishny, 1998, A model of investor sentiment, Journal of Financial
Bartholdy, J., and A. Riding, 1994, Thin trading and the estimation of betas: The efficacy of alternative
techniques, Journal of Financial Research 241-254.
Baskin, J.B., and P. J. Miranti, 1997, A history of corporate finance, Cambridge University Press.
Basu, Sanjoy, 1977, Investment performance of common stocks in relation to their price-earnings
ratios: A test of the efficient markets hypothesis, Journal of Finance 32, 291-311.
151

Basu, Sanjoy, 1983, The relationship between earnings yield, market value, and return for nyse
common stocks: Further evidence, Journal of Financial Economics 12, 129-156.
Beer, Francisca Marie, 1997, Estimation of risk on the brussels stock exchange: Methodological issues
and empirical results, Global Finance Journal 8, 83-94.
Bekaert, Geert, Campbell R. Harvey, and Christian T. Lundblad, 2007, Liquidity and expected returns:
Lessons from emerging markets, Review of Financial Studies 20, 1783-1831.
Black, F. , M. C. Jensen, and Myron Scholes, 1972, "The capital asset pricing model: Some empirical
tests in jensens(ed.), Studies in the theory of Capital Markets New York: Praeger.
Blitz, David C., and Pim Van Vliet, 2007, The volatility effect, Journal of Portfolio Management 34,
102-113.
Blume, Marshall E., 1971, On the assessment of risk, Journal of Finance 26, 1-10.
Blume, Marshall E., 1975, Betas and their regression tendencies, Journal of Finance 30, 785-795.
Bulkley, George, and Vivekanand Nawosah, 2009, Can the cross-sectional variation in expected
returns explain momentum?, Journal of Financial and Quantitative Analysis 44, 777-794.
Chabot, Benjamin, Eric Ghysels, and Ravi Jagannathan, 2009, Momentum cycles and limits to
arbitrage evidence from victorian england and post-depression us stock markets, (NBER).
Chan, K. C., and Nai-fu Chen, 1988, An unconditional asset-pricing test and the role of frim size as an
instrumental variable for risk, Journal of Finance 43, 309-325.
Chan, K. C., Nai-fu Chen, and D. A. Hsieh, 1985, An exploratory investigation of the firm size effect,
Journal of Financial Economics 14, 451-471.
Chan, L. K. C., and J. Lakonishok, 1993, Are the reports of beta's death premature?, Journal of
Portfolio Management 19, 51-62.
Chan, Louis K. C., and Josef Lakonishok, 1992, Robust measurement of beta risk, Journal of Financial
and Quantitative Analysis 27, 265-282.
Chatterjee, S., and W. Jacques, 1994, An outlier resistant approach to risk estimation Financial
Analysts Journal 69-75.
Cheffins, B. R., 2006, Dividend as a substitute for corporate law: The separation of ownership and
control, Washington and Lee Law Review 63, 1273-1338.
Cheffins, B.R., 2008, Corporate ownership and control: British business transformed, Oxford
Univerisity press.
Cohen, Kalman J., Gabriel A. Hawawini, Steven F. Maier, Robert A. Schwartz, and David K. Whitcomb,
1983, Estimating and adjusting for the intervalling-effect bias in beta, Management Science
29, 135-148.
Collins, D. W., J. Ledolter, and J. Rayburn, 1987, Some further evidence on the stochastic properties
of systematic risk, Journal of Business 60, 425-448.
152

