THE UNIVERSITY OF SHEFFIELD

DEPARTMENT OF ECONOMICS

THE DAY–OF–THE–WEEK EFFECT

IN THE COLOMBIA STOCK EXCHANGE

SUBMITTED BY
OSCAR DAVID GALLEGO CARVAJAL

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF

MA IN ECONOMICS OF
MONEY, BANKING AND FINANCE

SUPERVISOR: PROFESSOR DAVID CHAPPELL

MAY 2005
 ACKNOWLEDGEMENTS

I would first like to thank God for the blessings received finishing this
dissertation.

My mother has been my greatest source of inspiration. My brother, for their

moral and economic support in the whole course. I would like also to thank
Erika, for her love and support during the preparation of this dissertation and
throughout last year.

Many people on the faculty and staff of the Department of Economics assisted
and encouraged me in various ways during my course of studies. I am
especially grateful to the Professor Michael Dietrich, for the patience and
support during this Master course.

Finally, this dissertation would not have been possible without the expert
guidance of my esteemed supervisor, Professor David Chappell. Not only was
he readily available for me, he always read and responded to the drafts of each
chapter of my work more quickly than I could have hoped. All his written
comments are always extremely perceptive, helpful, and appropriate.

1
 TABLE OF CONTENTS

ABSTRACT 1

CHAPTER 1. INTRODUCTION 2

CHAPTER 2. LITERATURE REVIEW 6

2.1 UNCERTAINTY ASSOCIATED WITH THE 10

FUTURE: THE ARCH MODEL

CHAPTER 3. THE COLOMBIAN STOCK EXCHANGE 14

CHAPTER 4. DATA AND METHODOLOGY

4.1 DATA 20
4.2 METHODOLOGY 24

CHAPTER 5. EMPIRICAL RESULTS 33

CHAPTER 6. CONCLUSIONS 43

BIBLIOGRAPHY 47

APPENDICES

APPENDIX 1 51
APPENDIX 2 53
APPENDIX 3 53
APPENDIX 4 54
APPENDIX 5 55
APPENDIX 6 55
APPENDIX 7 56

2
 ABSTRACT

This study investigates the existence of day of the week effects on stock returns
in the Colombian Stock Exchange (CSE) for the period between June 2001 and
March 2005. The Bogotá Stock Exchange was established in 1928. However,
the two other main bourses in the country merged with this in 2001 to create the
CSE. Since then, the CSE is becoming a good diversification alternative for
both domestic and foreign investors.

The modelling in the study begins with linear regression analyses, but the data
generating process is shown to be non-linear. A non-linear GARCH model is
then applied, achieving a good explanation for the modelled rates of return.
Results obtained indicate the significant presence of day of the week effects in
both returns and volatility. The maximum return is on Friday whereas the
minimum is on Tuesday, with return variances at their highest on Monday.

3
 CHAPTER 1.

INTRODUCTION

Calendar anomalies in stock market returns have been of considerable interest

during the last three decades. The main anomalies can be listed briefly as the
weekend effect, the day of the week effect, and the January effect.

In several empirical studies concentrating on daily stock returns, the

distributions of stock returns are assumed to be identical for all weekdays.
However, numerous studies document that this assumption is not correct.
French (1980), Gibbons and Hess (1981), Kein and Stambaugh (1984),
Lakonishok and Levi (1982), and Rogalski (1984) demonstrate that the
distribution of stock returns varies according to the day of the week.

Sullivan et al. (2001) focus their attention on the case where among the social
sciences, economics predominantly studies non-experimental data and thus
does not have the advantage of being able to test hypothesis independently of
the data that gave rise to them in the first instance. Therefore, they criticize the
fact that none of the calendar effects were preceded by a theoretical model
predicting their existence.

The findings of some studies have shown that the average return for Monday is
significantly negative for countries like the United States, the United Kingdom,
and Canada. On the other hand for several Pacific rim countries, the lowest rate

4
of return tends to occur on Tuesday (Jaffe and Westerfield, 1985; Dobois and
Louvet 1996).The day of the week effect has been explained by examining
various kinds of measurement errors: the delay between trading and settlement
in stocks and in clearing checks; specialist related biases; the distinction
between trading and non-trading periods; the timing of corporate and
government news releases; and time zone differences between relevant
countries and markets. These are some of the possible explanations, but so far
none of the suggested explanations is entirely adequate.

These day of the week findings appear to conflict with the Efficient Market
Hypothesis since they imply that investors could develop a trading strategy to
benefit from these seasonal regularities. In other words, any predictable pattern
in stock returns and variances may provide investors with received returns
different from the stock market average.

However, Berument and Kiymaz (2003) argue that for a rational financial
decision market, returns constitute only one part of the decision-making
process. Another part that must be taken into account when one makes
investment decisions is the risk or volatility of returns. A formal test on the
variations of volatility across days of the week is useful because that enables us
to see whether the higher return on a particular day is just a reward for higher
risk on that day.

The purpose of this study is analyze the day of the week effect in stock market
returns and volatility by examining the Colombian Stock Exchange (CSE) index
during the period of June 2001 and March 2005. In the emerging markets, with
their unstable characteristics and relative few researches in this area, this study
can be considered as one of the pioneering studies for CSE that examines data
for the existence of a day of the week effect.

