Você está na página 1de 146

Modelling, Forecasting and Trading Power Futures Spread

PAUL AMENYAWU

This dissertation is submitted in partial fulfilment of the requirement for the degree

of Master of Science in International Banking and Finance in the Isle of Man

International School, Liverpool John Moore’s University.

September 2007

1
Declaration

“This is to certify that this dissertation is the result of an original investigation. The

material has not been used in the submission of any other qualification. Full

acknowledgement has been given to all sources used”

Signed: Paul Amenyawu …………………………………..

Signed: Dr Andreas Nabor ………………………………………

2
Modelling, Forecasting and Trading Power Futures Spread

By PAUL AMENYAWU

Dissertation Supervisor: Dr Andreas Nabor

3
Acknowledgements

I would like to show an appreciation to my supervisor Dr. Andreas Nabor for his

suggestion, ideas and feed back. His guidance helped me a lot.

The content of this dissertation was inspired from articles, journal and books; I

am indebted to the authors and wish to express my appreciation to all of them.

In addition I would like to thank Nord Pool for granting me the permission to use

their data, which made my work possible.

Finally, I would like to thank all IBS staff in general and Melanie Jones in

particular; she was always there for us and for providing me with most of the

articles and books that I needed; especially for the literature review. She is great!

4
Table of Content

MODELLING, FORECASTING AND TRADING POWER FUTURES SPREAD .................. 1

PAUL AMENYAWU .................................................................................................................... 1

DECLARATION............................................................................................................................... 2

MODELLING, FORECASTING AND TRADING POWER FUTURES SPREAD .................. 3

Acknowledgements .................................................................................................................... 4

ABSTRACT..................................................................................................................................... 12

CHAPTER 1 INTRODUCTION .............................................................................................. 13

CHAPTER 2 LITERATURE REVIEW ................................................................................. 16

EUROPEAN ELECTRICITY MARKETS: AN OVERVIEW...................................................................... 25


ECONOMIC FACTORS ..................................................................................................................... 26
WEATHER AND TEMPERATURE CONDITION ................................................................................... 26
COMMODITY P RICES ..................................................................................................................... 27
THE FINANCIAL MARKETS ............................................................................................................ 27
OPTION CONTRACTS ..................................................................................................................... 28
CONTRACTS FOR D IFFERENCE (CFD) ........................................................................................... 28
GROWTH ....................................................................................................................................... 28
S&P500 INDEX ............................................................................................................................. 29

TECHNICAL ANALYSIS: AN OVERVIEW.............................................................................. 30

CHAPTER 3 METHODOLOGY & METHODS ................................................................... 32

B OX JENKINS OR ARIMA METHODOLOGY.................................................................................... 32


ARMA (P , Q) AND ARIMA (P , D, P) MODELS ................................................................................ 32
ARIMA ( P, D, Q)........................................................................................................................... 33
VECTOR ERROR CORRECTION MODEL (VECM) & VECTOR A UTO REGRESSION (VAR)
METHODOLOGY ............................................................................................................................. 34

GENERAL VECEM (P) .................................................................................................................. 35


COINTEGRATION METHOD ............................................................................................................. 35
CO- INTEGRATION “FAIR VALUE” APPROACH-P AIRS TRADING MODELLING ................................. 36
VOLATILITY MODEL METHODOLOGY ............................................................................................ 38
INTRODUCTION ............................................................................................................................. 38
GARCH (1, 1) Model............................................................................................................... 38

5
TARCH Model with seasonal dummies............................................................................................. 39
SUMMARY ..................................................................................................................................... 40
SPREAD SERIES ............................................................................................................................. 40
DATA D ESCRIPTION ...................................................................................................................... 42
MEAN R EVERTING MODELS WITH JUMPS ...................................................................................... 43
P ARAMETER ESTIMATION .............................................................................................................. 46
MEAN R EVERTING MODELS WITH TIME VARYING MEAN ............................................................... 46

SIMULATION OF ELECTRICITY PRICES RETURNS.......................................................... 47

INTRODUCTION ............................................................................................................................. 47
MODELLING .................................................................................................................................. 49
MEAN R EVERSION PARAMETER ESTIMATION ................................................................................ 51
S IMULATION OF MEAN R EVERSION WITH JUMPS ........................................................................... 52
S INGLE F ACTOR MODEL WITH S EASONALITY ........................................................................... 54

S IMPLE MOVING AVERAGE (SMA)............................................................................................... 55


TRADING R ULES: .......................................................................................................................... 55
EXPONENTIAL SMOOTHED MOVING AVERAGE ............................................................................. 55
DOUBLE EXPONENTIAL MOVING AVERAGE (DEMA)................................................................... 56
MOVING AVERAGE C ONVERGENCE – DIVERGENCE (MACD)....................................................... 57
TRADING R ULES WITH MACD...................................................................................................... 58
B OLLINGER B ANDS (BB) .............................................................................................................. 58
TRADING R ULES: .......................................................................................................................... 59
RELATIVE S TRENGTH INDICATOR (RSI)........................................................................................ 59
ADX DIRECTIONAL MOVEMENT INDEX ....................................................................................... 59
ADX ALGORITHM......................................................................................................................... 60
TRADING RULES ............................................................................................................................ 61
MOVING AVERAGE METHODOLOGY .............................................................................................. 61
TRADING R ULES FOR THE EMA AND MA: .................................................................................... 62

REVIEW OF PERFORMANCE MEASURES............................................................................ 63

INTRODUCTION ............................................................................................................................. 63
FORECASTING MEASURES ............................................................................................................. 63

ROOT MEAN SQUARE ERROR (RMSE).................................................................................. 63

P ROFIT MEASURES ........................................................................................................................ 64


R ISK MEASURES ........................................................................................................................... 65
THE CORRELATION FILTER ............................................................................................................ 66
TRADING R ULES: .......................................................................................................................... 67
EQUITY CURVE ............................................................................................................................. 67

UNDER WATER CURVE ............................................................................................................. 68

6
SUMMARY ..................................................................................................................................... 68

CHAPTER 4 FINDINGS .......................................................................................................... 70

INTRODUCTION ............................................................................................................................. 70

RESULTS OF FORECASTING PERFORMANCE ................................................................... 70

A SUMMARY OF MODELLING PROCESS OF ARMA TIME SERIES ..................................................... 72


RESULT OF THE TRADING STRATEGY USING S&P500 .................................................................... 72
SYSTEM PRICE AND MACD .......................................................................................................... 76

CHAPTER 5 CONCLUSION................................................................................................... 78

REFERENCES................................................................................................................................ 82

APPENDIX 1................................................................................................................................... 86

COINTEGRATION JOHANSEN METHODOLOGY ................................................................................ 86

APPENDIX 2................................................................................................................................... 88

APPENDIX 3................................................................................................................................. 101

FORECASTING WITH ORDINARY LEAST SQUARE AS APPLIED TO POWER DATA


(CDF) ............................................................................................................................................. 105

APPENDIX 4................................................................................................................................. 109

FORECASTING WITH ARIMA MODEL .......................................................................................... 111


VECTOR E RROR C ORRECTION MODEL (VECM) ......................................................................... 120

FIGURE 47: MODELLING OF S&P500 SERIES.................................................................... 134

7
List of Tables

TABLE 1: SOURCE: HTTP://WWW . FUTURESANDOPTIONSTRADER. COM ............................................... 57


TABLE 2: A COMPARISON OF FORECASTING TOOLS ......................................................................... 71
TABLE 3: SUMMARY OF RATIOS OF THE TRADING STRATEGY -OLSQ APPLIED TO CDF...................... 71
TABLE 4: SUMMARY OF T RADING S TRATEGY ANALYSIS OF S&P500 INDEX ( WITHOUT FORECASTING
TOOL) ....................................................................................................................................... 73

TABLE 5: SUMMARY OF RATIOS OF THE RESULT OF S&P500 TRADING STRATEGY WITH ARMA ....... 74
TABLE 6: SUMMARY RATIOS OF PAIRS TRADING STRATEGY ............................................................... 75
TABLE 7: SUMMARY OF THE RATIOS OF MACD TRADING STRATEGY APPLIED TO YEARLY SYSTEM
INDEX ....................................................................................................................................... 76

TABLE 8: SUMMARY OF THE RATIOS OF MACD TRADING STRATEGY APPLIED TO YEARLY S&P500
INDEX ....................................................................................................................................... 77

TABLE 9: SUMMARY RATIOS OF THE TRADING STRATEGY - AS APPLIED TO THREE DIFFERENT DATA
SET. .......................................................................................................................................... 78

TABLE 10: A SUMMARY OF ADF TEST ............................................................................................. 90


TABLE 11: A SUMMARY OF PHILIPS -PERON TEST ........................................................................... 92
TABLE 12: A CORRELOGRAM OF CDF SERIES ................................................................................. 95
TABLE 13: FORECASTING WITH OLSQ........................................................................................... 109
TABLE 14: UNIT ROOT TEST ............................................................................................................ 112
TABLE 15: SUMMARY OF ESTIMATION OF ARMA (10, 10) .............................................................. 114
TABLE 16: GARCH-M MODEL ........................................................................................................ 116
TABLE 17 RESULT FROM COINTEGRATION TEST USING JOHANSEN COINTEGRATION TEST ......... 120
TABLE 18: RESULTS FROM THE UNIT ROOT TEST .......................................................................... 123
TABLE 19: OUTPUT OF THE ESTIMATION OF THE SHORT-RUN DYNAMICS ...................................... 126
TABLE 20: SUMMARY OF C O-INTEGRATION T EST............................................................................ 131
TABLE 21: OUTPUT OF THE UNIT ROOT TEST FOR STATIONARITY..................................................... 135
TABLE 22: OUTPUT OF AUTOCORRELATION TEST ......................................................................... 137
TABLE 23: OUTPUT OF AR AND MA TEST ...................................................................................... 138
TABLE 24: FORECAST OUTPUT (STATIC FORECAST -USING THE ARMA) ......................................... 140
TABLE 25: LIST OF R ATIOS .............................................................................................................. 141

8
List of Figures

FIGURE 1: A COMBINATION OF 30-DAY VOLATILITY AND SKEW OF RETURNS OF THE

SYSTEM PRICE ...................................................................................................... 49

FIGURE 2: A COMPARISON OF A SIMULATE PATH OF A MEAN REVERSION MODEL PATH

AND THE ORIGINAL PATH. .................................................................................... 52

FIGURE 3 JUMP DIFFUSION MODEL ............................................................................... 53

FIGURE 4 A MOVING AVERAGE STRATEGY ................................................................. 62

FIGURE 5 AN EXAMPLE OF A SUMMARY OF THE RESULTS OF A TYPICAL TRADING

STRATEGY. SOURCE : F UTURES & OPTIONS MAGAZINE ........................................ 66

FIGURE 6: EQUITY CURVE FOR A TRADING STRATEGY SHOWING FEW PAUSES. ............ 67

FIGURE 7: THE DIAGRAM SHOWS THE DRAWDOWN OF THE OUTPUT OF A TRADING

STRATEGY. ........................................................................................................... 68

FIGURE 8: SYSTEM PRICE DETERMINATIONS ................................................................ 29

FIGURE 9: A DAILY DATA OF RANGE PRICE SERIES OR THE CDF .................................. 88

FIGURE 10: A DATA COMPARISON OF SYSTEM AND HELSINKI DAILY PRICES ............... 89

FIGURE 11: A SUMMARY STATISTICS OF THE CDF ....................................................... 89

FIGURE 12: A LOG TRANSFORMATION OF THE RANGE SERIES....................................... 90

FIGURE 13: A SUMMARY STATISTICS OF LOG TRANSFORMATION OF THE CDF SERIES.. 90

FIGURE 14: QUANTILE-QUANTILE PLOT OF THE LOG-RETURNS .................................... 99

FIGURE 15: A LOG TRANSFORMATION OF HELSINKI PRICES ........................................ 99

FIGURE 16: LOG TRANSFORMATION OF SYSTEM PRICES ............................................... 99

FIGURE 17: S&P500 INDEX........................................................................................ 100

FIGURE 18: LOGARITHMS RETURNS OF THE SYSTEM PRICE ....................................... 100

FIGURE 19: A PLOT OF NUMBER OF JUMPS AGAINST RETURNS ................................... 101

9
FIGURE 20: A COMBINATIONS OF 30-DAY VOLATILITY AND LOGARITHMS RETURNS

OF THE S YSTEM PRICE ........................................................................................ 101

FIGURE 21: MAXIMUM DRAWDOWN OF S&P500 WITHOUT FORECASTING ................ 102

FIGURE 22: RETURNS AND 9-DAY EMA AND 20-DAY MA SERIES ........................... 102

FIGURE 23: MAXIMUM DRAWDOWN FOR S&P500 WITH ARMA FORECAST .............. 103

FIGURE 24: DAILY PROFIT FOR S&P500 WITH ARMA FORECAST ............................. 103

FIGURE 25: DAILY CUMULATIVE PROFIT FOR S&P500 WITH ARMA FORECAST ........ 104

FIGURE 26: COMPARING PRICE, MOVING AVERAGE, EXPONENTIAL MOVING AVERAGE104

FIGURE 27: EQUITY CURVE OF S&P500 WITHOUT FORECASTING ............................... 105

FIGURE 28: GRAPH OF MAXIMUM DRAWDOWN OF OLSQ APPLIED TO CDF ............. 106

FIGURE 29: GRAPH OF EQUITY CURVE ...................................................................... 106

FIGURE 30: 30-DAY VOLATILITY ............................................................................... 107

FIGURE 31: GRAPH OF MAXIMUM DRAWDOWN .......................................................... 107

FIGURE 32: GRAPH OF CUMULATIVE PROFIT .............................................................. 108

FIGURE 33: ERROR SQUARE ...................................................................................... 109

FIGURE 34: RESULT OF FORECAST USING OLSQ ........................................................ 111

FIGURE 35:................................................................................................................. 112

FIGURE 36: A SUMMARY OF FORECASTING RESULT OF BOX-JENKINS METHOD .......... 116

FIGURE 37: A SUMMARY OF RESULT OF A DYNAMIC GARCH-M FORECAST .............. 119

FIGURE 38: A SUMMARY OF RESULT OF A STATIC GARCH-M FORECAST .................. 119

FIGURE 39: ERROR CORRECTION TERM SERIES ....................................................... 123

FIGURE 40: RESULTS OF DYNAMIC FORECAST OF VECM MODEL ............................ 128

FIGURE 41: RESULTS OF STATIC FORECAST OF VECM MODEL ............................... 128

FIGURE 42: A COMPARISONS OF A SYSTEM PRICE AND SIMULATED PRICE WITH A

MEAN SPEED OF 0.47 AND A LONG RUN MEAN OF 949.27 WITH A HALF-LIFE 1.46129

FIGURE 43: ESTIMATING PARAMETER OF A MEAN REVERSION MODEL USING LEAST

10
SQUARE ESTIMATION......................................................................................... 129

FIGURE 44: SUMMARY STATISTICS OF RETURNS SERIES OF GARCH (1, 1)

ESTIMATION OF THE SYSTEM PRICE .................................................................... 130

FIGURE 45: VARIANCE OF THE SERIES RETURNS ........................................................ 131

FIGURE 46:................................................................................................................. 134

FIGURE 47: THE LOG SERIES OF THE S&P INDEX ........................................................ 135

FIGURE 48: OUTPUT FORECAST (DYNAMIC FORECAST)........................................... 140

11
Abstract

The main aim of this dissertation is to forecast short-term electricity price spread

and to apply a correlation filter to the forecasted returns with the aim of finding

initially the best forecasting tool. A MACD trading strategy is then applied to the

forecasted returns of the best tool, profitability and volatility ratios are then

calculated using Microsoft excel spread sheet set up. This set up consists of an

input data series-which can be varied with set of ratios generating different values.

The is an adaptation of Jason al et, (2002) spread sheet models. The applicability of

the models to trading is thus evaluated and conclusions drawn based on the values

of the various ratios.

It is concluded that forecasting tools can be used by traders in their daily trading

activities. Similarly market inefficacies do exist and trading strategies can generate

modest or reasonable amount of profits and losses from these markets.

12
Chapter 1 Introduction

The main objective of the dissertation is to apply academic forecasting tools to

stock market trading. Williams (2003) investigated the above and come out with

the following conclusion: “Regression model…, do have the ability to forecast

EUR/USD returns…, and to add value as a forecasting and quantitative tool “. The

dissertation set out to verify this, but with different data, with different

characteristics and from different markets.

The aim of the topic is to test the usefulness of academic forecasting tools, such as

ARMA to stock market traders. Market traders use two main tools: Fundamental

and Technical analysis. Fundamental analysis is based on their views of the market;

the trading decisions they make are based on for example; on the financial

statement of a particular company or a group of companies. Another example

includes US Federal monetary policies, especially with regard to interest rates.

Whilst technical analysis on the other hand uses day to day price movements as

their means of forecasting price trends- which can be upwards of downwards.