Conrad, Jennifer, and Gautam Kaul, 1998, An anatomy of trading strategies, Review of Financial
Studies 11, 489-519.
Cooper, Michael J., Huseyin Gulen, and Michael J. Schill, 2008, Asset growth and the cross-section of
stock returns, Journal of Finance 63, 1609-1651.
Cooper, Michael J., Roberto C. Gutierrez, Jr., and Allaudeen Hameed, 2004, Market states and
momentum, Journal of Finance 59, 1345-1365.
Daniel, K., D. Hirshleifer, and Avanidhar Subrahmanyam, 1998, Investor psychology and security
market under- and overreactions, Journal of Finance 53, 1839-1886.
Dimson, E., and P. Marsh, 1999, Murphy's law and market anomalies, Journal of Portfolio
Management 25.
Dimson, E., and P. Marsh, R.,, 1983, The stability of uk risk measures the problem of thin trading,
Journal of Finance 754-784.
Dimson, Elroy, 1979, Risk measurement when shares are subject to infrequent trading, Journal of
Dimson, Elroy, Paul Marsh, and Mike Staunton, 2002. Triumph of the optimists. 101 years of global
investment returns (Princeton University Press).
Doukas, J. A., Chansog (Francis) Kim, and Pantzailis Christos, 2010, Arbitrage risk and stock mispricing,
Journal of Financial and Quantitative Analysis 45, 907-934.
Eisenbeiss, Maik, Gran Kauermann, and Willi Semmler, 2007, Estimating beta-coefficients of german
stock data: A non-parametric approach, European Journal of Finance 13, 503-522.
Elton, E. J., and M. J. Gruber, 1983, A simple examination of the emipirical relationship between
dividend yields and deviation from the capm, Journal of Banking and Finance 7, 135-146.
Faff, R. W., and H John, 1992, Time stationarity of systematic: Some australian evidence, Journal of
Business Finance and Accounting 19, 253-270.
Fama, E. F., and K. R. French, 1992, The cross-section of expected stock returns, Journal of Finance
47, 427-465.
Fama, E. F., and K. R. French, 1993, Common risk factors in the return on stocks and bonds, Journal of
Fama, Eugene F., and Kenneth R. French, 1996, Multifactor explanations of asset pricing anomalies,
Journal of Finance 51, 55-84.
Fama, Eugene F., and Kenneth R. French, 2008, Dissecting anomalies, Journal of Finance 63, 1653-
1678.
Fama, Eugene F., and James MacBeth, 1973, Risk, return, and equilibrium: Empirical tests, Journal of
Political Economy 81, 607-636.
153

Fowler, D. J., and C. H. Rorke, 1983, Risk measurement when shares are subjected to infrequent
trading, Journal of Financial Economics 280-283.
Fu, Fangjian, 2009, Idiosyncratic risk and the cross-section of expected stock returns, Journal of
Gervais, Simon, and Terrance Odean, 2001, Learning to be overconfident, Review of Financial Studies
14, 1-27.
Ghysels, E., 1998, On stable factor structures in the pricing of risk: Do time-varying betas help or
hurt?, Journal of Finance 53, 549-573.
Gibbons, Michael R., Stephen A. Ross, and Jay Shanken, 1989, A test of the efficiency of a given
portfolio, Econometrica 57, 1121-1152.
Goyal, Amit, and Pedro Santa-Clara, 2003, Idiosyncratic risk matters!, Journal of Finance 58, 975-
1007.
Gregory-Allen, Russell, C. Michael Impson, and Imre Karafiath, 1994, An empirical investigation of
beta stability: Portfolios vs. Individual securities, Journal of Business Finance and Accounting
21, 909-916.
Grossman, Richard S., and Stephen H. Shore, 2006, The cross section of stock returns before world
war i, Journal of Financial and Quantitative Analysis 41, 271-294.
Gwilym, Owain AP, Andre Clare, James Seaton, and Stephen Thomas, 2009, Dividends and
momentum, The Journal of investing 42-47.
Hanna, J. Douglas, and Mark J. Ready, 2005, Profitable predictability in the cross section of stock
returns, Journal of Financial Economics 78, 463-505.
Harvey, David, Stephen Leybourne, and Paul Newbold, 1997, Testing the equality of prediction mean
squared errors, International Journal of Forecasting 13, 281-291.
Haugen, R. A., and N. L. Baker, 1996, Commonality in the determinants of expected stock returns,
Hawawini, G., and P. Michel, 1979, An assessment of risk in the thinner market, belgium case, Journal
of Economics and Business
Hong, Harrison, Terence Lim, and Jeremy C. Stein, 2000, Bad news travels slowly: Size, analyst
coverage, and the profitability of momentum strategies, Journal of Finance 55, 265-295.
Hong, Harrison, and Jeremy C. Stein, 1999, A unified theory of underreaction, momentum trading,
and overreaction in asset markets, Journal of Finance 54, 2143-2184.
Horowitz, Joel L., Tim Loughran, and N.E. Savin, 2000, The disappearing size effect, Research in
Huber, P. J., 2004, Robust statistics, Hoboken, New Jersey, Wiley.
154