Some of their tools include moving average, range trading- trading between low

price and high price; other indicators are used to reach some of these decisions.

Two different data sources will be used. These are Nord Pool electricity and US

S&P500 index prices. These markets are entirely different, one is power based,

which exhibits peculiar characteristics such as: seasonal cycles, price spikes and

high volatility, whilst the other mainly based on equities of 500 companies from

wide range of backgrounds. In this regard S&P500 will be used as proxy for the

perfect market. Firstly each of these will be forecasted, with the aim of using the

ratios to compare the forecasting ability of these tools. Secondly, a moving average

13
trading strategy will applied using the out-of-sample data to measure the risk and

profit of the forecasting tools. Finally the moving average strategy will be applied

directly to the prices of both markets as a means of testing whether traders can

profit without the use of the academic forecasting tools. An Excel spread sheet will

be set up to do the moving average strategy calculations – which include ratios,

graphs and charts.

400 prices from 01-January-2002 to 31-December-2002 will be used. 250 of this

will be used in-sample for estimation of the various models; the rest will be used as

out-of-sample data mainly for forecasting and testing moving average strategy.

Four main forecasting tools will be use to forecast the Nord Pool and S&P 500

prices. The in-sample data will be used for estimation and optimization purposes,

whilst the out of sample data will be used for the forecasting. The various ratios

will be rank and compared with the aim of finding the best forecasting methods.

Secondly a moving average trading strategy will be tested using the forecasted data

values, with the various risk and profitability ratios computed. The result of the

strategy will be then compared to find out which of the forecasted tools provided is

the best. Dissertation is divided into five chapters. Chapter 2, Literature Review

examines previous articles, books and other works that have been studies by

several researches. Their approach and methodologies will examine with the aim of

identifying strength or weakness in each of them and if possible extend and

improve on them.

Chapter 3: “Methodology & Method”-describes the various research

methodologies, tools and ratios that are going to be used in the dissertation. Some

of these includes, forecasting tools, technical analysis ratios. Simulation of

electricity prices will be examined with the aim of identifying, prices hikes,

14
volatility and any other characteristics. A review of the Nord Pool as well as

S&P500 markets will also be examined. Data analysis and dynamic of each of the

price data will be look at. Finally, in addition to comparing forecasting techniques

as applied to electricity, the relationship between Spot/System price, Futures,

Forwards and CDF of Nord Pool will be examine using simulation techniques and

looking for evidence of unique characteristics. Chapter 4: “Findings” – discuss on

the basis of data analysis the usefulness and limitation of “academic” forecasting

tools to market participants. A general summary of the output of the results which

includes description of statistics of the various test will be in this section.

Chapter 5: “Conclusion”-This section provides a summary of the observation of

price developments, estimation, procedures and the relevance of the forecasting

tools and their usefulness to market traders in the two markets.

15
Chapter 2 Literature Review

Introduction

Futures sport price exhibit features such as: mean reversion, seasonal effects,

time varying volatility and volatility clustering, extreme values/spikes, time of the

day effects i.e. price variation during the day, week days, weekends and the

beginning of the week. Couple with the fact the electricity is non-storable in other

words cannot be use to arbitrage price. However, Clewlow al et, (2000), argued

that inputs such as coal, gas, nuclear fuels, are in fact storable, therefore cost

associated with transmission, purchases, storages, tolling and other charges, must

be taking into account in the arbitrage pricing model. This introduces complexities

of power data and modelling must be therefore be understood and embraced by the

modeller. In addition the demand and supply shocks introduce regime-switches.

There is therefore the need to obtain an accurate forecast of the electricity spot

prices and to take into account the feature listed above. The next section examines

some of the main research which has been done. In addition the various

methodologies and conclusion drawn will also be looked at.

Jason et al. (2004) studied the use of correlation filter as a measurement of

performance of the spread trading models. They used a data which consist of daily

closing prices for the period 1995 till 2004 of WTI Light, sweet crude oil futures

contract and Brent Crude oil futures contract. Neural network Regression was used

as a bench mark model against other methods of forecasting such as Moving

Average, ARMA, Co-integrating ”fair value” model. The forecasted sample was

then exported to excel and both a standard and a correlation filter were applied.

Their findings suggested that ARMA model is the best. The ideas and concepts

from this paper inspired the dissertation. They concluded that “the correlation

16
filter can provide traders with a useful way of improving the performance of

spread trading and subsequently should be studied further”. However, their

conclusions were based on single market. In other words data from different market

such as electricity which has features like volatility clustering, jumps, seasonal and

regional variation are likely to give different results if theses are not taken into

consideration. Their research was from an academic view point which is usually

frowned upon by market traders. In addition part of their research was based on

moving average which is trend recognition and unidirectional which has a

disadvantage of lag in the data. This is usually resolve by practitioners using

counter-trend systems such as Bollinger bands which takes into consideration

volatility, peaks and trough within a certain range. This also includes protective

stops which minimises unlimited losses. I would therefore like to extend this

research by making this relevant to market by using tools used by traders on a daily

basis. The other aspect of Evans et, al; that is of interest is the extension of Holmes

et, al (2003) “Fair Value” approach. In their conclusion they stated that the fair

value principle is an example of a test in market efficiency. In which both arms of

the spread contracts are “efficiently priced”. Based on this they tested the predicted

value of the spread using Technical analysis models. The data on which their

research was based is oil futures which are considered as a commodity and whose

price follows a random walk. The general acceptance of their conclusion could be

stronger if they had used data from other sources or markets. Their approach will

be used but with different data and product-electricity.

Hadsel (2006) uses TARCH to examine the returns in electricity futures volatility.

Evidence from his study confirmed; “asymmetric reaction of traders to new

information”. Traders reacted by “increasing trading volume and adjusting prices

to a new equilibrium”.

17
Weber et, al (2007) extended ARMA models by using it to forecast day-ahead in

electricity markets. The combination of Gaussian-mixtures and switching-regime is

similar to the ARMA models. They concluded that GARCH approach captures the

fat tail in the price distribution. In other words “high volatility coincides with

strong price change”.

Mark et, al (2003) used EUR/USD returns to investigate regression models as a

tool for trading and investment. Neural network (NNR) was used as a bench mark

model against Box-Jenkins type of regression models. They concluded that,”

regression models and in particular NNR do have forecasting abilities and an

added tool for quantitative trading”. The literature above assumes deterministic

price volatility, but evidence from electricity markets suggest otherwise. It is

therefore appropriate for the modeller to use models which incorporate loads,

weather and plant variables. In other words, cyclical patterns that dominate electric

spot markets must be considered. This includes the so-called: hour-of the-day

effects, day-of-the-week as well as seasonal effects. These will be included in

models by the use of dummy variables. The various models used by previous

researchers for forecasting commodity or electricity will be discussed in this

section. Both linear and non-linear series will be considered. Each of these will be

derived and discussed. In addition the suitability of the model to the data will also

be looked at.

The dissertation will also examine and apply a paper by Bachelor, et, al, 2007.

They examine sport and forward rates in freight rate assuming shipping is a non-

storable service; using the Forward Freight Agreement (FFA) market using

forecasting methods such as vector equilibrium correction (VECM), ARIMA and

VAR. They found out that the ARIMA or VAR provided a better forecast of spot

rate. The test was based on the following assumptions: firstly the FFA market is

18
small and new; secondly they regarded the forward rate as a services which is not

storable, suggesting that there is no arbitrage between the spot and futures rates –

non existence of cost and arbitrage relationship; finally, the FFA market is made up

of hedges rather than speculators – as speculation often leads to biased with regard

to forecasting. They concluded that the spot and forward rate do not cointegrate

which was contrary to their expectation.

Tai-Leng (2006) studied the profitability of directional indicators such as

Directional Movement Index (DMI) trading rule. Their aim was to test DMI trading

strategy to Hang Seng, KOSPI, Nikkei 255 and TWSE then compare this with the

UK and US stock market. Their result concluded that the strategy performed well

with the Asian market with the DMI. In addition they found that buy-and-hold

strategy did provide more positive returns for the Asian, UK and US data. In effect

their research was to test market efficiency; they found out that the US/UK markets

are more efficient than the Asian markets. If that is the case then one will be forced

to conclude that, technical analysis does not have any predictable ability. However,

their result was inconclusive in this regard.

Weber (2007) applied ARMA in conjunction with GARCH, Gaussian-mixtures to

model electricity price German data. Their conclusion was that the “extended

model” did actually capture the characteristics often exhibited by power data, and

hence recommended the forecasting abilities of theses models in general and

ARMA in particular.

Williams M (2003) investigated the predictability of regression models for the

forex markets. Their results confirmed the forecasting ability of these tools;

however their study did not test the suitability of these tools to different markets or

data.

19
Burgess (2003) used daily closing price of the STOXX 50 index to examine the

potential of Cointegration techniques to hedging and trading international equities.

He used the properties of Cointegration, which assumes that if Cointegration exist

between two of more variables, it is possible to establish a long-term equilibrium

relationship, which included an error correcting (ECT). This error term can be

classified as mispriced which opens up opportunities for arbitrage. I addition the

ECT can used in statistical arbitrage trading strategy in general and pairs trading in

particular. Since the ECT model is an equilibrium model; which tend to revert to its

long term mean; it is this property which enables it to be used for hedging and

trading purposes. The result from his study provided a further proof of the

usefulness of, statistical forecasting tools to trading.

Chaun et al. 2001 applied two technical trading rules- moving average (MA) and

Channel rules (CR). They applied this to 13 Latin American currencies. In addition,

the use of their research is to find out whether theses markets are efficient. They

found out that four of the currencies showed positive returns for MA and only three

for CR. The four countries include: Brazil, Mexico, Peru and Venezuela and the

remaining three, Brazil, Mexico and Venezuela. Their result suggested that the

difference in the returns of the trading rules was due to statistical properties

peculiar to the various currencies, which calls for further research.

Bracj et al. 92, tested technical rules namely moving average (MA) and Trading

range break (TRB), using Dow Jones Industrial index data from 1897 to 1980.

Their result proved highly positive returns from these technical rules. However the

returns from “academic” tools such as AR(1), GARCH-M, random walk and

Exponential GARCH models, were not consistent with the trading rules. Their

studies focused on simple trading rules we the hope that it can be further extended.

Their result proved that technical analysis does have predictable ability over the

20
academic ones. In addition the found out that technical analysis do account for the

patterns in stock prices which cannot be accounted for by AR models. These

patterns include returns for long signals are larger than that for short.

Marshal et al. 2006 used CRISMA a technical trading system, which is a

combination of momentum, volume and relative index indicators to examine the

profitability of trading system. They used CRSP stock from 1976 to 2003. Their

funding suggested that the system did not provide consistence returns. Previous

researcher however used small samples which provided consistent returns. Thus in

their final conclusion they suggested that profitability of any trading system

depends on the sample size.

Burgess (1990) in his thesis examines the Statistical Arbitrage (StatArb) model as a

low risk trading tool. A StatArb is an application of statistical and computational

techniques to exploit the short term fluctuations or mispricing of stock prices. A

simpler form is pairs trading; which involve buying one share and selling the other,

unlike pairs trading, StatArb exploits the mispricing of more than two shares in a

portfolio, by buying one set and selling the other. Neil Burgess used Cointegration

as the main statistical tool for his thesis. He selected series of cointegrated stocks

and then modelled the mispricing using the Error Correction model concept, with

the aim of capturing the mispricing of the stock in the portfolio. Finally he used a

risk minimizing techniques to select the best stock for the portfolio. The closing

prices of the FTSE 100 index and other indices were the main source of data. He

concluded that statistical and computational methods do have practical application.

Furthermore the model ensured profitability with or without transaction cost under

certain risk limits. However in his views the result was inconclusive, since the data

used were historical, which did not take into consideration, real data in a trading

environment. In spite of this, he concluded that computational and statistical tools

21
do in his own words, ”demonstrate significant potential”, hence calls for further

research.The next section gives a brief description of the models that are widely

used by researchers.

Autoregressive moving average

The ARMA model assumes a set of data with an error term that normally

p q
distributed. The general model is of the form: Yt = iYt i i 
t i 
~.
t
i
1 i0

~ is a random error component with zero mean and variance 


Where:  ~ 2 ; with p

denoting the maximum number of lags on Yt .This model is made up of two

p
components, autoregressive (AR): Y = iYti 
~ and moving average
t
i1

q
(MA): Y = it i . The observations are usually calculated based on previous
i0

values. An extended form of the ARMA model is ARIMA (p, d, q)

d Yt 0  1 ... 
1 Yt 
d
p Yt p 
d
t 1t1 ... q t q

Where: Yt Yt Yt 1 . The parameters are estimated by applying the maximum

likely hood estimation procedure to the density function of the distribution.

The generalized-autoregressive-conditional heteroscedasticity (GARCH)

This GARCH(r, p) model is as follows: rt t t

r p
t2 it2r d ti it2i . The parameters of this model are usually
i0 i0

estimated using the maximum likelihood principles. The error terms are assumed to

be random and follow a normal distribution. There are different types of GARCH

22
models; which are distinguished by their parameters and most are suitable for

electricity price modelling; as they tend to capture the extreme volatilities. The

simple GARCH (1, 1) is sufficient in most hence it is going to be used in this

dissertation; however different parameters can be used to determine an appropriate

model.

Naïve methodology

The Naïve or random walk model is of the form: Yˆ


t Y t1 + t , where t , is a

random error. Random walk is not mean-reverting, in other words the price does

not return to the mean of the prices. Its trend cannot therefore be predictable. The

Naïve model is simple and enables us compare actual values with forecasted ones.

It does not take into considerations factors such as trends, seasons and in addition

requires a minimal amount of data.

Error Correction Model (ECM) methodology

The ECM assumes that if variables Yt and X t are non-stationary it is possible to

find a linear combination of both variables which will be stationary. A combination

of first difference and lagged levels of co-integrated variables forms equilibrium

model: Yt o X t 1 (Yt1 X t1 ) t or three variable

version: Yt o X t 1 t 3 (Yt1 0 X t 1 1 t 1 ) t , were Yt1 X t1 is

the error correction term (ECT). Furthermore provided that Yt and X t are co-

integrated with co-integration coefficient , theYt1 X t 1 , will be I (0) even

though the constituents are I (1). ECT is a short term deviation from the long-run

equilibrium position. If such deviation exist, then it is possible to exploit it for

speculative profit. Burgess (2003) examined this technique in his thesis. Similarly

Law et al (2004) used the above techniques to test for market efficiency of futures

market. They used a simplify version of the model:

23
Y=a +bX +e

According to them for the requirement of market efficiency the following four

conditions must hold:

i. Y and X must be of the same order

ii. Y and X must cointegrate

iii. B the slope equation must be equal to 1, with the constant a=0

iv. The forecasted error must not be correlated.

In other words for a short term efficiency, ECT must exist. They held the view

shared by other researchers, by stating that for inefficient market traders can take

advantage of this to make profits. Their final conclusion stated that a fair amount of

profit and losses could be made by using trading strategies. This conclusion will be

tested using different trading strategies and data. The trading strategy will be tested

whether the market is efficient or not.

The recent growth of energy exchange such as Nord Pool has generated a lot of

interest in modelling and pricing electric futures. The next section gives an

overview of markets associated with this exchange.

24
European Electricity markets: an Overview

Nord Pool power markets embraces Finland, Sweden, Denmark and Norway,

emerge as a result of deregulation that is sweeping throughout Europe and the rest

of the world. Deregulation broke government monopoly over power generation,

paving the way for private sector participating in the sector thereby increasing

generating capacity and transmission efficiency. Power is now considered to be a

tradable competitive commodity - it can be index like any other commodity;

making it an important tool in a free market economy. This has made Nord Pool a

multi billion market which is rapidly evolving and developing. Risk management

practices which are common with financial assets such as stokes and bonds in the

new market. The Nord Pool sports market is called Espot Market. It is an auction

based for trading in physical power, which uses bid for buying and selling. There

three types of bids: hourly, block and flexible. Market participants after submitting

their bids a spot price usually called system Price is calculated by using aggregate

supply and demand curves. The intersection of these two curves forms the system

Price. The system price is a reference price for all contracts traded in the derivate

markets. It takes into consideration consumptions variations – plans closure, fuel

prices, seasonal variation, reservoir levels and other bottlenecks that affect the four

Nord Pool members. To be able to model power prices effectively there is the need

to understand factors influencing power prices. There is a general view held by

researchers about the unique features exhibited by electric prices. These views

include: non storability of electricity; seasonality; price spikes; transmission

bottlenecks. Nord Pool specific reasons for power prices various are not so much

different. Their annual report of 2006 provided reasons for theses price fluctuations

25
excerpts of which follows:

Economic Factors

The main economic factors which influences prices was identified in the report

these are: Level of economic activity; currency movements and generating

capacity. Increasing generating capacity brings into its wake increase in supply

with its attendant low prices, depending of other relating cost. Business cycles with

Europe and the rest of the world affects power generating inputs such as coal, any

fluctuation in prices of any of theses do affect power prices in the Nord Pool

generating area. Most of the raw materials used in power generation are priced

using the US dollar which is subjected to fluctuations. Nord Pool benefits if the

exchange rate is low but loses if it is high.

Weather and temperature condition

Nord pool electricity is generated from sources such as hydro, nuclear and coal;

any factor the affect theses source is likely to affect the supply and consequently

the prices. Similarly, rain, snow tends to raises the level of the rivers that feed the

turbines consequently increasing supply thereby reducing prices. Thus, periods of

high precipitation leads to high supply and lower prices; while a low precipitation

means lower supply and high prices.

Finally temperature trends within the year do affect prices. For instance colder

temperatures in winter month’s results I high demand for electric and consequently

high prices.

Electricity transmission bottle necks affect prices in two ways; firstly limited

capacity in the networks tends to affect supply and demand. If the demand exceeds

supply in such areas, prices would tend to rise. Secondly, other non-Nordic markets

26
trade with Nord Pool with recorded inflows and outflows, power supply and

demand in these markets also affects prices.

Commodity Prices

The cost of commodities such as coal, raw which serves as inputs or raw

materials for power generation affects prices significantly. Regional generation

stations in European rely on identical inputs - these two markets compete for this

essential commodity, thus pressures due to supply and demand have an enormous

influence in price determination in the entire region.

Carbon dioxide emission allowance (EUA), a recent phenomenon is a tradable.

Inputs such as coal, gas do emit carbon dioxide, thus generating stations are

required to buy EUA and the cost is therefore incorporated in pricing of power.

Nuclear power stations account for 30% of the power supply, any factor the affects

the power stations will also affects power prices.

Thus the key factor listed above explains, seasonality, cyclical pattern and

occasional dramatic price change – spikes followed by return to normal levels

explains the System Price path. These unique features of power markets present a

challenge and thus call for new models development and study.

The Financial Markets

The Nord pool ASA is one of the third arms of the Nord Group; this financial

Market is made up power producers, retailers and stock brokers. The aim of theses

participants is to use the contracts traded in this market to manager their risks.

Contracts traded includes: base load futures, forwards, options and contracts for

difference (CFD).The futures contracts consist of: day contracts base load for the

period of 24 hours; weekly contracts; base load for a period of 7 days. The

27
Contracts period ranges from eight to twelve months. The future settlement uses a

daily mark-to-market settlement and a final settlement of cash which is a difference

between final closing prices and the System Price. Daily gains and losses are

reflected in the mark-to-market settlement. The Forward on the other hand consist

of three seasonal contracts, which is also contract mark-to-market; however any

profit and loss is accumulated and realised at the end of the trading period.

Option Contracts

Options contracts are one of the derivatives traded at Nord Pool ASA which offers

risk management strategies. Option traded is mainly European in nature, which is

exercisable only at the exercise date. Most of the contracts consist of quarterly and

yearly forward contracts. New contracts are listed at the expiration of the old ones.

Exercise day is usually the third Thursday of the month before delivery. Strike are

normally set buy the pool, with a price which is usually quoted in EUR/MWH and

a tick size of 0.01 EUR. Well know option strategies such as straddles; strangle

which combines calls and puts are used by traders wanting to make profit.

Contracts for Difference (CFD)

An alternative to futures and forwards contract is the CFD, which is the difference

between and area price ant Nord Pool System Price. CFD is considered to be a

forward contract that is used to exploit prices differential in the System price.

Growth

According to the 2006 Account Statement, Nord Pool is the largest Power

Exchange in Europe with volume of trade amounting to 2220 with market value of

Euro 79.2 billion traded in the Financial (OTC) trades alone- an increase of 3% and

51.35% in the previous year. Similarly contracts in the financial market in 2006

28
amounted at a value of 766Twh with a corresponding value of Euro 36.3 billion

and a total transaction of 89147 and increase of -2.55%, 52.57% and 16.81%

respective love the previous years. Thus the market has seen a phenomenon growth

over the reporting period.

S&P500 Index

The S&P 500 index is made up of 500 leading companies in the US, with about

70% of all US stocks/equities. It has a market capitalization of 12,867 billion

dollars. Some of the sectors include energy, industries, information technology,

telecommunication services, consumer staples, health care and banks. Theses

companies include world renowned ones such as: Microsoft, Cisco Systems,

General Electric, AT&T, just to mention a few. Thus the S&P500 can be used as

proxy of a perfect market; it also has an added advantage of reflecting the common

risk and returns of the various sectors as enumerated above.

Figure 1: System Price determinations

Source: Nord Pool

29
A few researchers has research into mainly modelling and pricing of electric

futures; but little in terms of modelling and forecasting an important product such

as Contract of Difference(CDF), which is actively traded in that exchange and

other exchanges. The aim of this dissertation is to apply the methodologies and

tools enumerated in the literature review, to power price data. Common statistical

and forecasting tools which are available to most researchers will be used to model

and forecast the price data with the aim of finding out any predictability power and

whether the results can be exploited by market traders in identifying opportunities

presented by these tools.

Technical Analysis: An overview

Technical analysis is a method used to analyse past data with the intention taking

investment and trading decision. Techniques have been developed to determine for

example: trends, cycles commonly found in market data. Market traders generate

profits by following the market trend which can be an uptrend, downtrend or range

bound. Chapter 3 introduces some of theses trend measures and the signals they

often generate and finally the main strategies often associated with them. The

indicators are constructed by using a formula to transform a price data series into

another. Theses indicators are of graphical nature usually plotted on the price data

or in most cases below it.

The measures that are going to be considered includes: Relative Strength

Indicator, Moving Averages, Exponential Moving Averages, Moving Average

Convergence-Divergence and Bollinger Bands, since these are the commonest

tools used by market practitioners for analysing these trends.

30
Figure 2: Graph showing moving average on the price data; Volume and MACD

indicators at the bottom, traders use these indicators for short term forecasting.

31
Chapter 3 Methodology & Methods

Introduction

This chapter examines the forecasting models in detail, their properties, method of

constructing, evaluating and comparison. The final section briefly examines models

of some of the technical indicators and a list of trading rules often used in

conjunction with them. The remainder of the section provides a list of some of the

profitability as well as volatility measures which is used to compare the forecasting

abilities of the trading strategies.

Box Jenkins or ARIMA methodology

Box Jenkins classical methods are one of the best econometric forecasting tools. It

embodies model identification, evaluating or estimating and model validating. Box

Jenkins methods unlike random walk model are mean reverting and stationary.

They have other properties such as, trending, deterministic, which makes them

useful forecasting tools. This section examines some of these models.

ARMA (p, q) and ARIMA (p, d, p) models

The autoregressive (AR) model is a class of time series denoted by AR (p) and of

the form: ~
Yt 1Yt 1 ... p Ytp  t =1, 2… T. 3. 1
t

p
= i Yt i 
~
t 3. 2
i1

~ is a random error component with zero mean and variance 


Where:  ~ 2 ; with p

denoting the maximum number of lags on Yt .

32
Similarly the Moving average (MA) model is of the form:

Yt t 1t 1 ... q tq 3. 3

q
= i ti
i0

ARMA (p, q) model is a combination of moving average (MA) and autoregressive

(AR) process. It is unvaried in nature and therefore suitable for modelling spread

series. It is capable of predicting futures values using past values of a data,

provided an appropriate values for p and q are chosen, it also has the capability of

mimicking any stationary series of interest without an economic theory. ARMA

models are traditional method of modelling electricity prices. The ARMA (p, q)

model is of the form:

p q
Yt = i Yt i i  ~
t i t 3. 4
i1 i 0

ARIMA (p, d, q)

ARIMA (p, d, and p) model is and extension of ARIMA (p, q) model with an

additional parameter d which is the difference factor responsible for transferring

non-stationary series to a stationary one. It is of the form:

d Yt 0  d
1 ... 
1 Yt  p Yt p 
d
t 1t1 ... q t q 3. 5

Where: Yt Yt Yt1 .

Fitting an ARMA model to a data involves: model identification or test for

stationary; estimation of parameters; finally a diagnostics test which includes

checking for the significance of the parameters.

33
Vector error correction model (VECM) & Vector Auto regression

(VAR) methodology

VAR can be used to forecast electricity prices since it does not need any theoretical

justification for its use. It is capable of explaining the random disturbances that

influences the spreads, since it treats the two arms of the spread as endogenous

variable. In addition is also enables the impact of demand and supply as well as

regime-switches between static and dynamic volatile periods to be incorporated

into the model. The model for Spot(S) and Forward (F) are:

VAR (p);

p p

S t o i S t i i Ft j t . 3. 6
i
1 j1

p p

Ft o i S t i i Ft j t 3. 7
i
1 j1

VECM (p)

p p

S t o i S t i i Ft j 1 ( St 1 0 1 Ft 1 ) t 3. 8
i
1 j1

p p

Ft o i S t i 

 
 

i Ft j 


2 ( S t1 
0


1 Ft 



1 ) t 3. 9
i
1 j1

34
General VECEM (p)

Pt o 1 Pt1 2 Ft 1 3 Gtl ( Pt 1 0 Ft1 
1 G t1 ) 
ˆt ; 3. 10

With error correction model term (ECT):

ECT= Pt 1 0 Ft 1 1 Gt 1 or

R t =Rt 1 t where: t ~N (0,  ) 3. 11

The error correction term defines the long term relationship between variables

and the speed of adjustment is denoted by the parameters: 1 , 2 , a value of zero

reduces the VECM to VAR.

Cointegration method

Long term relationship between variables can be modelled using co-integration.

If variables are co-integrated then the spread deviation will not depart from the

long equilibrium position in the long-run. Trading strategies can then be set up if

the variable co-integrated otherwise the spread will deviate without bounds thereby

making the spread a risky one.

Testing for co-integration requires the series to be non-stationary at the prices

level with the same order of integration. A stationary series has the additional

property of a constant mean and a constant variance and independent of time. We

are going to test for co-integration using Johansen approach. Let consider the

following equation:

Yt Yt 1 t 3. 12

If 1 , equation (1) is said to be integrated of order 1- I(1) and non-stationary and

of unit root. A unit root can be tested by using augmented Dickey Fuller (ADF) and

Phillips Peron tests. The test uses the following model:

35
n 3. 13
Y t  0  1 Y t 1   i 2 Y t 1 i
i 3

Where, Y indicates the various prices and Yt Yt Y t 1 with n the number of

lags. The test will examine whether the coefficient 1 is < 0, an indication of

stationary series. The null hypothesis to be tested is: H o : 0 against an

alternative of: H 1 : 0 . A rejection of the null hypothesis indicates the presence

of a unit root. If the first difference of the series is stationary, ADF and Phillip-

Peron test will suggest, the presence of unit root.

In addition a test of the presence of auto correlation is to be conducted using LM

test. If the test suggests the presence of autocorrelation, further test should be done

by adding lags variables until autocorrelation was removed.

The first difference rejected the null hypothesis of a unit root at levels, 1%, 5% and

10% respectively. This is an indication that System and Helsinki spread series are

stationary and thus can be modelled in a co-integration relationship.

The next section introduces pairs trading; an application of ECT concept; which is

a Statistical Arbitrage method in its simplest form. The System and Helsinki prices

will be used to model a portfolio which will be tested later using a moving average

strategy.

Co-integration “Fair Value” Approach-Pairs Trading modelling

Cointegration test enable us find whether an economic relationship exit between

the two variables in the spread. If, it indeed exits, then there must be a temporary

widening of the spread in either direction and in addition the two arms of the

spread must be capable of reverting to the long term mean. This will enable traders

to short (buy) the spread. This amounts to testing for statistical arbitrage. Johansson

36
co-integration test is use to find a long-term spread equilibrium. Firstly an in-

sample is used to find the co integration vector (ECT):

ECT =PP t- GPt - PFt = PP t - St 3. 14

“The fair value of the spread is then calculated”, based on the above relation

The trading signals are as follows:

If PP t < S t , then go long the spread until fail value is gained, similarly

If PP t> S t then go short the spread, until a fail value is regained.”

Evans & Laws (2004), the standard and correlation filters are then applied. The

result of the Cointegration test of the System and Helsinki prices proved that they

cointegrated with [1,-0.019635] as the Cointegration vector. As a result it is

possible to form a portfolio, V of the two asserts of the form: V (t) =100S t -

1.9635H t .The trading strategy is: buy if V>0(mean), sell otherwise. The mean in

this regard can be the moving average of the portfolio values. If transaction cost k

is included; the new trading rule becomes: Buy if V MV K , sell otherwise;

1 m 1
with the moving average, MV= V m i .
m i 0

The next section examines the two main volatility models which are going to be

used in the forecasting processes.

37
Volatility Model methodology

Introduction

On examination of the System price reveal that they are very high during winter

and summer months. Other period tended to be relatively stable but price spikes

appear to be evenly distributed throughout the period under study. This behaviour

patterns suggest a non-constant variance or time dependence, which is an

indication of a random volatility. This particular behaviour is commonly seen in

commodities prices such as crude oil, interest rates and equities. The best model

which explains this phenomenon is Autoregressive Conditional Heteroscedasticity

(ARCH), they have been shown to model leptokurtic distribution effectively. Some

of these models will be considered in next section.

Comparing simulated futures price volatility and GARCH (1, 1)

The Schwartz single factor model can be used to simulate the System Price:

The following model is going to be used: X t X t 1 


ˆ(
ˆX t 1 )t t  t

GARCH (1, 1) Model

GARCH (1, 1) is one of the stochastic volatility models. It models data with

varying volatility it is of the form: r t = c + t with, t2 w0 w1t21 w2 t21 .

Where; t2 is the conditional variance of the returns and t21 ; the squared random

St
error component, r t is the logarithm of the returns, define as: r t =log ( ) ; and
S t1

S t is the System Price.

38
TARCH Model with seasonal dummies

Electricity prices have been observer to exhibit a period of static volatility

followed by period of dynamic or large volatility of volatility clustering due to

seasonal patterns. There is therefore the need to use a model which captures

theses features. In addition the assumption of constant variance no longer

appropriate in this situation. A GARCH model which allows the variances to be

conditioned on the past errors is the most is the best in this situation. TARCH or

Threshold ARCH is one variant of GARCH models has additional property of

capturing asymmetric response in electricity futures prices. In addition it

captures persistence and mean-reversion. The following model will be used:

rt t t

q r p
t2 j t2j i t2r d t i i t2i Di . 3. 15
j
1 i 0 i0

Where d t =1 if t 0 and 0 otherwise; where: D i represents the various

monthly dummies. According to E-views 5 TM :

“Good news means t i 0, and bad news t i0 , have differential effects on

the conditional variance; good news has impact on i , while bad news has

impact of

i i . If i 0, bad news increases volatility, and we say that there is a

leverage effect for the i-th order. If i 0 , the news impact is asymmetric”

TARCH (1, 1) model is of the form: t2 t21 t2l d t 1 t21 Seasonal

dummy variables and a mean equation: rt t t .Long-term variance =


2  3. 16
1 0.5

39
The GARCH-M volatility model has been found to incorporate the mean equation,

and therefore provides a good forecast, thus it will be used as the main forecasting

tool for volatility.

Summary

We have examined the properties of the main forecasting tools, which include

Random walk, Box Jenkins and GARH volatility models. The strength and

limitation and the context under which they can be applied was also explained.

Finally the concept of Cointegration was also examined. In the next section the

actual series will be model, examine, with an additional aim of finding properties

such as: stationarity and mean-reverting. The existence of these properties will lead

to a possible model for the data.

Spread Series

This section will examine the suitability of the mean-reversion model as applied to

electricity futures prices, using test for normality and mean reversion. The

logarithms returns are going to be examined for leptokurtic behaviours using

quintile plots, kurtosis and skewness. If the statistics indicate a small negative

skewness and high kurtosis the jump diffusion process is appropriate for modelling

the future price. Finally, a clear understanding of sport futures price dynamics will

be explored in this section. The data to be used consist of closing prices of the

following futures: System Price (SP), Helsinki Price (HP). The period of the data

is from 29/12/03 to 31/12/04, with all data coming from Nord Pool (NP).

40
The following formula show how the spread series R t - CFD is defined as:


HP HPt1  
SP SPt 1 
“R t =  t - t  3. 17
 HPt   SPt 1 

Where:

SP t = System spot price at time t

SP t 1 =System spot price at t-1

HP t = Helsinki spot price at time t

HP t 1 = Helsinki at time t-1

Laws et, al (2004).