Ibbotson, R. G., P. D. Kaplan, and J. D. Peterson, 1997, Estimates of small-stock betas are much too
low, Journal of Portfolio Management 23, 104-111.
Jagannathan, R., and Z. Wang, 1996, The conditional capm and the cross-section of expected returns,
Jegadeesh, N., 1990, Evidence of predictable behavior of security returns, Journal of Finance 45, 881-
898.
Jegadeesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and selling losers:
Implications for stock market efficiency, Journal of Finance 48, 65-91.
Jegadeesh, Narasimhan, and Sheridan Titman, 2001, Profitability of momentum strategies: An
evaluation of alternative explanations, Journal of Finance 56, 699-720.
Jegadeesh, Narasimhan, and Sheridan Titman, 2002, Cross-sectional and time-series determinants of
momentum returns, Review of Financial Studies 15, 143-157.
Keim, D. B., 1985, Dividend yields and stock returns: Implications of abnormal january returns,
Klemkosky, R. C., and J. Martin, D., 1975, The adjustment of beta forecasts, Journal of Finance 1123-
1128.
Kothari, S. P., Jay Shanken, and Richard G. Sloan, 1995, Another look at the cross-section of expected
stock returns, Journal of Finance 50, 185-224.
Lakonishok, J., and Alan C. Shapiro, 1986, Systematic risk, total risk and size as determinants of stocks
market returns, Journal of Banking and Finance 10, 115-132.
Lakonishok, J., A. Shleifer, and R. W. Vishny, 1994, Contrarian investment, extrapolation, and risk,
Lee, Charles M.C., and Bhaskaran Swaminathan, 2000, Price momentum and trading volume, Journal
of Finance 55, 2017-2069.
Lesmond, David A., Michael J. Schill, and Chunsheng Zhou, 2004, The illusory nature of momentum
profits, Journal of Financial Economics 71, 349-380.
Levy, R., A., 1971, On the short term stationarity of beta coefficient, Financial Analysts Journal 55-62.
Lintner, John, 1965, Security prices, risk and maximal gains from diversification, Journal of Finance
20, 587-615.
Lintner, John, 1965, The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets, Review of Economics and Statistics 47, 13-37.
Lo, Andrew W., and A. Craig MacKinlay, 1990, Data-snooping biases in tests of financial asset pricing
models, Review of Financial Studies 3, 431-468.
Luoma, M., T. Martikainen, J. Perttunen, and S. Pynnnen, 1994, Different beta estimation
techniques in infrequently traded and inefficient stock markets, Omega 22, 471-476.
155