41
Data Description

Daily closing settlement price from the Nord Pool Power Exchange, with data

period spanning between 01/06/2005-29/12/2006 will be used. This data will be

split into two samples. The first is in-sample and the second is out-of-sample. The

in-sample is used for initial test and modelling. The in-sample will consist of the

first 250 sample. Whilst the out-of-sample is going to be use to evaluate the

forecasting accuracy and testing the trading strategies, the last 150 prices will be

used as the out-of the sample data. On the examination of CDF daily prices, it was

found to exhibit characteristics already identify by Swider (2006), where they

listed: “daily, weekly and seasonal cycles; high volatility; mean-reversion;

frequent prices spikes”. They attributed this short term fluctuations to non-

storability of electricity which leads to peak supply and demand not being able to

match. Furthermore a summary statistics of the range series, suggested a non-

normal nature of the data, this was confirmed by the Jarque-Bera statistics. The

data is highly skewed with very high kurtosis which also suggests non-stationarity.

Theses evidences calls for the data to be transformed to make it stationary. A log

difference of the CDF series also showed the spiky nature of the series. The

summary statistics showed an improvement with Jarque-Bera statistics still

indicating a non-normal series with very low skew, which is a characteristic of

stock market prices. Swider (2006) and others viewed this characteristic as fat

tails, hence call for a new form of modelling which incorporates conditional

heteroscedasticity modelling. Hence they suggested using ARCH test to the

presence of heteroscedasticity.

On examination of the system price, the Jarque-Bera Statistics gave a value of

20596.20 which lead to the rejection of the normality of the price distribution. The

42
data was transformed by taking the logarithms of each of the prices. A log

transformation of the data suggested stationarity. A test of unit root to confirm

stationarity using Dickey Fuller (ADF) test and Philips-Peron test statistics rejected

unit root at significant level of 1%, 5% and 10% respectively. Thus the log-returns

are stationary. The summary for the log-returns showed a very large kurtosis and

skewed data. The Jarque-Bera statistics rejected the normality of the returns. It

exhibits leptokurtosis and volatility clustering which is a common feature of

financial data. A close examination of the Quantile-Quantile plots proved that the

log-returns is not normal, as there were some data at the extreme ends of the

reference line- a clear departure from normality. Thus the log returns is leptokurtic

and hence follows a jump diffusion mean reverting process.

Mean Reverting Models with jumps

Merton (1970) initiated the idea of jumps in financial models and subsequently a

lot of researches followed suite. However, the approach used in this section is

based on the one developed by Lumberton et, al (1995). They argued that Black-

Scholes model assumes complete market scenarios; rare events such as earthquakes

cannot be modelled using Black-Scholes. They proposed discontinuous stochastic

model that in their view captures this jumps as they called it. Thus financial data

with jumps can be modelled by extending the Black-Scholes model to cater for

jumps, which a modelled using the Poisson process.

Let N t be a sequence of non-negative independent and identical distributed

random variables, the random variable N t is said to follow a Poisson process with

parameter with the following probability distribution:

(t ) x
P( N t x ) e t x=0, 1 …n
x!

43
With the following properties:

E (N t ) = t, Var (N t ) = t

Let begin the derivation of the jump process by using a risk less asset define by S

(t) =e rt

Where, r is the interest rate and t the time. Let define the proportion of jumps by

consideration the following random variables: 1 , 2 ...k ... at the within jumps

times’ define by: 1 ,2 ,....k .. and assuming each of the jumps follows a Poisson

process.

Let assume the variables: w1 , w2 ,... for t>0 follows a Brownian motion. Finally let

assume the price of the asset is X t for t> 0. According to D. Lumberton, B Lapeyre

In the interval [ k ,k 1 ]

dX t =X t dt X t dWt

At time k the jump of X t is given by:

X k X k X k  X k U k ,

Thus X j =X  (1 U k )
k

So we have for t [0,1 ]

X t X o e (  / 2 ) t wt
2

Consequently the left-hand limit of 1 is given by:

X 1 X o e (  / 2 )t w1
2

and X 1 X o (1 U1 )e (  / 2) t w1


2

Then for t[1 ,2 ]

44
( 2 / 2)( t 1 )
X t X 1 e ( wt1 wt1 )

The above process is repeated till we obtain:

X t X 0 (J 1 (1 u J ) e (  wt


nT 2
/ 2) t

0
With the convention 1 j 0

The solution of the deferential equation satisfies the equation:

t N f

X 0 ( ds dws ) X j U j


X t X 0 
0 j
1

Mean reversion is the ability of energy or electricity spot prices to return to a

long term level. Theses are models base on Ornstein-Uhlenbeck process and are of

the form:

dP t = 
ln Pt 
Pt dt Pt dWt (2)

The model can be amended by adding another term with takes care of jump

diffusion or spikes. This model is the Clewlow & Strickland (2000) jump-diffusion

mean-reversion SDE

dP t = 
K ln Pt 
Pt dt Pt dWt +P t dQt (14)

Where, Q t is a Poisson process with intensity k, and jump size  that is distributed

normally with mean j and standard deviation j . The forecast model is:

Pˆt 1 Pˆ1 


t j . (15)

Clewland & Strickland (2000) estimated these parameters using a Recursive Filter

Which caters for autocorrelation, where, Pt is the price at time t, the long term

ˆe , the speed of reversion rate  and W t a Wiener process.


mean is 

45
Integrating the equation gives:

t
P t = e t S o (1 e t ) 
e ( st )dWs
t0

(3)

The discrete version is:

Pt =c + Pt 1 t

(4)

Where c= (1 e ) , e 

(5)

This equation is an autoregressive process of order 1 i.e. AR (1).

The spread is going to be modelled to include seasonal trend:

12
R t =Dt AR (1) t
t 1

(6)

Seasonal dummies are included in the model to capture the effect of weather

and to explain the impact of the demand side effects.

Parameter estimation

The parameters that are going to be estimated include: Mean reversion rate, jump

rate, standard deviation of jumps, and average number of jumps. Finally diagnostic

tests will be conducted on all the equations for parameter breaks, outliers and

ARCH effects.

Mean Reverting models with time varying mean

These models incorporate seasonal variation found in electricity prices

dP t = 
t ln Pt 
Pt dt Pt dWt (7)

46
Where:

t = Dayt + Montht + Hourt +Trend t (8)

This model has a discrete form:

Pt = t + t Pt 1 t . (9)

This is an ARMA (1, 0) model with dummy exogenous variable which takes

into consideration an hour, day of a week, month of a year.

Rt = ARMA (0, 1) + seasonal dummies.

Simulation of Electricity Prices Returns

Introduction

Monte Carlo method is going to be used to simulate future price using its historic

parameters, since the simulated returns are realistic and closer to the market data.

The parameters to be estimated from the actual market data includes: mean,

deviation, mean reversion rate, jump rates. This is an aid to choosing the right

model. Figure 3.0 shows the logarithms returns of the system price. Theses returns

do fluctuates around zero, which must be due to supply and demand, weather and

market related information.

47
30-Day Volatility

0.080
0.070
0.060
volatilty

0.050
0.040
0.030
0.020
0.010
0.000
0 50 100 150 200 250 300 350 400
time

Figure 1: 30-day volatility of logarithms returns of the System price

Figure 1 shows the system price volatility using a 30-day rolling returns. The

simulated path appears to capture seasonal patterns due to weather fluctuations,

which is prevalent in most electricity markets

Figure 3.15 show a comparison of the logarithm of the system prices and 30-day

rolling standard deviation. An inspection of these two paths show that prices are

much higher in summers and winter as shown by the peaks or spikes. Theses peaks

are spread throughout the whole year.

48
30-day volatility & skew

6.00
5.00
4.00
3.00
Returns

2.00
1.00
0.00
-1.00 1 39 77 115 153 191 229 267 305
-2.00 30-day skew
-3.00
30-day volatility
Time

Figure 3: A combination of 30-day volatility and skew of returns of the

System price

Figure 3.3 shows a plot of 30-day volatility and skew. There appears to be a

relationship between skew and volatility, if the volatility is positive skew tends to

be the same, however if the volatility decreases the skew decreases and sometimes

with a negative value. In other words there is a correlation between volatility and

the skew; this is an interesting finding which needs to be investigated further.

Modelling

Schwartz, E (1997) describes a one factor model of the logarithm of

Spot price of commodity prices assuming the prices follows Ornstein-Uhlenbeck

mean. The reverting process of the form: dS= (


ˆln S )Sdt Sdz .Equation (3.1)

can be written if a different format by letting X =lnS:

dX = (
ˆX ) dt dz

3.2

2
With 
ˆ
2

3.3

49
The speed of adjustment a non-negative constant , which measures the rate at

which the mean reverts to its long term mean of 


ˆ. If a market risk  is

introduced, equation (3.1) now becomes:

dX = (
X ) dt dz 3.4

Where: 


The model is and extension of Geometric Brownian Motion(GM) by incorporating

mean reversion, which takes into consideration the convergence of the spot prices

to its long term mean, thus balancing the supply and demand fluctuations. S is the

spot prices, ̂ is the mean and  the volatility and dz is a Wiener process;  is the

mean reversion rate, which is always positive. The long term mean of the spot price

is e . The drift is given by ( 


ˆln S ) , which is a measure of how the spot price

will drift to its long term mean.

ln( 2)
The half-life is this measures the time taken for the series to revert to its long

term position.

Equation 3.2 can be stated in the form:

dX t X t dt dWt 3.5

It can be show to have a solution of the form:

t
e 
t t z
X t  xe e dW z 3.6
0

With expectation and variance of:

E(X t ) =X o e t 3.7

1 e 2 t
Var(X t ) =E ((X t -[E ( X t )] 2 ) =2 3.8
2

50
For the purpose of simulation a discretised form of equation (3.1)

X i = X i 1 + ( 
ˆX i1 )t 
i t With S (t) =exp(X i ) 3.9

Mean Reversion parameter estimation

The various parameters of the mean reversion model can be estimated using a

linear regression. Clewlow & Strickland (2000), method is simple yet robust means

of estimating the parameters of the model, we are therefore going to employed it.

The estimation procedures use discrete form of the continuous model:

Yt o 1 Yt t

By plotting Yt against Y t the main parameters o (intercept) and 


1 (slope) can

be from historical price change Yt ; using Excel function STDEV and annualised

assuming a 252 trading days, with the annualised standard deviation taking the

form:  252 .

Similarly excel function STEYX - returns the standard error of the predicted y-

value for each x in the regression, was also calculated. The remaining parameters

were calculated using the following equations:

 = speed =- slope

Log Run mean = ̂ =- Intercept/speed

Volatility =STEYS/log run mean

Half-life =ln2/speed

The path is simulated using Monte Carlo method in combination with variance

reduction techniques. This methodology involves using another path which is

negatively correlated with the first one; one additional equation will be needed:

51
X j =X j 1 + (
ˆX j1 ) t (j ) t 3.3

The average of the two paths is then calculated using the following relation:

X k 0.5*(X i + X j ). The following estimated parameters of , , :


ˆ0.04, 
ˆ447.054, 
ˆ0.0024 . The following equation was used to simulate

the path: X i X i1 .04(447.054 X i1 )t 0.0024 t .Where ~ N (0,1)

t is defined as T/N where T is the time and N the number of intervals. The path is

generated by generating randomly from a normal distribution.

1600.00

1400.00

1200.00

1000.00
simulate values
800.00
Original values
600.00
400.00

200.00

0.00
1 36 71 106 141 176 211 246 281 316 351

Figure 4: A comparison of a simulate path of a mean reversion model path

and the original path.

Figure 3.0 shows the simulated path and the original, clearly this are similar, thus

show that the original system price is mean reverting.

Simulation of Mean Reversion with jumps

The model to be used in simulation of the jumps is the Merton diffusion Model:

dS SdS Sdz kSdq

A discrete form of the model is:

52
X i X i 1 
ˆ(
ˆk X i 1 )t i t (
j )( i t )

Where:

k = average number of jumps

= standard deviation of the number of jumps

 = average number of jumps

 = number of jumps return/time of jumps

i a random variable from a normal distribution

Jump Diffusion

1600.00
1400.00
1200.00
System Price

1000.00
Jump_dif fusion
800.00
Non_jump path
600.00
400.00
200.00
0.00
1 46 91 136 181 226 271 316 361
Tim e

Figure 5: Jump diffusion model

3.1 Two factor model

The two-factor model was developed by Gibson and Schwartz. It consisted of

two models; with the first following Geometric Brownian Motion (GBM):

dS ( )Sdt 1 dz1 (1)

and the second is the instantaneous convenience yield model which is mean

reverting Orenstein-Uhlenbeck equation:

dS ()Sdt 2 dz2 (2)

53
where: dz 1 and dz 2 following a Browning Motion and correlated with correlation

coefficient .

If we let X =ln(S) equation (1) can be written as:

dS ( )Sdt 1 dz1

1
dX ( 12 )dt 1 dz1 (3)
2

Where:  is the dividend rate with risk adjustment factor r- , thus the electricity

price, this factor incorporate the storability of assert. Both equations can be

expressed as:

dS ( )Sdt 1 dz1

dS (() )Sdt 2 dz 2

With dz 1 *dz2 
dt

A discretised form of both equations can be written in the form:

1 2
X i1 X i (r i   ) t 1 t
2

Single Factor Model with Seasonality

A Schwartz model can be used to simulate prices by incorporating seasonal

volatility. A single model is of the form:

dX t 
ˆ(
ˆX t ) dt t dz

With z discretised form used by Clewlow and Strickland:

ˆ
X t X t 1 (X t 1 )t t t t

The time varying volatility is estimated using a 30-day standard deviation with the

54
sole aim of including the seasonal cycles.

Simple Moving Average (SMA)

A simple moving average technique involves adding a market price a of a n-period

and diving the result by n. SMA always lags behind the market, however it is

useful visual aid as well as determining market trend.

SMA can be used to indicate a buy or a sell signal. For example if the market price

falls below the SMA at time t gives a sell signal and the reversal indicates a sell

signal.

Trading Rules:

n
SMA n Pi / n ; where P t is the stock price at time t.
i1

Rate of return is defined as: R t X t t1 , where:

1 If Pt SMA t , and if RSI t 30% go short 



B t1 =  
- 1 If Pt SMA t and if RSI t 70% go long 

Exponential Smoothed Moving Average

The Exponential Moving Average (EMA) is a method of moving averaged

developed out of the short comings of the ordinary moving average (MA). EMA

incorporates older data often discarded by MA. The deletion of earliest trading data

according to Dahlquist (2007) leads to significant false signals, since information

on recent prices is often ignored. Thus EMA was developed to address these by

generating more buy and sell signals.

The formula for EMA is calculated as follows:

Smoothing factor (weight) = wt 2/(umber of days in moving average +1)

55
For example a 26-day EMA, the weight: w 26 =2/26+1 =7.4%

The general weight is calculated as follows: EMA t 100% wt

Similarly the formula for calculating the EMA for a particular t is:

EMAt wt * Pt EMA weight * MAt 1 , Where; P t = today’s closing price.

MAt1 =simple moving average of the previous day.

Double Exponential Moving Average (DEMA)

The Double Exponential Moving Average is a techniques initially developed by G

Mullory in 1994. Fot Staff applied this in the December (2007) of Futures &

Options Trader Journal. According to him DEMA is a “responsive indicator”,

which aims at reducing lag which is commonly found in all moving average

indicators such as SMA and EMA. DEMA formula incorporates a smoothing

constant or factor which provides an advantage of adding weight to the price. Thus

the formula is:

EMA t = w t *P t + (1-w t ) *EMA t 1

Where: 0<w t 1 and w t 2 / n 1 , with n representing the number of periods.

DEMA=2*EMA(x)-EMA(EMA(x) ).

According to Fot Staff:

“The DEMA calculates the difference between two components: a doubling of the

basic EMA value and “double smoothing” of the EMA value-an EMA smoothed by

a second EMA”

Fot Staff computed DEMA and EMA using two MACD systems. The main

difference between the two was the components. The first system used DEMA and

the second EMA, was used as the standard. A summary of the result of the two

56
systems is shown below:

Table 1: Source: http://www.futuresandoptionstrader.com

The net profit for the DEMA system appeared to be much better than the standard.

This trading system can be improved by an addition of Bollinger Bands with the

aim of including price volatility.

Moving Average Convergence – Divergence (MACD)

The Moving Average Convergence-Divergence (MACD) is widely used by

technical analyst as a trending tool. It is an application of the EMA methodology. It

uses initially a 12-day and 26-day EMA; however it can be adjusted for short or

longer periods. The 26-day EMA is subtracted from the 12-day EMA:

MACD t =EMA 12 EMA26

In addition to the MACD line there is another line often referred to as the signal

line, which is a nine-period EMA. MCAD oscillates about zero; if it is above the

signal line (SL) then a buy signal is generated; similarly if it is below zero a sell

signal is generated. In other words a buy signal is generated if the MACD line

crosses the signal line from below. Similarly a buy signal is generated if the

MACD crosses the signal line from above. However downward signals are not

57
always reliable therefore other measures such as RSI are used in conjunction with

the above strategy.