Malkiel, B. G., and Yexiao Xu, 2006, Idiosyncratic risk and security returns,, University of Texas at
Dallas.
Martin, R. Douglas, and Timothy T. Simin, 2003, Outlier-resistant estimates of beta, Financial Analysts
Journal 59, 56-69.
McLean, R. David, 2010, Idiosyncratic risk, long term reversal and momentum, Journal of
Quantitative and Financial Analysis 45, 883-906.
Merton, Robert C., 1987, A simple model of capital market equilibrium with incomplete information,
Morgan, G., and S. H. Thomas, 1998, Taxes,dividend yields and returns in the uk equity market,
Journal of Banking and Finance 405-423.
Moskowitz, Tobias J., and Mark Grinblatt, 1999, Do industries explain momentum?, Journal of
Finance 54, 1249-1290.
Mossin, Jan, 1966, Equilibrium in a capital asset market, Econometrica 34, 768-783.
Newey, Withney K., and Kenneth D. West, 1987, A simple, positive semi-definite, heteroskedasticity
and autocorrelation consistent covariance matrix, Econometrica 55, 703-708.
Neymarck, A., 1911. Les finances contemporaines. L'epargne franaise et les valeurs mobilires 1872-
1910 (Felix Alcan, Paris).
Nijman, Theo, Laurens Swinkels, and Marno Verbeek, 2004, Do countries or industries explain
momentum in europe?, Journal of Empirical Finance 11, 461-481.
Reinganum, M. R., 1981, A new empirical perspective on the capm, Journal of Financial and
Quantitative Analysis 439-462.
Reinganum, M. R., 1983, Portfolio strategies based on market capitalization, Journal of Portfolio
Management 9, 29-36.
Reinganum, Marc R., 1981, Misspecification of capital asset pricing: Empirical anomalies based on
earnings' yields and market values, Journal of Financial Economics 9, 19-46.
Ritter, J. R., and Navin Chopra, 1989, Portfolio rebalancing and the turn-of-the-year effect, Journal of
Finance 44, 149-166.
Rosenberg, Barr, Kenneth Reid, and Ronald Lanstein, 1985, Persuasive evidence of market
inefficiency, Journal of Portfolio Management 11, 9-17.
Rouwenhorst, K. Geert, 1998, International momentum strategies, Journal of Finance 53, 267-284.
Rouwenhorst, K. Geert, 1999, Local return factors and turnover in emerging stock markets, Journal of
Finance 54, 1439-1464.
Scholes, Myron, and Joseph Williams, 1977, Estimating betas from nonsynchronous data, Journal of
156

Scholliers, P., 1997, Industrial wage differentials in nineteenth-century belgium, in H. Van der Wee,
and J. Blomme, eds.: The economic development of belgium since 1870 (Edward Elgar,
Cheltenham).
Schwert, G. William, 2003, Anomalies and market efficiency, in G. M. Constantinides, Milton Harris,
and Rene Stulz, eds.: Handbook of the economics of finance (North-Holland, Amsterdam).
Shanken, Jay, 1992, On the estimation of beta-pricing models, Review of Financial Studies 5, 1-33.
Sharpe, W. F., and G. M. Cooper, 1972, Risk- return class of new york stock exchange, Financial
Analysts Journal 28, 46-52.
Sharpe, William F., 1964, Capital asset prices: A theory of market equilibrium under conditions of risk,
Journal of Finance.
Subrahmanyam, Avanidhar, 2010, The cross-section of expected stock returns: What have we learnt
from the past twenty-five years of research?, European Financial Management 16, 27-42.
Van der Wee, Herman, 1996, The industrial revolution in belgium, in M. Teich, and Roy Porter, eds.:
The industrial revolution in national context. Europe and the USA (Cambridge University
Press, Cambridge).
Van Nieuwerburgh, Stijn, Frans Buelens, and Ludo Cuyvers, 2006, Stock market development and
economic growth in belgium, Explorations in Economic History 43, 13-38.
Vasicek, Oldrich, 1973, A note on using cross-sectional information in bayesian estimation of security
betas, Journal of Finance 28, 1233-1239.

157

NEDERLANDSTALIGE SAMENVATTING
In dit proefschrift bestuderen we cross-sectionele patronen in aandelenrendementen, gebruik
makende van een volledig onafhankelijke databank gebaseerd op gegevens van de Brusselse
Beurs (BSE: Brussels Stock Exchange) tijdens de 19
de
eeuw en de eerste jaren van de 20
ste

eeuw. Deze dataset laat ons toe de invloed te bestuderen op aandelenrendementen van sterk
varirende omstandigheden in de economische en institutionele omgeving. De tijdsperiode die
deze dataset bestijht, vermijdt mogelijke kritiek op datamining. Met deze data:
- Testen we de validiteit van het CAPM.
- Daarenboven testen we of andere kenmerken een verklaring kunnen bieden voor de
cross-sectionele variatie in aandelenrendementen.
Op het moment van het schrijven van dit proefschrift was de boekhoudkundige en
transactiedata van de BSE nog niet gedigitaliseerd voor de 19
de
eeuw. Daarom was het niet
mogelijk de verklaringskracht van boekhoudkundige gerelateerde kenmerken voor
aandelenrendementen te onderzoeken. Daarentegen onderzoeken we of grootte
(marktkapitalisatie), momentum (voorbije korte termijn rendementen), totaal risico
(bedrijfsspecifiek risico) en dividendrendement (als indicator voor de waarde van de activa,
net zoals de book-to-market ratio) de aandelenrendementen cross-sectioneel kunnen
voorspellen.
Er werden vier empirische studies uitgevoerd aan de hand van de BSE data uit de 19
de
eeuw
en het begin van de 20
ste
eeuw. In studie 1 behandelen we de beoordeling van de bta. In
studie 2 testen we de validiteit van het CAPM en het grootte-effect. De derde studie test het
momentum-effect. De gecombineerde impact van grootte, momentum, totaal risico en
dividendrendement op de cross-sectionele aandelenrendementen wordt behandeld in studie 5.
158