Trading Rules with MACD

Rate of return is defines as: R t X t t 1 , where

1 If MACD > 0, and SL t , long 



B t1 =  
- 1 If MACD < 0, and SL t short 

In addition to MACD line, there is one other line which represents the signal line- a

9-day period EMA. This is also plotted on the same chart. The difference between

these two lines is often depicted below the main charts. The histogram trend is

similar to the MACD line.

In situations where the MACD is above zero or below zero for the entire period;

an upward trend with buy signals should be followed. Buy signals are located

where the upward trend starts and the intersection of the EM 9 signal line-buy

signals are above this line. Downward trend in this situation are not profitable

therefore ignored.

Finally, the peaks and troughs as illustrated in the histogram can provide bits and

pieces of information regarding divergence analysis which identifies directional

changes over both long and short periods.

Bollinger Bands (BB)

A Bollinger is made up of two lines one above and one below a moving average

Line. Each line is calculated based upon a certain number of standard deviation

from the moving average. Market

Traders use 20-period simple moving average. Two standard deviation added to

plot the upper band similarly two standard deviation are subtracted to plot the

58
lower band. The bands are adjusted automatically with regard to the price changes

or volatility.

Trading Rules:

Rate of return is defines as: R t X t t 1 , where:

1 if P > SMA 20 + 2 and if RSI t < 30% Sell


 
B t1 =  
- 1 if P < SMA 20 + 2 if RSI t > 70% Buy 

Relative Strength Indicator (RSI)

RSI measures the strength of the market by removing erratic price movements.

Furthermore, it has the advantage of detecting market reversal. RSI scale range

from 0% to 100%.If 30 %< RSI< 70% indicates a warning signal. If RSI > 70%

indicate an overbought- provides an opportunity for buying of selling. If RSI <

30% indicate an oversold. A RSI value of 80% and 20% are preferred by market

practitioners. In addition if RSI >70% sell signal is indicated similarly if RSI<30%

a buy signal is generated. In other words RSI compares “up days” to “down days”

of a stock price and generates signals as a result. RSI is calculated as follows:

Up = (Sum of gain over N periods)/N

Down= (sum of losses over n periods)/N

RS=Up/Down

100
RSI =100-( ); with 0<RSI<100. Traders normally use a 14-day period.
1 RS

ADX Directional Movement Index

Introduction

ADX system was created by Wells Wilder Jr. It can be used as a stand alone

filter or in conjunction with a trend following system like moving average. The

59
Directional Movement Index (DMI) indicates trend and in addition measures it.

DMI is made up of three lines:

+DM measures an upward trend

 DM measures a downward trend

ADX signals the presence of a trend.

DMI is on a scale of 0-100, the higher the better for a potential tread.

ADX is used together with +DI and –DI. A signal is generated when +DI and

–DI crosses each other.

ADX Algorithm

Let Y h , Y l denotes previous day high and low prices similarly T h T l

denotes today’s high and low values respectively. Using these values +DM and

–DM are calculated as follows:

max(Th Yh ,0)ifT h Yh Yl Tl 



+DM =  

0, otherwise 

 max(Yl - Tl , 0) ifTh Yh Yl Tl 


-DM =  

0 otherwise 

Another measure is the True Range (TR) which is defined as follows:

TR=max (|T h –T l |, | T h - Y c |, |T l -Y c |)

Where Y c , is the previous day closing price.

The N-day Directional Indicators are defined as follows:

DM i
it N 1
+DI (N) t = t

TR
it N 1
i

60
t

DM i
it N 1
-DI (N) t = t

TR i
itN 
1

the N-day Directional Movement Index(DX) at time t is defines as follows:

DI ( N ) t (DI ( N ) t )
DX (N) t = * 100
DI ( N ) t ( DI ( N ) t )

The N-day Average Directional Movement Index (ADX) is similarly defined

at a time t as follows:

DX (N )
itN 
1
i
ADX (N) t =
N

Trading rules

Rate of return is defines as: R t X t t 1 , where:

1 if DI(N) t - DI(N) t-1 and DI(N) t DI ( N ) t 


and 
 
ADX (N ) t 20 
 
Buy 
B t1 =  
 
- 1 if DI(N) DI(N) and DI(N) DI ( N ) 
 t t-1 t t

and 
 
ADX (N ) t 20 

Sell 

Moving Average methodology

For the purpose of this research, a combination of two moving average-an

exponential (EMA) and a standard (MA) will be used to identify the trend and to

reduce lag thereby providing a robust trading system. Both of the moving average

has the advantage of smoothing out data with EMA being the more aggressive and

hence tracking price effectively, however both of them lag the data.

61
Moving Average

91.40
91.20
91.00 sell buy Prices
90.80
10-Day MA
90.60
90.40 10-Day EMa
90.20
90.00
0 5 10 15 20 25 30
Time

Figure 6: A Moving Average strategy

The system will be traded on the buy side only. A buy signal is generated if the

EMA cross the MA-the signal line and above the price curve. All other buy signals

are ignored until an exit signal is triggered; this is when the EMA line crosses

below the signal line. However, traders usually wait for confirmation from the

price movement and other indicators such as stochastic, relative index, before

entering the trade. This is a simplest trading strategy that can be executed by any

trader, and are found in almost all trading software systems.

Trading Rules for the EMA and MA:

Rate of return is defines as: R t X t t 1 , where:

1 if P > MA 10 & EMA10 Sell


 
B t1 =  

- 1 if P < MA 10 & EMA10 Buy 

62
Review of Performance measures

Introduction

In this section we list some of the performances measures statistics and the

information they captures. We are going to look at three main groups of measures:

Forecasting, Profit and Risk measures.

Forecasting Measures

There are three main ratios which are use in practice to compare the forecasting

abilities of the various models. The ratio measures the difference of the forecasted

values and the original observations. Theses ratios are:

Root Mean Square Error (RMSE)

1 N
RMSE= 
N t 1
( y t yˆ
t)
2

Mean Absolute Error (MAE)

1 N
MAE= ( y t yˆt ) 2 .
N t 1

Where n is the number of observations; y t the actual observation and yˆ


t is the

forecast value a t time t.

RMSE and MAE according to E-views 5 TM are used to compare the “forecast for

the same series across different models; the smaller the error, the better the

forecasting ability of that model according to that criterion”.

Mean Absolute Percentage Error (MAPE)

N t
1
MAPE=
N
y t y t / t
t N 1

63
Theil Inequality Coefficient (TIC)

1
yˆt y t 2 / h
N
TIC=
yˆ/ h  yˆ / h
2
t t

Where n is the number of observations; y t the actual observation and yˆ


t is the

forecast value a t time t.

MAPE and TIC are similarly, according to E-views 5; “lies between zero and one,

where zero indicates a perfect fit”. There three other ratio statistics: Bias

Proportion, Variance proportion and Covariance Proportion. For a good forecast

theses ratio should add up to one.

Profit Measures

Profit statistics are used to indicate the profitability of a strategy. Any rule based

strategy are meant to generate profit, therefore an appropriate rule must be put in

place to ensure profitability, which must stand the test of time. The following

statistics are used to measure the profitability of a strategy. The ratios will be used

in the spread sheet.

Total Pr ofit
Profit Factor (PF) =
TotalLoss

If PF>1, this indicates the system is making profit otherwise it is losing.

1. Percentage winning trades-usually between 30%-50%. For a winning

system this ration must be greater the 60%.

2. Annualized rate of return compare with a market benchmark

AverageTrade
3. Payoff ratio (PR) = . A PR greater the 5 is
AverageLos sin gTrade

preferable.

64
Netprofit
4. Efficiency Factor = a value between 38% and 69% is
Grossprofit

preferable.

Risk Measures

Risk is usually inherent in any trading strategy. Therefore any rule put in

place must be able to account for risk. This section looks at some of theses

measures.

1. The maximum cumulative drawdown- a correction from the maximum

curve, or a deviation of the equity curve from a peak and is a measure of

risk.

Netprofit
2. Net profit drawdown ratio = .
max imumDrawndown

n
Where Maximum drawn down = min [ Rt Max(rt ) ]
t
1

This ratio is also referred to as a recovery ratio and a value greater than two is

preferable.

Re turns
3. Sharp ratio=
SandardDeviation

4. Calmar Ratio (CR) – A measure of downside risk which provide an

added advantage over Sharpe ratio which measures both downward and

upside volatility.

Anuualised rate of retrun


CR=
| max imum drawdown |

5. Profit factor=gross profit/gross loss

Where gross profit is the profit from profitable trades and gross loss is the

total of all losses.

65
6. Percentage winning trades= Average of winning trade.

7. Cumulative profit=sum of daily profit.

8. Annualised volatility =Standard deviation of daily trading profit/square

root of number of trading days.

9. Maximum Daily profit =maximum of daily trading profit

10. Maximum daily loss= Minimum of daily trading profit.

11. Maximum drawdown Minimum of maximum drawdown.

Figure 7: An example of a summary of the results of a typical trading strategy.

Source: Futures & options Magazine

The Correlation filter

The aim of this method is to separate dynamic spread from static spread based on

the correlation of the spread. Dynamic spread corresponds to a decreasing

correlation whilst a static one corresponds to a decreasing correlation. Correlation

filter can be used instead of moving average technique.

66
Trading Rules:

Rate of return is defines as: R t X t Bt 1 , where:

N DRCt 1 , is the N-day rolling correlation, which is obtained by assign the

correlation of the two spreads and dividing by N.

1 if N - DRC t-1 < Spread t and RSI < 30% Sell 


Bt 1 =  
- 1 if N - DRC t-1 Spread t and RSI t > 70% Buy

Equity Curve

The Equity curve is a plot of cumulative profit over time. It provides a visual

representation of profit and losses over the testing period. A perfect equity line is a

straight line through the origin. Any deviation from this is usually due to

drawdown and pauses.

Equity Curve

50.00%
Account Balance

40.00%
30.00%
Account
20.00%
10.00%
0.00%
1 40 79 118 157 196 235 274 313 352
Time

Figure 8: Equity curve for a trading strategy showing few pauses.

67
Under Water Curve

Drawdowns

0%
-20%
-40%
% Drawdowns

-60%
-80%
Drawdown
-100%
-120%
-140%
-160%
-180%
Time

Figure 9: The diagram shows the drawdown of the output of a trading strategy.

The Under Water Curve is a plot of percentage drawdown. It provides a visual

representation of the draw downs over time. The percentage draw downs measure

the maximum percentage of decline of equity from the highest value. The smaller

the ratio, the better the system or strategy. A value less than 20% is effective in

minimizing risk. The output of a trading strategy show on the diagram gives a

summary of the draw downs, which shows several draw downs greater 20% with

most of them long-lived relatively.

Summary

Kirkpatrick, CD and others [2006] sums up what a good trading system is:

“In his book, Beyond Technical Analysis, Tushar Chande discusses the

characteristics of a good trading system. Chand’s Cardinal Rules for a trading

system are the following:

 Positive expectation-Greater than 13% annually

 Small number of robust trading rules-Less than ten each is best for

entry and exit rules

68
 Able to trade in multiple markets- Can use baskets fro determining

parameters, but rules should work across similar markets, different

stocks, different commodities futures , and so on

 Incorporates good risk control – Minimum risk as defied by

drawdown should not be mores than 20% and should not last more

than nine months

 Fully mechanical-No second-guessing during operation of the

system”.

The next section discusses results of the application of the out-of-sample prices to

the moving average trading strategy. The S&P500 price index will be compared

with Nord pool electricity prices. Since these indexes are similarly influenced by

factors such as demand and supply, changes in prices of commodities such as gas,

oil, carbon, short and long term weather forecast.

69
Chapter 4 Findings

Introduction

This section summary the results of the forecasting tools, and the application to the

moving average strategy. On the examination of the electricity prices, it was observed

that they do indeed exhibit volatility clustering. This was confirmed by Jarque-Bera

statistics, which rejected normality. In addition the quintiles distribution showed a

few outliers, a further confirmation of the departures from normality. Thus the data

fat tailed which called for a model which would incorporate this property. Finally, the

results of the GARCH test showed that all the terms including the asymmetric ones

proved significant. The coefficient of the asymmetric terms was negative, thus

pointing to a high variance in the following period. In addition the sum of the

coefficients was less that one, which provided the evidence of the persistence of the

conditional variance.

Results of forecasting performance

Each of the forecasting tools was tested with regard to their forecasting ability using

RMSE. The one with the least RMSE was then chosen. The moving average trading

strategy is then applied using the out-of-sample from the best forecasting tool. The

performance of the forecasting model models is compared using the out-of-sample

period, the reason being that market participants use a day–ahead forecast which

consist of 24 hour auctions, where orders are bid until demand are met. Out-of–

sample periods are chosen with or without regards to jumps.

Ordinary least square forecast was found to have the least RMSE with Naïve method

the worse; Table 4.1 shows a summary of the results. Hence it was made the model of

choices for the trading strategy. Table 4.2 similarly shows the ratios of the trading
strategy.

Table 2: A comparison of forecasting tools

Tool RMSE Rank

Naïve 45.0633 6

GARCH-M(dynamic) 4.854295 3

Box-Jenkins 7.971633 4

VECM(dynamic) 20.418411 5

VECM(static) 4.631233 2

OLSQ 2.5788 1

Table 3: Summary of ratios of the trading strategy-OLSQ applied to CDF

Ratios

Cumulative Profit -11201.9% Number of Periods Daily returns Rise 52.00

Annualised Volatility 16740.0% Number of Periods Daily returns Fall 30.00

Max. Daily Profit 8.4 Number of Winning Up Periods 35

Max. Daily Loss -7.8 Number of Winning Down Periods 12

Maximum Drawdown -17.0 % Winning Up Periods 67%

% Winning Trades 18.2% % Winning Down Periods 23%

Number of Up Periods 35 Correct Directional Change 0.24

Number of Down Periods 35 Average Gain in Up Periods 86.69

Number of Transactions 23 Average Loss in Down Periods -83.3

71
A summary of modelling process of ARMA time series

A graph of the time series of the S&P500 suggested the presence of cycles and

some trends. This was however transformed into a stationary one using logarithms

and first differences. ADF test on the data confirmed stationarity. Similarly, the

sample correlation and partial sample autocorrelation suggested an ARMA (p, q)

model. Model estimation was carried out, with the result indicating significance of

the model parameters. An ARMA (6, 6) model was then used to produce an out-of-

sample forecast. The mean square error of forecast indicated that, the static forecast

was better than the dynamic forecast, since it provided the smallest MSE. In the

next section the forecasted out-of-sample data will be used as an input for the

moving average trading strategy.

Result of the trading strategy using S&P500

The trading strategy was applied to the S&P500 index; without the use of any

forecasting tool. These section summaries this result with the aim of using this to

compare the output from other data where academic tools have been applied to,

different data from different environment. The equity curve shows some losses

initially but was increase in a passage of time with a steady trend. There was

however reversal which was due to drawdown’s. The trade system had a low

cumulative profit of 44.7%. This provides an indication of a good trading system,

since it is far greater the 13%-the standard. There were a lot of drawdowns which

was lower than -20% recommended for a robust system. In particular there were

several drawdown’s in excess of -100%, but were short-lived. The maximum

drawdown reported by the system was about -2.5% which is less than 20%

recommended. From the statistical ratios the trading system befitted from both
direction, however the system could minimize loss due to potential reversals if

protection offered in the form of protective stops.