In de eerste studie focussen we op de beoordeling van de bta. Bta is de belangrijkste
inputvariabele van het CAPM. Het is een geschatte variabele die mogelijk gemeten wordt met
een bepaalde statistische fout, waardoor de testresultaten van het CAPM vertekend kunnen
zijn. De instabiliteit, vertekening en niet-robuustheid naar outliers toe van bta zijn dan ook
belangrijke onderzoekstopics geworden sinds de ontwikkeling van het CAPM. In deze studie
onderzoeken we de relatieve prestatie van de verschillende methodes om de bta te schatten,
gebaseerd op hun vermogen om de erop volgende bta te voorspellen. Meer specifiek
vergelijken we de btas van het marktmodel met btas die geschat werden door de auto-
regressieve technieken van Blume (1971) en Vasicek (1973).
We tonen aan dat de individuele btas van het aandelenmarktmodel niet stabiel zijn. De
voorspelbaarheid van deze btas kan worden verbeterd door het vormen van portefenilles
met ten minste tien of meer aandelen. Opvallend genoeg zijn er geen significante verschillen
in de verklarende nauwkeurigheid tussen de Blume en Vasicek aangepaste btas. Wanneer
er gebruik gemaakt wordt van de Dimson-methode om de btas te schatten, blijken een klein
aantal aandelen rendementen te hebben die een voorsprong (lead) of achterstand (lag) habben
op de marktrendementen. Om rekening te kunnen houden met outliers werden iterative
reweighted least square (IRLS) technichen gebruikt om de btas te schatten. De btas van
de IRLS zijn klein qua grootte in vergelijking met de btas van het marktmodel, maar hebben
wel dezelfde verklarende nauwkeurigheid als het marktmodel.
In de derde studie werd er gebruik gemaakt van de sorteermethode en de Fama en MacBeth
(1973) (FM) cross-sectionele regressiemethode om te kunnen onderzoeken of het CAPM
geldig is op data van vr de Eerste Wereldoorlog. Er werd eveneens getest of het grootte-
effect (de neiging van kleine aandelen om hogere rendementen te hebben dan grote aandelen)
reeds in de 19
de
eeuw aanwezig was op de BSE.
159

De sorteermethode en de FM cross-sectionele regressiemethode leveren geen empirische
bewijskracht voor het CAPM. We tonen eveneens aan dat de relatie tussen de bta en de
rendementen varieert doorheen de tijd. De nulhypothese die de gelijkheid van de geschatte
hellingen veronderstelt van de cross-sectionele regressie en de abnormale marktrendementen
wordt niet verworpen in de periode 1868-1893. Daarentegen wordt door gebruik te maken van
de op grootte gesorteerde portfolios (met gelijk gewicht) in de cross-sectionele regressie, een
verband vastgesteld tussen bta en abnormaal rendement met bta als enige verklarende
variabele (het CAPM is geldig). De voorwaardelijke dubbele sorteermethode die toegepast
wordt door Fama en French (1992) scheidt echter het effect van bta en grootte op de
verwachte rendementen. Deze methode maakt de gemiddelde cross-sectionele helling van de
bta niet significant, zowel wanneer ze met als zonder de grootte in de regressie wordt
geplaatst.
We vinden dat grootte negatief gerelateerd is aan abnormale rendementen. Er bestaat een
significant negatief verband met grootte (grootte-effect), maar de bta is niet gerelateerd aan
abnormale rendementen wanneer beide variabelen worden opgenomen in de cross-sectionele
regressies. Een gedetailleerde analyse van de data onthult dat het grootte-effect voornamelijk
wordt veroorzaakt door de kleinere aandelen, die meetellen voor slechts 0.35% van de totale
marktkapitalisatie. Wanneer deze kleine aandelen buiten beschouwing worden gelaten wordt
er geen relatie gevonden tussen abnormale rendementen enerzijds en btas of grootte
anderzijds. Zowel de sorteermethode als de cross-sectionele regressie onthult dat het grootte-
effect verdwijnt wanneer de aandelen op waarde worden gewogen bij het vormen van de
portefenilles. Samenvattend wil dit zeggen dat het CAPM niet geldig is voor de BSE tijdens
de 19
de
eeuw. Door het schatten van de btas met het marktmodel en de Dimson en Vasicek
methodes zal het model niet tot stand gebracht worden. Het grootte-effect bestaat, maar is
160