Table 4: Summary of Trading Strategy Analysis of S&P500 Index (without

forecasting tool)

Ratios

Cumulative Profit 19.4% Number of Periods Daily returns Rise 32

Annualised Volatility 4.6% Number of Periods Daily returns Fall 45

Max. Daily Profit 0.9% Number of Winning Up Periods 62

Max. Daily Loss -0.5% Number of Winning Down Periods 36

Maximum Drawdown -0.5% % Winning Up Periods 193.75%

% Winning Trades 71.4% % Winning Down Periods 113%

Number of Up Periods 33 Correct Directional Change 12.25%

Number of Down Periods 36 Average Gain in Up Periods 20.95%

Number of Transactions 8 Average Loss in Down Periods -1.55%

73
Table 5: Summary of ratios of the result of S&P500 trading strategy with ARMA

Ratios

Cumulative Profit 16.4% Number of Periods Daily returns Rise 34.00

Annualised Volatility 13.9% Number of Periods Daily returns Fall 44.00

Max. Daily Profit 1.9% Number of Winning Up Periods 23

Max. Daily Loss -2.5% Number of Winning Down Periods 28

Maximum Drawdown -7.3% % Winning Up Periods 68%

% Winning Trades 27.8% % Winning Down Periods 82%

Number of Up Periods 37 Correct Directional Change 0.18

Number of Down Periods 34 Average Gain in Up Periods 34.03%

Number of Transactions 19 Average Loss in Down Periods -17.6%

74
Summary of Trading Strategy Analysis of Pairs trading

Table 6: Summary ratios of pairs trading strategy

Ratios

Cumulative Profit 0.1% Number of Periods Daily returns Rise 110

Annualised Volatility 297590.7% Number of Periods Daily returns Fall 143.00

Max. Daily Profit 29.0% Number of Winning Up Periods 105

Max. Daily Loss -34.2% Number of Winning Down Periods 0

Maximum Drawdown -65.5% % Winning Up Periods 95.45%

% Winning Trades 5.7% % Winning Down Periods 0%

Number of Up Periods 36 Correct Directional Change 11.86%

Number of Down Periods 34 Average Gain in Up Periods 68.63%

Number of Transactions 24 Average Loss in Down Periods -42.08%

75
System price and MACD

Table 7: Summary of the ratios of MACD trading strategy applied to yearly System

index

Ratios

Cumulative Profit 2.0% Number of Periods Daily returns Rise 100

Annualised Volatility 125842.4% Number of Periods Daily returns Fall 148.00

Max. Daily Profit 20.7% Number of Winning Up Periods 94

Max. Daily Loss -13.9% Number of Winning Down Periods 0

Maximum Drawdown -34.6% % Winning Up Periods 94.00%

% Winning Trades 2.9% % Winning Down Periods 0%

Number of Up Periods 33 Correct Directional Change 16.21%

Number of Down Periods 37 Average Gain in Up Periods 234.81%

Number of Transactions 16 Average Loss in Down Periods -48.25%

76
Table 8: Summary of the ratios of MACD trading strategy applied to yearly

S&P500 index

Ratios

Cumulative Profit 2.8% Number of Periods Daily returns Rise 108

Annualised Volatility -48918.7% Number of Periods Daily returns Fall 145.00

Max. Daily Profit 88.9% Number of Winning Up Periods 102

Max. Daily Loss -120.4% Number of Winning Down Periods 0

Maximum Drawdown -356.6% % Winning Up Periods 94.44%

% Winning Trades 5.7% % Winning Down Periods 0%

Number of Up Periods 34 Correct Directional Change 11.46%

Number of Down Periods 36 Average Gain in Up Periods 1424.14%

Number of Transactions 21 Average Loss in Down Periods -867.68%

77
Chapter 5 Conclusion

In this dissertation the forecasting ability of a series of forecasting tools was tested

with the aim of finding their use in a trading environment. The returns from

forecasting using out-of-sample data was applied to a moving average trading

strategy (MACD) the various forecasting, profitability and other ratios was then

generated. The CDF series was modelled using ordinary least square, the out of

sample forecast was then simulated by the MACD strategy. Similarly S&P500

prices were modelled by ARMA and the out–of-sample was simulated using the

MACD. The trading strategy was simulated using S&P500 price without the use of

any tool. Finally, the usefulness of pairs trading concepts was tested using MACD.

A summary of the ratios generated by trading strategy is shown in the Table 5.1

below.

Table 9: Summary ratios of the trading strategy- as applied to three different

data set.

Ratios Nord CDF with OLSQ S&P500 S&P500 ARMA

Cumulative Profit -11201.9% -13095.0% 16.4%

Annualised Volatility 16740.0% -9230.7% 13.9%

Max. Daily Profit 8.4 0.9% 1.9%

Max. Daily Loss -7.8 -0.5% -2.5%

Maximum Drawdown -17.0 -0.5% -7.3%

% Winning Trades 18.2% 71.4% 27.8%

Number of Up Periods 35 33 37

78
Number of Down Periods 35 36 34

Number of Transactions 23 8 19

Number of Periods Daily returns Rise 52.00 32 34.00

Number of Periods Daily returns Fall 30.00 45 44.00

Number of Winning Up Periods 35 62 23

Number of Winning Down Periods 12 36 28

% Winning Up Periods 67% 193.75% 68%

% Winning Down Periods 23% 113% 82%

Correct Directional Change 0.24 12.25% 0.18

Average Gain in Up Periods 86.69 20.95% 34.03%

Average Loss in Down Periods -83.3 -1.55% -17.6%

Maximum drawdown is a measure risk that needs to be taken into consideration

before the strategy starts making profit. For the purpose of comparison we assume a

maximum drawdown of 20% or less, any other ratio represents a risk to the trader.

The table suggest all the drawdown ratios are reasonable, since they are all less than

20% the limit. However the best drawdown has a value of -0.17%; this was from

the OLSQ technique. The worse was from the trading pair’s strategy which gave a

value of -65.5% The best maximum daily profit was 29%, which was from the pairs

trading strategy, followed by 20.7% when the MACD was applied to the System

price. The two ratios were more than the minimum of 16% suggested by analyst.

The OLSQ ratio was 8.4%, with the rest generating lesser ratios. Thus this result

suggested that the more risky a strategy is, the better the returns.

The Correct Direction Index (CDI) is an additional tool used in conjunction with

79
RMSE to measure the direction ability of the forecasted variable; a positive CDI

indicates an increase price value. All the types produce positive values, with OLSQ

generating the biggest ratio of 0.24; the pairs trading strategy produce a least value

of 0.1186.

We conclude based on the above a ratio that OLSQ is the best forecasting tool for

electricity prices and can be used in all markets. However the ratios used were

calculated from past data, with different sample size. The applicability of these

ratios in a trading environment cannot be relied on, since it did not take into

consideration the dynamic nature of the trading environment. In addition

transactions cost was not taken into consideration. The losses generated by the

strategies were very huge, thus any future research in this area should take into

consideration the risk premium.

Thus in spite of the above weakness, “academic tools” are relevant in any trading

environment and this cannot be over emphasised. This conclusion seems to agree

with the finding of Jason et, al (2004). The simulation of the CDF process using

Ornstein-Uhlenbeck method; also provided evidence of the use of stochastic

process in the modelling and pricing of CDF similarity which, which suggest its

consideration for future research.

Future Research

This dissertation was meant to point out some of the weakness of previous studies;

however there were some limitations which future research could consider. For

example the, both the short side and the long of the trades can be considered in

future since only the long side was used.

The usefulness of this study can be more relevant if the number of portfolios is

80
increased, this is in regard to the pairs trading strategy, which is very relevant to

today’s live trading environment. Similarly the period for the moving average can

be varied and adapted to different trading styles, such as weighted moving average.

Data from other major exchanges in the rest of Europe, Japan could also be

considered in the future, since the price data behaviours varies for instance,

currencies data behaves differently from commodities such as oil and gas.

The reliability of the conclusions drawn could question on the basis of the one year

data which was used. It is a well known fact that long term data captures all kinds of

important information such as interest rate cuts, business cycles which is relevant to

both short and long term traders.

81
References

ANSIL, M. ISLAM, SYED & AHMED M (1999): The Purchasing Power Parity

Relationship: Causality and Cointegration Test Using Korea-U.S Exchange Rate

and Prices. Journal of Economic Development Vol. 24(2)

BATCHELOR, R. (2007): Forecasting Spot and Forward prices in the

International Freight Market. International Journal of Forecasting Vol. 23, pp101-

114

BROCK, B, LAKONISHOK, J, LeBARON, B (1992): Simple Technical Trading

Rules and the Stochastic Properties of Stock Returns. The Journal of Finance, Vol.

47, No. 5 (Dec., 1992), pp. 1731-1764

BROOKS, C, ALISTAIR, G, REW & RITSON, S (2001): A trading strategy based

on the lead-lag relationship between the spot index and futures contract for the

FTSE 100. International Journal of Forecasting17, 31-44

BROOKS, C & BURKE, SP (2003): Information criteria for GARH model

selection. The European Journal of Finance 9, 557-580.

BYSTRON, HNE (2003): The hedging performance of electricity futures on the

Nordic exchange. Allied Econometrics Vol. 35, pp. 1-11

CHAN IL, KIMBERLY, C, GLEASON, MUTHAR I(2001): Trading rules in Latin

American currency spot rates. International Review of Financial Analysis

10(2001)135-156

CHOWDHURY, AR (1991): Futures Market Efficiency: Evidence from

Cointegration Tests. The Journal of Futures Markets, Vol. 11 No. 5, 577-589.

CIRA, P & MARILENA S (2008): Mathematical and Statistical Methods in

82
Insurance and Finance. Springer-Verlag Italia

CLEWLOW, L & STRICKLAND, C (2000): Energy Derivatives: Pricing and Risk

Management. Lcima Publications

DERK, JS & CHRITOP W (2007): Extended ARMA models for estimating price

development on day-ahead electricity markets. Electric Power Systems Research

Vol. 77, pp. 883-893

DUNIS, C, LAWSON, J, & NA 


I
MP (2003): Applied quantitative methods for

trading and investments. John Wiley & Sons Ltd

EVANS, B, LAWSON, C, DUNNIS, and LAWS, J (2004): Trading Futures

Spreads: An Application of correlation. CIBEF

LAWSON,C, DUNIS & MORRISION, V(2004): The Economic Value of

Advanced Time Series methods for modelling and trading 10-year Government

Bonds. Liverpool Business School (CIBEF)

LAWSON, C. D. & Mark, W (2003): Application of Advances Regression for

trading and Investment (CIBEF)

LAWSON, C, DUNIS & MORRISION, V (2004): The Economic Value of

Advanced Time Series methods for modelling and trading 10-year Government

Bonds. Liverpool Business School (CIBEF)

LINDEMAN, LAWSON, C. DUNIS, LISBOA, PL (2004): Probability

Distribution and Leveraged Trading Strategies: An application of Gaussian

Mixtures Models to the Morgan Stanley High Technology 35 Index. Liverpool

Business School (CIBEF)

MARSHALL BR, JARED M. CAHAN B, CAHAN, RH (2006): Is the CRISMA

83
technical trading system profitable? Global Finance Journal 17 (2006) 271–281.

KIRKPATRICK, KD & DHALQUIST JR (2006): Technical Analysis. Pearson

Education Inc.

SONDERMANN, D (2006): Introduction to Stochastic Calculus for Finance.

Springer-Verlag, Berlin Heidelberg.

HSIAO-TIEN (2007): Forecasting electricity market using artificial neural

networks. Energy Conversion & Management Vol. 48, pp.907-912

ICEBLING, K, & MILT, B (1966): Designing a neural network for forecasting

financial and economic series. Neural Computing Vol. 10, pp.215-236

LAWS, J & THOMSON J (2004): The efficiency of financial futures markets: Tests

of prediction accuracy. European Journal of Operational Research 155, 284-298.

LAMBERTON-LAPEYRE (1995): Introduction to Stochastic Calculus, applied to

Finance. Chapman and Hall, London.

LESTER, H (2006): A TARCH examination of return volatility-volume relationship

in electricity futures. Applied Financial Economics Vol.12, pp. 893-901

MAGEE J, EDWARDS RD &BASSETTI WHC (2007): Technical Analysis of

Stocks Trends. CRC Press

PRING M J (2002): Technical Analysis Explained. McGraw-Hill

PAUL, BG & ALBERT S.P (1999): Risk Arbitrage Opportunities in Petroleum

Futures Spreads. The Journal of Futures Market Vol. 19, pp 931-955

SASCHA W & WIMSCHULTE J (20007): The Pricing of Electricity Futures:

Evidence From European Energy Exchange. The Journal of Futures Markets, Vol.

27, No. 4 387-410

84
SIBHES, CB & TIEN L (2004): Short-term electric load forecasting an artificial

neural network: a case of Northern Vietnam. International Energy Research Vol.

28, pp. 463-472

WING-SHING, L, & TAI-LEUNG TC (2006): Profitability of Directional

Indicators. Allied Financial Economics Letters, Vol. 2, 401-406

Leading Methods for Forecasting Energy Demand. Available from:

http://www.newenergyassoc.com/documents/forecasting_wp.pdf

[Accessed: 27 June, 2007]

Nord Pool. Available from:http://www.nordpool.no/nordpool/financial/index.html

[Accessed: 10 June, 2007]

Centre for International Banking, Economics and Finance (CIEBF).

Available from: http://cwis.livjm.ac.uk/AFE/CIBEF/67756.htm

[Accessed: 10 November, 2007]

Futures & Options trading magazine.

Available from: http://www.futuresandoptionstrader.com/huj32/FOT1007.pdf

[Accessed: 10 November, 2007]

85
APPENDIX 1

Cointegration Johansen methodology

If a set of variables are co-integrated, then there is an error correction model

(ECM) that links all the variables. In other words if there is a long term relationship

between two co-integration variables, it is possible to find a linear combination of

theses variable in a vector representation format. Engle and Grange [1987] method

is often applied but by far the most superior is the method used by Johansen [1988]

is going to be employed in the study. Johansen used a maximum likelihood

approach using vector auto regression representation (VAR) which is specified in

the following form:

X t = i X t 1 ... k X t k i (1)

Where, X i is a p-dimensional vector of price; with k lags and i an independent

identical distributed random variable, with mean zero and variance 1.

For p=3, equation (1) becomes:

X i 1 X t 1 2 X t 2 t

(2)

2
X i i X t i t
i1

(3)

The tow variables are chosen because of the two spread variables. According to

Johansson if, the X t variable is non-stationary, it is possible to find a variable X t

which is stationary. As a result of this, equation (1) can be transformed into its

86
EMT by using the L operator defined by: 1 L

k1
ΔXt Γi ΔXt i Xt k μεt (4)
i1

Where: Γi 
(1 Π1 ... Πi ), I=1... K-1,

Is n x n coefficients matrix which includes information about the variable Y t

Φ 
(I Π1 .. Πk ) i=1 ... k-1 (5)

Both X t and ik tends to zero in the long run according to Johansson.

The matrix is an nxp and can be expressed as:

'
Παβ̂, with both αand β̂ representing a p x r matrices with αrepresenting

the speed of adjustment of the coefficients. If the variables are co-integrated then

there exists a stationary relation of the form β̂


X t even though Xt is non stationary.

The rank of Π is r < p depending on the number of co-integrating vectors. The

rank determines the number of co-integrating equations. A rank of two suggests that

the variables are jointly stationary. The hypothesis of interest is:

H o (r) : Πα
βˆ

If the hypothesis is significant, a rank r is found then there is likely to be a linear

combination of the stationary variables. The α vector is the speed of adjustment

with β̂ representing the co-integrating vector. The estimation of the parameters of

Π is determined by regressing X t and X t k on lagged difference -

X t1 ,..., X t k 1 . Johansson formed a product of residual of the

form: S ij T1 ij


jt , i, j =0, k

The estimates of βis found by equating the determinant of the Eigen value to zero:

87
S kk Sko S 1 Si0 | =0

Johansson normalised this equation and applied the maximum likelihood principle

to obtain a likelihood function or a likelihood ration test statistics of the form:

p
trace 2 ln(Q) T ln(1 
ˆ
i)
i r 1

ˆ
With T, the sample size and  ˆ ˆ
1 ,2 ,..., p representing the Eigen values.

This statistics is a chi-squared distribution with N-r degree of freedom, where N is

the number of variables and r the rank of the matrix Π. The null hypothesis of at

most r co-integrating vector is tested against an alternative of r grater than r.

Similarly the maximum likelihood Eigen value for the null hypothesis with exactly r

co-integrating equations against r+1 is:

ˆ
max T ln(1  ˆ
r 1 ) , with 
ˆ
1 ... 
r p

APPENDIX 2

90
80
70

60
50
40
30
20

10
0
50 100 150 200 250 300 350 400

CDF

Figure 10: A daily data of range price series or the CDF

88
350

300

250

200

150

100

50
50 100 150 200 250 300 350 400

SYSTEM HELSINKI

Figure 11: A data comparison of System and Helsinki daily prices

100
Series: CDF
Sample 1 400
80 Observati ons 370

Mean 13.42475
60 Median 11.46933
Maximum 86.13453
Minimum 3.334269
40 Std. Dev. 7.355829
Skewness 3.535252
Kurtosis 28.96757
20
Jarque-Bera 11166.39
Probability 0.000000
0
12.5 25.0 37.5 50.0 62.5 75.0 87.5

Figure 12: A summary statistics of the CDF

60

40

20

-20

-40

-60

-80
50 100 150 200 250 300 350 400

DCDF

89
Figure 13: A log transformation of the range series.

160
Series: DCDF
140 Sample 1 400
Observations 369
120
Mean -0.005930
100 Median 0.032009
80 Maximum 55.67898
Minimum -68.52856
60 Std. Dev. 6.925025
Skewness -1.331122
40 Kurtosis 39.50347

20 Jarque-Bera 20596.20
Probability 0.000000
0
-60 -40 -20 0 20 40 60

Figure 14: A summary statistics of log transformation of the CDF series.

Table 10: A summary of ADF test

Null Hypothesis: D(DCDF) has a unit root

Exogenous: Constant

Lag Length: 7 (Automatic based on SIC, MAXLAG=16)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -15.42240 0.0000

Test critical values: 1% level -3.448312

5% level -2.869351

10%

level -2.570999

*MacKinnon (1996) one-sided p-values.

90
Augmented Dickey-Fuller Test Equation

Dependent Variable: D(DCDF,2)

Method: Least Squares

Date: 02/18/08 Time: 14:00

Sample (adjusted): 12 371

Included observations: 360 after adjustments

Coefficie

Variable nt Std. Error t-Statistic Prob.

9.89547

D(DCDF(-1)) 5 0.641630 -15.42240 0.0000

7.58449

D(DCDF(-1),2) 8 0.608208 12.47024 0.0000

5.99664

D(DCDF(-2),2) 6 0.540608 11.09242 0.0000

4.39986

D(DCDF(-3),2) 6 0.448245 9.815769 0.0000

2.87512

D(DCDF(-4),2) 9 0.339461 8.469694 0.0000

1.65367

D(DCDF(-5),2) 8 0.228883 7.224993 0.0000

0.68970

D(DCDF(-6),2) 7 0.129044 5.344728 0.0000

91
0.21152

D(DCDF(-7),2) 5 0.052174 4.054205 0.0001

0.01435

C 5 0.339779 0.042247 0.9663

0.89302 0.02534

R-squared 6 Mean dependent var 5

Adjusted R- 0.89058 19.4900

squared 8 S.D. dependent var 0

6.44679 6.58972

S.E. of regression 5 Akaike info criterion 5

14587.9 6.68687

Sum squared resid 7 Schwarz criterion 8

1177.15 366.273

Log likelihood 1 F-statistic 0

2.02061 0.00000

Durbin-Watson stat 3 Prob(F-statistic) 0

Table 11: A summary of Philips-Peron test

Null Hypothesis: D(CDF) has a unit root

Exogenous: Constant

92
Bandwidth: 132 (Newey-West using Bartlett kernel)

Adj. t-Stat Prob.*

Phillips-Perron test statistic -102.2090 0.0001

Test critical values: 1% level -3.447914

5% level -2.869176

10%

level -2.570905

*MacKinnon (1996) one-sided p-values.

44.2693

Residual variance (no correction) 0

1.14383

HAC corrected variance (Bartlett kernel) 3

Phillips-Perron Test Equation

Dependent Variable: D(CDF,2)

Method: Least Squares

Date: 02/18/08 Time: 14:05

Sample (adjusted): 4 371

Included observations: 368 after adjustments

Variable Coefficie Std. Error t-Statistic Prob.

93
nt

1.27629

D(CDF(-1)) 4 0.050231 -25.40849 0.0000

0.00092

C 2 0.347786 -0.002650 0.9979

0.63819 0.01560

R-squared 4 Mean dependent var 4

Adjusted R- 0.63720 11.0765

squared 5 S.D. dependent var 4

6.67167 6.63903

S.E. of regression 2 Akaike info criterion 8

16291.1 6.66027

Sum squared resid 0 Schwarz criterion 8

1219.58 645.591

Log likelihood 3 F-statistic 3

2.16825 0.00000

Durbin-Watson stat 8 Prob(F-statistic) 0

94
Table 12: A Correlogram of CDF series

Date: 02/18/08 Time: 14:11

Sample: 1 400

Included observations: 368

Partial Q-

Autocorrelation Correlation AC PAC Stat Prob

- -

****|. | ****|. | 1 0.528 0.528 103.49 0.000

- -

.|. | ***|. | 2 0.045 0.450 104.26 0.000

.|* | ***|. | 3 0.085 0.326 106.96 0.000

95
- -

*|. | ***|. | 4 0.060 0.363 108.32 0.000

.|* | **|. | 5 0.116 0.194 113.39 0.000

- -

*|. | ***|. | 6 0.179 0.402 125.47 0.000

.|** | **|. | 7 0.215 0.202 142.84 0.000

- -

*|. | **|. | 8 0.110 0.203 147.42 0.000

.|. | .|. | 9 0.021 0.043 147.59 0.000

- -

*|. | *|. | 10 0.071 0.166 149.53 0.000

.|* | .|. | 11 0.102 0.011 153.46 0.000

- -

.|. | .|. | 12 0.047 0.048 154.32 0.000

- -

*|. | *|. | 13 0.071 0.146 156.23 0.000

.|* | *|. | 14 0.150 0.137 164.83 0.000

- -

*|. | *|. | 15 0.080 0.107 167.29 0.000

.|. | *|. | 16 0.007 0.129 167.31 0.000

96
- -

.|. | *|. | 17 0.021 0.085 167.48 0.000

.|. | .|. | 18 0.045 0.007 168.26 0.000

.|. | .|* | 19 0.019 0.083 168.39 0.000

- -

*|. | .|. | 20 0.080 0.017 170.88 0.000

.|* | *|. | 21 0.074 0.137 173.02 0.000

.|. | *|. | 22 0.061 0.114 174.48 0.000

- -

.|. | .|. | 23 0.032 0.021 174.89 0.000

- -

*|. | *|. | 24 0.102 0.091 179.00 0.000

.|* | .|. | 25 0.101 0.026 183.06 0.000

.|. | .|. | 26 0.025 0.012 183.30 0.000

- -

*|. | *|. | 27 0.072 0.097 185.39 0.000

.|* | .|. | 28 0.152 0.055 194.60 0.000

*|. | .|. | 29 0.105 0.014 199.06 0.000

97
-

.|. | .|. | 30 0.025 0.006 199.32 0.000

.|. | .|. | 31 0.011 0.005 199.37 0.000

.|. | .|. | 32 0.009 0.048 199.40 0.000

.|. | .|. | 33 0.009 0.030 199.43 0.000

- -

*|. | .|. | 34 0.071 0.040 201.47 0.000

.|* | .|. | 35 0.109 0.011 206.32 0.000

.|. | .|. | 36 0.030 0.022 206.68 0.000

The oretical Quantile-Quantile


8

4
Normal Quantile

-4

-8

-12
-80 -60 -40 -20 0 20 40 60
DCDF

98
Figure 15: Quantile-Quantile plot of the log-returns

1.2

0.8

0.4

0.0

-0.4

-0.8

-1.2
25 50 75 100 125 150 175 200 225 250

LHELSINKI

Figure 16: A log transformation of Helsinki prices

.4

.3
.2

.1
.0

-.1

-.2
-.3

-.4
-.5
25 50 75 100 125 150 175 200 225 250

LSYSTEM

Figure 17: Log transformation of system prices

99
1240

1200

1160

1120

1080

1040
25 50 75 100 125 150 175 200 225 250

CL OSE

Figure 18: S&P500 index.

LOG RETURNS

0.40
0.30
0.20
0.10
logs

0.00
-0.10 0 50 100 150 200 250 300 350 400

-0.20
-0.30
time

Figure 19: Logarithms returns of the System price

100
Jumps

0.40
0.30
0.20
R e tru n s

0.10
Jumps
0.00
-0.10 0 50 100 150 200 250 300 350 400
-0.20
-0.30
Time

Figure 20: A plot of number of jumps against returns

Volatity & Returns

0.40

0.30
0.20
R e tu rn s

0.10

0.00
0 100 200 300 400
-0.10
-0.20

-0.30 Log Returns


Time 30-day rolling Volatility

Figure 21: A combinations of 30-day volatility and logarithms returns of the

System price

APPENDIX 3

101
Maximum Drawdown

0%
-500%
-1000%
Drawdown

-1500%
Maximum Drawdown
-2000%
-2500%
-3000%
-3500%
Time

Figure 22: Maximum Drawdown of S&P500 without forecasting

.015

.010

.005

.000

-.005

-.010

-.015

-.020
10 20 30 40 50 60 70

RETURNS
9-Day EMA
20-Day MA

Figure 23: Returns and 9-Day EMA and 20-Day MA series

102
.00

-.01

-.02

-.03

-.04

-.05

-.06

-.07

-.08
10 20 30 40 50 60 70

Maxi mum Drawdown

Figure 24: Maximum drawdown for S&P500 with ARMA forecast

.02

.01

.00

-.01

-.02

-.03
10 20 30 40 50 60 70

Daily Trading Profit

Figure 25: Daily profit for S&P500 with ARMA forecast

103
.25

.20

.15

.10

.05

.00

-.05
10 20 30 40 50 60 70

Cummulative Profit

Figure 26: Daily cumulative profit for S&P500 with ARMA forecast

1240

1200

1160

1120

1080

1040
2004M04 2004M07 2004M10

MA EMA Adj Close

Figure 27: Comparing price, moving average, exponential moving average

104
Equity Cur ve (%)

50.0%
45.0%
40.0%
Account Balance

35.0%
30.0%
25.0%
20.0%
15.0%
10.0%
%

5.0%
0.0%
-5.0% 1 25 49 73 97 121 145 169 193 217
Tm e

Figure 28: Equity curve of S&P500 without forecasting

Forecasting with Ordinary Least Square as applied to power data

(CDF)

15
Forecast:LSYSTEMF_OLSQ
10 Actual: LSYSTEM
Forecastsample: 290 400
5 Adjusted sample: 290 371
Included observations: 82
0
Root Mean Squared Error 2.204098
Mean Absolute Error 1.602345
-5
Mean Abs. Percent Error 270.1602
Theil Inequality Coefficient 0.355895
-10
Bias Proportion 0.000113
Variance Proportion 0.030467
-15
Covariance Proportion 0.969420

-20
290 300 310 320 330 340 350 360 370

LSYSTEMF_OLSQ

105
0

-4

-8

-12

-16

-20
10 20 30 40 50 60 70 80

Maximum Drawdown

Figure 29: Graph of Maximum Drawdown of OLSQ applied to CDF

40

30

20

10

-10

-20

-30
10 20 30 40 50 60 70 80

C umulative Profi t(%)

Figure 30: Graph of Equity curve

106
30-day volatily & returns

0.4
log returns
0.2
Returns

0 30-day
-0.2 0 100 200 300 400 volatility
-0.4
time

Figure 31: 30-day volatility

.000

-.001

-.002

-.003

-.004

-.005

-.006
2004M09 2004M10 2004M11 2004M12

Maximum D rawdown

Figure 32: Graph of maximum drawdown

107
Equity Curve

0.018%
0.016%
0.014%
0.012%
% P rofit

0.010%
Profit
0.008%
0.006%
0.004%
0.002%
0.000%
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65
Time

Figure 33: Graph of cumulative profit

Append Public Function MV(X, a, u, t, sigma, nsim)

Dim rnmut, sigt, Sum, S1, S2

Dim i As Integer

Randomize

rnmut = X + a * (u - X) * t

sigt = sigma * Sqr (t)

For i = 0 To nsim

rands = Rnd + Rnd + Rnd + Rnd + Rnd + Rnd + Rnd + Rnd + Rnd + Rnd + Rnd +

Rnd - 6

S1 = rnmut + rands * sigt

108
S2 = rnmut - rands * sigt

Next i

MV = 0.5 * (S1 + S2)

End Function

APPENDIX 4

350

300

250

200

150

100

50

0
250 275 300 325 350 375 400

ERROR_S QUARE

Figure 34: Error Square

Table 13: Forecasting with OLSQ

Dependent Variable: LSYSTEM

Method: Least Squares

109
Date: 03/06/08 Time: 11:17

Sample (adjusted): 2 250

Included observations: 249 after adjustments

Coefficie

Variable nt Std. Error t-Statistic Prob.

0.42508

LHELSINKI 7 0.023186 18.33387 0.0000

5.79763

C 1 0.427433 13.56383 0.0000

0.57642

R-squared 4 Mean dependent var -0.059474

Adjusted R- 0.57471

squared 0 S.D. dependent var 6.871133

4.48096

S.E. of regression 1 Akaike info criterion 5.845552

4959.51

Sum squared resid 6 Schwarz criterion 5.873804

725.771

Log likelihood 2 F-statistic 336.1309

1.24412

Durbin-Watson stat 6 Prob(F-statistic) 0.000000

110
Figure A summary of output of ordinary least square estimation of the model

30
Forecast: LSYSTEMF
Actual: LSYSTEM
20 Forecast sample: 251 400
Adjusted sample: 251 371
Included observations: 121
10
Root Mean Squared Error 2.578828
Mean Absolute Error 1.789171
0 Mean Abs. Percent Error 235.8585
Theil Inequality Coefficient 0.324692
Bias Proportion 0.004656
-10 Variance Proportion 0.257563
Covariance Proportion 0.737781

-20
275 300 325 350

LSYSTEMF

Figure 35: Result of forecast using OLSQ

Forecasting with ARIMA model

40

30

30 20

20 10

10 0

-10
250 275 300 325 350

Resi dual Actual F itted

111
Figure 36:

Table 14: Unit root test

Null Hypothesis: RESID06 has a unit root

Exogenous: Constant

Lag Length: 1 (Automatic based on SIC, MAXLAG=15)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -7.287126 0.0000

Test critical values: 1% level -3.456730

5% level -2.873045

10%

level -2.572976

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(RESID06)

Method: Least Squares

Date: 03/06/08 Time: 11:29

Sample (adjusted): 4 250

Included observations: 247 after adjustments

Variable Coefficie Std. Error t-Statistic Prob.

112
nt

0.51206

RESID06(-1) 5 0.070270 -7.287126 0.0000

0.17651

D(RESID06(-1)) 3 0.062993 -2.802108 0.0055

0.00416

C 6 0.261331 -0.015940 0.9873

0.33247 0.00743

R-squared 9 Mean dependent var 2

Adjusted R- 0.32700 5.00649

squared 8 S.D. dependent var 0

4.10712 5.67539

S.E. of regression 9 Akaike info criterion 7

4115.91 5.71802

Sum squared resid 6 Schwarz criterion 1

697.911 60.7658

Log likelihood 5 F-statistic 7

1.98302 0.00000

Durbin-Watson stat 9 Prob(F-statistic) 0

113
Table 15: Summary of estimation of ARMA (10, 10)

Dependent Variable: DCDF

Method: Least Squares

Date: 03/06/08 Time: 12:00

Sample (adjusted): 13 250

Included observations: 238 after adjustments

Convergence achieved after 13 iterations

Backcast: 3 12

Coefficie

Variable nt Std. Error t-Statistic Prob.

0.03073

C 4 0.438951 0.070016 0.9442

0.41715

AR(10) 1 0.375670 1.110417 0.2680

0.52888

MA(10) 1 0.353149 -1.497612 0.1356

0.01336 0.01710

R-squared 0 Mean dependent var 8

Adjusted R- 0.00496 7.93309

squared 3 S.D. dependent var 6

S.E. of regression 7.91338 Akaike info criterion 6.98751

114
5 3

14716.0 7.03128

Sum squared resid 9 Schwarz criterion 1

828.514 1.59106

Log likelihood 1 F-statistic 9

2.55500 0.20589

Durbin-Watson stat 0 Prob(F-statistic) 4

Inverted AR Roots .92 .74+.54i .74-.54i .28-.87i

.28+.87i -.28-.87i -.28+.87i -.74-.54i

-.74+.54i -.92

Inverted MA Roots .94 .76-.55i .76+.55i .29+.89i

.29-.89i -.29+.89i -.29-.89i -.76-.55i

-.76+.55i -.94

115
40
Forecast: CDF F
Actual: CDF
30 Forecast sample: 251 400
Adjusted sample: 251 381
Included observations: 121
20
Root Mean Squared Error 7.971633
Mean Absolute Error 7.166969
10 Mean Abs. Percent Error 66.33961
Theil Inequality Coefficient 0.244146
Bias Proportion 0.579859
0 Variance Proportion 0.090201
Covariance Proportion 0.329940

-10
275 300 325 350 375

CDFF

Figure 37: A summary of forecasting result of Box-Jenkins method

Table 16: GARCH-M model

Dependent Variable: CDF

Method: ML – ARCH (Marquardt) - Normal distribution

Date: 03/06/08 Time: 12:17

Sample (adjusted): 2 250

Included observations: 249 after adjustments

Convergence achieved after 18 iterations

Variance backcast: ON

GARCH = C(3) + C(4)*RESID(-1)^2 + C(5)*GARCH(-1)

Coefficie

nt Std. Error z-Statistic Prob.

@SQRT(GARCH) 0.27927 0.103521 -2.697787 0.0070

116
9

14.1238

C 2 0.584897 24.14756 0.0000

Variance Equation

21.5422

C 2 2.626939 8.200500 0.0000

0.70764

RESID(-1)^2 8 0.068842 10.27931 0.0000

0.08378

GARCH(-1) 0 0.047718 -1.755740 0.0791

0.16536 13.7191

R-squared 9 Mean dependent var 2

Adjusted R- 0.18447 8.35327

squared 4 S.D. dependent var 2

9.09116 6.60525

S.E. of regression 0 Akaike info criterion 6

20166.4 6.67588

Sum squared resid 0 Schwarz criterion 8

- 0.68483

Log likelihood 817.354 Durbin-Watson stat 9

117
4

Output from the parameter estimation of the model proved significant, and in

addition the coefficients of the sum of the Arch and Garch were closer to one, thus

proving that the model is stable.