voornamelijk toe te schrijven aan een kleine groep van aandelen die slechts een klein
percentage van de totale marktkapitalisatie vertegenwoordigen.
In de vierde studie onderzoeken we of een momentum-strategie een abnormale winst kan
realiseren op de 19
de
-eeuwse BSE. De momentum-strategie houdt in dat aandelen worden
gekocht (verkocht) die sterk (zwak) gepresteerd hebben in de laatste 3 tot 12 maanden. Er
bestaat overtuigend bewijs dat deze strategie winstgevend is op de 19
de
-eeuwse BSE. Het
vinden van een momentum-effect in deze periode bevestigt de bewering dat momentum-
winsten gevonden op markten in de periode na de Tweede Wereldoorlog, niet enkel toe te
schrijven zijn aan vertekeningen door data-snooping. Een gedetailleerde analyse onthult dat
het momentum-effect niet bestaat in de groep van kleine aandelen. Een bijkomende analyse
met betrekking tot de momentum-winsten in elke kalendermaand toont aan dat de winst
positief was voor elk van deze maanden. Een omkening in januari (January reversal effect)
dat gevonden wordt op naoorlogse Amerikaanse markten kan niet gevonden worden op de
19
de
-eeuwse BSE. In feite laat de maand januari zelfs de vierde hoogste momentum-winsten
optekenen in vergelijking met de andere maanden van het jaar. We vinden dat de momentum-
winsten niet sterk zijn in de eerste twintig jaar die we bestudeerden.
Teneinde de oorzaak na te gaan van de momentum-winsten hebben we gebruik gemaakt van
de Jegadeesh en Titman (2001)-benadering om de rendementen te bestuderen van de
momentum-portrfenilles in de post-holdingperiode. De momentum-rendementen keren om in
het tweede tot vijfde jaar na de holdingperiode. Verder onderzoek toont aan dat deze
omkering voornamelijk veroorzaakt wordt door kleine aandelen. We hebben eveneens getest
of de momentum-winsten en de lange termijn ommekeer in de cross-sectie van
aandelenrendementen afhankelijk is van de markttoestand. We vonden dat de 6-maanden
formatiestrategie en de 6- tot 12-maanden holdingperiodestrategie enkel winstgevend waren
in perioden van marktwinsten.
161

In de vijfde studie onderzoeken we of grootte, momentum, total risico en dividendrendement
de cross-sectie van aandelenrendementen kunnen verklaren in de periode 1868-1913. Grootte
en momentum worden opnieuw in de analyse opgenomen om de standvastigheid en de
gecombineerde impact van deze kenmerken te testen op de cross-sectie van abnormale
rendementen. Het sorteren op grootte test eveneens of het effect van de andere
karakteristieken niet enkel van toepassing is op de groep van kleine en illiquide aandelen. Het
totaal risico voor elk aandeel wordt gemeten als de standaardafwijking van de residuen van
het Dimson-model voor het schatten van de btas van de aandelen. In elk jaar is het
dividendrendement van een aandeel gelijk aan de som van alle dividenden betaald tijdens de
laatste 12 maanden gedeeld door de eindejaarsprijs. We onderzochten de standvastigheid van
de relaties over alle dividendrendement-, grootte-, totaal risico en momentum-groepen heen.
De sorteermethode en FM cross-sectionele regressie methodes werden toegepast. We
bevestigen dat grootte een significant negatief verband heeft met abnormale rendementen.
Echter, wanneer deze analyse op groepen van verschillende grootte wordt uitgevoerd, komt
aan het licht dat dit negatieve verband volledig is toe te schrijven aan zeer kleine aandelen die
meetellen voor slechts 3.67% van de totale marktkapitalisatie. Dit bevestigt de bevinding dat
het grootte-effect voornamelijk toe te schrijven is aan de eerste deciel grootte-portefenille,
waarvan deze zeer kleine aandelen deel uit maken. We hebben geen consistente verbanden
gevonden voor totaal risico. Momentum toont een consistent positief verband met abnormale
aandelenrendementen, die meetellen voor 96% van de totale marktkapitalisatie. Verder
vonden we een negatief verband tussen dividendrendement en momentum bij dividend-
betalende aandelen, maar elk van hen is positief gerelateerd aan gemiddelde abnormale
rendementen. Dividendrendement en momentum zijn positief gerelateerd aan gemiddelde
abnormale rendementen bij onze grote aandelen, die meetellen voor ongeveer 96% van de
totale marktkapitalisatie.
162

Hoewel elke conclusie over historische periodes en verschillende systemen heen met enige
voorzichtigheid genterpreteerd moet worden, bevestigen de gelijkenissen met de 19
de
-eeuwse
BSE de conclusies van actueel onderzoek naar cross-sectionele voorspellingen van
aandelenrendementen. Meer specifiek bevestigen we, door het niet vinden van een verband
tussen verwachte rendementen en de bta in onze data, de twijfel die rust over het CAPM in
de hedendaagse markten. In dit onderzoek is het grootte-effect niet van toepassing op de
gehele markt, maar eerder op de kleine aandelen. Dit toont aan dat het effect niet te wijten is
aan marktinefficintie zoals bevestigd wordt in hedendaagse markten. Daarom sluiten we
grootte uit als een voorspeller van de cross-sectie van verwachte rendementen van effecten.
Totaal risico mag niet beschouwd worden als een voorspeller van de cross-sectie van
aandelenrendementen en is slechts beperkt tot een arbitrage rol in onze data. Dit is omdat
totaal risico geen consistent verband toont met de gemiddelde rendementen. Bovendien zijn
rendementen van het dividendrendement-effect niet positief gerelateerd aan totaal risico.
Onze bevindingen in verband met het momentum-effect, dividendrendement-effect en het
interactie-effect van beide op de 19
de
-eeuwe BSE tonen de robuustheid van deze kenmerken
aan als voorspellers van de cross-sectie van aandelenrendementen doorheen de tijd. Afgaande
hierop beschouwen we momentum en waarde als de belangrijkste voorspellers van de cross-
sectie van aandelenrendementen. Ook al veranderde de omgeving waarin deze historische
markten actief waren enorm in vergelijk met de huidige marktomstandigheden, waarde en
momentum blijken de gemeenschappelijke kenmerken te zijn die rendementen kunnen
voorspellen.
We maken dus een onderscheid tussen karakteristieken die consistente voorspellers zijn voor
rendementen en deze die toe te schrijven zijn aan data-snooping of statistische artefacten. In
het bijzonder onthullen we in dit proefschrift dat karakteristieken zoals grootte, totaal risico,
163

momentum en dividendrendement niet robuust zijn in tijdzeeksen en cross-secties in de
voorspelling van dividendrendementen.

Cross-Sectional Predictability of Stock Returns: Pre-World War I Evidence

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Cross-Sectional Predictability of Stock Returns: Pre-World War I Evidence

Enviado por

Direitos autorais:

Formatos disponíveis

FACULTY OF APPLIED ECONOMICS

DEPARTMENT OF ACCOUNTING AND FINANCE

= + + for j=1,,N (5)

. The error term

Você também pode gostar