120
Forecast: CDFF
Actual: CDF
80
Forecast sample: 251 400
Adjusted sam ple: 251 372
40 Included observations: 121

Root Mean Squared Error 4.854295


0
Mean Absolute Error 3.447139
Mean Abs. Percent Error 25.84455
-40 Theil Inequality Coefficient 0.185171
Bias Proportion 0.002594
-80 Variance Proportion 0.695304
Covariance Proportion 0.302101

-120
275 300 325 350

CDFF

500

400

300

200

100

0
275 300 325 350

Forecast of Variance

118
Figure 38: A summary of result of a dynamic GARCH-M forecast

120
Forecas t: CDF F
Ac tual: CDF
80
Forec ast s ample: 251 400
Adjusted sample: 251 372
40 Included observations: 121

Root Mean Squared Error 4.85429


0
Mean Absolute Error 3.44713
Mean Abs. Percent Error 25.8445
-40 Theil Inequality Coefficient 0.18517
Bias Proportion 0.00259
-80 Variance Proportion 0.69530
Covariance Proportion 0.30210

-120
275 300 325 350

CDFF

500

400

300

200

100

0
275 300 325 350

Forec ast of Variance

Figure 39: A summary of result of a static GARCH-M forecast

119
Vector Error Correction Model (VECM)

Table 17 Result from Cointegration test using Johansen Cointegration

test

Date: 03/06/08 Time: 12:33

Sample (adjusted): 7 371

Included observations: 365 after adjustments

Trend assumption: Linear deterministic trend (restricted)

Series: LSYSTEM LHELSINKI

Lags interval (in first differences): 1 to 4

Unrestricted Cointegration Rank Test (Trace)

Hypothesize

d Trace 0.05

No. of CE(s) Eigenvalue Statistic Critical Value Prob.**

None * 0.416156 214.7072 25.87211 0.0000

At most 1 * 0.048882 18.29287 12.51798 0.0049

Trace test indicates 2 cointegrating eqn(s) at the 0.05 level

* denotes rejection of the hypothesis at the 0.05 level

**MacKinnon-Haug-Michelis (1999) p-values

Unrestricted Cointegration Rank Test (Maximum Eigenvalue)

120
Hypothesize

d Max-Eigen 0.05

No. of CE(s) Eigenvalue Statistic Critical Value Prob.**

None * 0.416156 196.4143 19.38704 0.0001

At most 1 * 0.048882 18.29287 12.51798 0.0049

Max-eigenvalue test indicates 2 cointegrating eqn(s) at the 0.05

level

* denotes rejection of the hypothesis at the 0.05 level

**MacKinnon-Haug-Michelis (1999) p-values

Unrestricted Cointegrating Coefficients (normalized by

b’*S11*b=I):

LSYSTEM LHELSINKI @TREND(2)

-0.538467 -0.010573 0.000447

-0.231916 0.185189 0.000963

Unrestricted Adjustment Coefficients (alpha):

D(LSYSTEM

) 4.445258 0.321946

D(LHELSIN

KI) 6.526396 -0.654185

121
1 Cointegrating Log

Equation(s): likelihood -2236.576

Normalized cointegrating coefficients (standard error in

parentheses)

LSYSTEM LHELSINKI @TREND(2)

1.000000 0.019635 -0.000830

(0.02035) (0.00111)

Adjustment coefficients (standard error in parentheses)

D(LSYSTE) -2.393623

(0.15612)

D(LHELSIKI) -3.514246

(0.23655)

Estimation of Error Correction Model (ECT)

ECT=LYSTEM+0.019635*LHELSINKI-0.000830*@TREND (2)

122
40
30

20
10

0
-10

-20
-30

-40

-50
50 100 150 200 250 300 350 400

ECT

Figure 40: Error correction term series

Table 18: Results from the Unit root test

Null Hypothesis: D(ECT) has a unit root

Exogenous: Constant

Lag Length: 12 (Automatic based on SIC, MAXLAG=16)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -11.53205 0.0000

Test critical values: 1% level -3.448518

5% level -2.869442

10%

level -2.571047

*MacKinnon (1996) one-sided p-values.

123
Augmented Dickey-Fuller Test Equation

Dependent Variable: D(ECT,2)

Method: Least Squares

Date: 03/06/08 Time: 12:55

Sample (adjusted): 16 371

Included observations: 356 after adjustments

Coefficie

Variable nt Std. Error t-Statistic Prob.

14.9898

D(ECT(-1)) 5 1.299843 -11.53205 0.0000

12.7483

D(ECT(-1),2) 6 1.275552 9.994390 0.0000

11.4061

D(ECT(-2),2) 9 1.223897 9.319572 0.0000

9.92615

D(ECT(-3),2) 0 1.146680 8.656424 0.0000

8.29249

D(ECT(-4),2) 7 1.046997 7.920271 0.0000

6.58639

D(ECT(-5),2) 9 0.925844 7.113941 0.0000

D(ECT(-6),2) 4.93571 0.782420 6.308269 0.0000

124
9

3.68697

D(ECT(-7),2) 9 0.624312 5.905668 0.0000

2.70349

D(ECT(-8),2) 7 0.473787 5.706147 0.0000

1.81876

D(ECT(-9),2) 0 0.339830 5.351970 0.0000

1.10958

D(ECT(-10),2) 4 0.224429 4.944022 0.0000

0.61870

D(ECT(-11),2) 9 0.127910 4.837061 0.0000

0.25985

D(ECT(-12),2) 4 0.052313 4.967293 0.0000

0.01509

C 0 0.295237 -0.051112 0.9593

0.89702 0.02714

R-squared 8 Mean dependent var 8

Adjusted R- 0.89311 17.0373

squared 3 S.D. dependent var 8

5.57011 6.31124

S.E. of regression 9 Akaike info criterion 2

10610.9 6.46362

Sum squared resid 7 Schwarz criterion 6

125
-

1109.40 229.175

Log likelihood 1 F-statistic 2

2.04954 0.00000

Durbin-Watson stat 9 Prob(F-statistic) 0

Table 19: Output of the estimation of the short-run dynamics

Dependent Variable: D(LSYSTEM)

Method: Least Squares

Date: 03/06/08 Time: 13:04

Sample (adjusted): 4 250

Included observations: 247 after adjustments

Coefficie

Variable nt Std. Error t-Statistic Prob.

0.08743

D(LHELSINKI(-1)) 7 0.033883 2.580541 0.0105

ECT(-1) 1.26491 0.078100 -16.19598 0.0000

126
2

0.54004

C 1 0.430751 -1.253719 0.2111

0.59436 0.00795

R-squared 9 Mean dependent var 0

Adjusted R- 0.59104 10.5543

squared 4 S.D. dependent var 2

6.74945 6.66887

S.E. of regression 5 Akaike info criterion 2

11115.4 6.71149

Sum squared resid 5 Schwarz criterion 6

820.605 178.765

Log likelihood 7 F-statistic 8

2.03973 0.00000

Durbin-Watson stat 3 Prob(F-statistic) 0

127
150
Forecast: LSYSTEMF
100 Actual: LSYSTEM
Forecast sample: 251 400
50 Adjusted sample: 251 372
Included observations: 121
0
Root Mean Squared Error 20.41844
Mean Absolute Error 17.15579
-50
Mean Abs. Percent Error 2707.690
Theil Inequality Coefficient 0.818158
-100
Bias Proportion 0.698389
Variance Proportion 0.114911
-150
Covariance Proportion 0.186701

-200
275 300 325 350

LSYSTEMF

Figure 41: Results of dynamic forecast of VECM model

20
Foreca st: LSYSTEMFS
A ctual: LSYS TEM
Foreca st sample : 2 51 400
10 A djusted samp le: 251 37 2
Include d observatio ns: 121

Root Mean Squared Error 4.631233


0
Mean Absolute Error 3.252960
Mean Abs. Percent Error 184.4239
Theil Inequality Coefficient 0.806010
-10 Bias Proportion 0.003729
Variance Proportion 0.561783
Covariance Proportion 0.434488

-20
275 300 325 350

LSYSTEMFS

Figure 42: Results of static forecast of VECM model

128
System & Simulated Price

900.00
800.00
700.00
600.00
500.00 System Price
T im e

400.00 Simulated Price


300.00
200.00
100.00
0.00
1 23 45 67 89 111 133 155 177 199 221 243 265 287 309 331 353
Price

Figure 43: A comparisons of a System Price and simulated price with a mean

speed of 0.47 and a long run mean of 949.27 with a half-life 1.46

A 30-day rolling standard deviation was used instead of a constant standard

deviation

Estimating parameters

800.00
700.00
600.00
Y -v al u e s

500.00
Series1
400.00
Linear (Series1)
300.00
200.00
100.00
y = -0.4743x + 450.24
0.00
R2 = 0.0189
-150.00 -100.00 -50.00 0.00 50.00 100.00 150.00 200.00
X-values

Figure 44: Estimating parameter of a mean reversion model using Least

Square Estimation.

129
Comparing Returns
Log Returns of Sytem price
simulated returns

0.40
0.30
0.20
R e tu rn s

0.10
0.00
-0.10 0 50 100 150 200 250 300 350 400
-0.20
-0.30
Time

70
Series: Standardized Residuals
60 Sample 3 366
Observations 364
50
Mean -0.112784
40 Median -0.206567
Maximum 4.649894
Minimum -3.812500
30
Std. Dev. 0.995172
Skewness 0.793321
20
Kurtosis 6.304892
10
Jarque-Bera 203.8361
Probability 0.000000
0
-2.5 0.0 2.5

Figure 45: Summary statistics of returns series of GARCH (1, 1) estimation of the

system price

130
.4

.3

.2

.1

.0

-.1

-.2

-.3
50 100 150 200 250 300 350

RETURNS Resi duals

.05

.04

.03

.02

.01

.00
50 100 150 200 250 300 350

GARCH_VA RIANCE

Figure 46: Variance of the series returns

Table 20: Summary of Co-integration Test

Date: 03/03/08 Time: 11:16

Sample (adjusted): 7 250

Included observations: 244 after adjustments

Trend assumption: Linear deterministic trend

131
Series: LSYSTEM LHELSINKI

Lags interval (in first differences): 1 to 4

Unrestricted Cointegration Rank Test (Trace)

Hypothesize

d Trace 0.05

No. of CE(s) Eigenvalue Statistic Critical Value Prob.**

None * 0.430697 228.9098 15.49471 0.0001

At most 1 * 0.312582 91.45426 3.841466 0.0000

Trace test indicates 2 cointegrating eqn(s) at the 0.05 level

* denotes rejection of the hypothesis at the 0.05 level

**MacKinnon-Haug-Michelis (1999) p-values

Unrestricted Cointegration Rank Test (Maximum Eigenvalue)

Hypothesize

d Max-Eigen 0.05

No. of CE(s) Eigenvalue Statistic Critical Value Prob.**

None * 0.430697 137.4556 14.26460 0.0001

At most 1 * 0.312582 91.45426 3.841466 0.0000

Max-eigenvalue test indicates 2 cointegrating eqn(s) at the 0.05

level

* denotes rejection of the hypothesis at the 0.05 level

132
**MacKinnon-Haug-Michelis (1999) p-values

Unrestricted Cointegrating Coefficients (normalized by

b’*S11*b=I):

LSYSTEM LHELSINKI

-57.72870 5.542096

80.05778 -62.11374

Unrestricted Adjustment Coefficients (alpha):

D(LSYSTE) 0.048016 0.009584

D(LHELSIKI) 0.067214 0.052222

1 Cointegrating Log

Equation(s): likelihood 665.2841

Normalized cointegrating coefficients (standard error in

parentheses)

LSYSTEM LHELSINKI

1.000000 -0.096002

(0.04145)

Adjustment coefficients (standard error in parentheses)

D(LSYSTE) -2.771889

133
(0.21814)

D(LHELSIKI) -3.880190

(0.45750)

1240

1200

1160

1120

1080

1040
2004M04 2004M07 2004M10

Adj Close

Figure 47: Modelling of S&P500 series

134
.020

.015

.010

.005

.000

-.005

-.010

-.015

-.020
2004M04 2004M07 2004M10

LSP

Figure 48: The log series of the S&P index

Table 21: Output of the unit root test for stationarity

Null Hypothesis: D(ADJ_CLOSE) has a unit root

Exogenous: Constant

Lag Length: 0 (Automatic based on SIC, MAXLAG=15)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -15.41091 0.0000

Test critical values: 1% level -3.457173

5% level -2.873240

10%

level -2.573080

*MacKinnon (1996) one-sided p-values.

135
Augmented Dickey-Fuller Test Equation

Dependent Variable: D(ADJ_CLOSE,2)

Method: Least Squares

Date: 03/10/08 Time: 21:03

Sample (adjusted): 1/07/2004 12/22/2004

Included observations: 243 after adjustments

Coefficie

Variable nt Std. Error t-Statistic Prob.

0.99310

D(ADJ_CLOSE(1)) 6 0.064442 -15.41091 0.0000

0.35113

C 7 0.509783 0.688796 0.4916

0.49633 0.01098

R-squared 9 Mean dependent var 8

Adjusted R- 0.49424 11.1638

squared 9 S.D. dependent var 2

7.93928 6.98972

S.E. of regression 4 Akaike info criterion 0

15190.7 7.01846

Sum squared resid 7 Schwarz criterion 9

- 237.496

Log likelihood 847.250 F-statistic 0

136
9

1.99814 0.00000

Durbin-Watson stat 1 Prob(F-statistic) 0

Table 22: Output of autocorrelation test

Date: 03/10/08 Time: 21:11

Sample: 1/05/2004 12/22/2004

Included observations: 243

Partial Q-

Autocorrelation Correlation AC PAC Stat Prob

- -

****|. | ****|. | 1 0.464 0.464 52.936 0.000

- -

*|. | ***|. | 2 0.073 0.367 54.242 0.000

.|* | **|. | 3 0.081 0.198 55.889 0.000

- -

*|. | **|. | 4 0.102 0.265 58.501 0.000

.|. | **|. | 5 0.036 0.234 58.821 0.000

.|* | *|. | 6 0.068 0.128 59.977 0.000

.|. | *|. | 7 - - 60.311 0.000

137
0.036 0.094

- -

.|. | *|. | 8 0.041 0.143 60.740 0.000

.|* | .|. | 9 0.134 0.058 65.292 0.000

- -

**|. | *|. | 10 0.211 0.161 76.636 0.000

.|* | *|. | 11 0.118 0.088 80.188 0.000

- -

.|. | *|. | 12 0.016 0.143 80.255 0.000

Table 23: Output of AR and MA test

Dependent Variable: D(LSP)

Method: Least Squares

Date: 03/10/08 Time: 21:24

Sample (adjusted): 1/15/2004 9/01/2004

Included observations: 159 after adjustments

Convergence achieved after 11 iterations

Backcast: 1/07/2004 1/14/2004

138
Coefficie

Variable nt Std. Error t-Statistic Prob.

-9.09E-

C 05 0.000545 -0.166806 0.8677

0.79603

AR(6) 3 0.048746 16.33022 0.0000

0.93451

MA(6) 3 0.032376 -28.86471 0.0000

0.13417 4.24E-

R-squared 3 Mean dependent var 05

Adjusted R- 0.12307 0.01029

squared 3 S.D. dependent var 2

0.00963 6.42751

S.E. of regression 8 Akaike info criterion 1

0.01449 6.36960

Sum squared resid 1 Schwarz criterion 7

513.987 12.0873

Log likelihood 1 F-statistic 1

2.89384 0.00001

Durbin-Watson stat 1 Prob(F-statistic) 3

139
Inverted AR Roots .96 .48+.83i .48-.83i -.48+.83i

-.48-.83i -.96

Inverted MA Roots .99 .49-.86i .49+.86i -.49-.86i

-.49+.86i -.99

Table 24: Forecast output (static forecast-using the arma)

.04
Forecast: LSPF_STATIC_ARMA
.03 Actual: LSP
.02 Forecast sample: 9/01/2004 12/22/2004
Included observations: 79
.01
Root Mean Squared Error 0.008832
.00
Mean Absolute Error 0.007069
-.01 Mean Abs. Percent Error 449.4588
Theil Inequality Coefficient 0.663762
-.02 Bias Proportion 0.000372
Variance Proportion 0.000263
-.03
Covariance Proportion 0.999364
-.04
-.05
2004M09 2004M10 2004M11 2004M12

LSPF_STATIC_ARMA

.00018

.00016

.00014

.00012

.00010

.00008

.00006
2004M09 2004M10 2004M11 2004M12

Forecast of Variance

Figure 49: Output forecast (dynamic forecast)

140
.12
Forecast: LSPF_DYNAMIC
Actual: LSP
.08
Forecast sample: 9/01/2004 12/22/2004
Included observations: 79
.04
Root Mean Squared Error 0.010026
Mean Absolute Error 0.008392
.00
Mean Abs. Percent Error 789.5745
Theil Inequality Coefficient 0.645704
-.04 Bias Proportion 0.553835
Variance Proportion 0.170345
-.08 Covariance Proportion 0.275819

-.12
2004M09 2004M10 2004M11 2004M12

LSPF_DYNAMIC

Table 25: List of Ratios

141
142
143
144

Você também pode gostar