There exist two types of analysis for calendar anomalies. On the one hand, the
most common is the study of the presence of the day of the week effect in stock
market returns. Most of the studies investigating the day of the week effect in
returns employ the standard OLS methodology by regressing returns on five

5
daily dummy variables. On the other hand, GARCH (Generalized
Autoregressive Conditional Heteroskedasticity) models are used for the time
series behaviour of stock prices in terms of volatility. They take into account the
possibility that the variance is time dependent: a feature that is common in
stock return series and in financial time series in general.

Furthermore, the CSE will be modeled with linear and non-linear models in
order to capture any possibility of non–linear structure in the data generating
process. Accurate results will suggest the possibility of a trading strategy in
order to take advantage of the market inefficiency, buying and selling strategies
formulated accordingly to increase returns due to better timing. (For example,
buy on Tuesday, sell on Friday).

Initially, the rate of return is tested for market efficiency. As the tests applied for
efficiency show a departure from the random walk hypothesis, if efficiency is not
achieved, the sample is further examined in a more general linear model to look
for the presence of anomalies in the CSE returns. The usual linear regression
method error assumptions on daily stock returns are violated and the Brock,
Dechert and Scheinkman (1987) BDS test indicates that this model is
inadequate in explaining rates of return. Therefore, GARCH models are fitted
and tested; insignificant BDS statistics prove that these non-linear models
explain all the structure in the CSE returns data.

Empirical findings show that the day of the week effect is interrupted by a
structural change in the CSE on May 13th 2004, the day that the CSE had the
largest fall in the last 5 years. So the data should be split and analysed in two
sub-groups to achieve accurate results. The day of the week effect is
presented in both the return and the volatility. While the highest and lowest
returns are observed on Friday and Tuesday, the highest and the lowest
volatilities are observed on Monday and Tuesday, respectively.

6
The dissertation is organized as follows: Chapter 2 presents the literature
review; Chapter 3 describes the CSE; Chapter 4 introduces the data and
methodology used; Chapter 5 presents and discusses the empirical results and,
finally, Chapter 6 gives a summary of the findings of the study and offers some
conclusions.

7
 CHAPTER 2.

LITERATURE REVIEW

The Efficient Market Hypothesis (EMH) originated in the random walk theory
that emerged in the security price literature in the 1950’s. It states that the size
and direction of a price change at a particular time is random with respect to the
knowledge available at that point of time; future prices of a security are no more
predictable than a series of random numbers. Therefore, Markiel (2003) argues
that the logic of the random walk idea is that if the flow of information is
immediately reflected in stock prices, then tomorrow’s price change will reflect
only tomorrow’s news and will be independent of the price changes today. But
news is by definition unpredictable and therefore resulting price changes must
be unpredictable and random.

A number of researchers have found different and strange performance of the

market in relation to the day of the week, public holidays, change of month, and
even for the hour of the day; all of these peculiarities are known as “Calendar
Anomalies”.

French (1980) originally discussed two hypotheses. Firstly, the calendar time
hypothesis suggests that the average Monday return (the day following the

8
weekend) should be 3 times the average returns that occur on other days of the
week. In contrast, the trading time hypothesis postulates that the Monday
returns should not be significantly different from the return available on any
other day of the week. He observed that stock returns are higher than average
on the last trading day of the week and lower than average on the first, in the
Standard and Poor’s index over the 25 years from 1953 to 1977.

Many researchers have attempted to explain what has come to be known as the
“weekend” or “day-of-the-week” effect. Contrary with French’s findings, Rogalski
(1984) discovers that all the average negative returns from Friday close to
Monday close documented in the literature for stock market indexes occurs
during the non-trading period from Friday close to Monday open. He calls these
Monday effects, the non-trading weekend effect. His evidence also suggests a
January effect; which was found by segmenting the day of the week returns into
January’s versus the rest of the year. It reveals that the Monday effect and the
non-trading effects are on average positive in January and on average negative
for the rest of the year. Finally, a relation between the Monday/January effect
and firm size was found, where the close to close returns of small firms on
Monday in January are on average positive (and greater than the corresponding
positive returns of large firms), rather than the rest of the year where small and
large firms have negative returns. However, a relation between Monday and
firm size is not at all evident.

Furthermore, in response to previous results, Lakonishok and Levi (1982)

presented an explanation based on the delay between trading and settlements
in stocks, and in clearing checks. Basically, they propose an explanation to
measure daily returns, that should depend on the day of the week and that
adjustment for interest gains on certain days over adjacent business days that
should be made. In the 1980´s, the United States settlement on traded stocks
took place five business days after trading, nowadays it takes just three
business days. Checks that clear via the United States Federal Reserve System
take one business day from the time they are delivered to the commercial
banks, to the time that usable funds are debited and credited. Normally, the in-
clearing delay means that in weeks without a holiday, stocks purchased on

9
business days other than Friday give the buyer eight calendar days before
losing funds for stock purchases. These eight days are the five business days
for settlement, the two weekend days, and the check clearing day. However,
when the trade is taken on Friday the purchase will not actually occur until the
second following Monday, ten calendar days after the trade. Consequently, it is
important to understand the position of the buyer and seller of stocks; buyers
should therefore be prepared to pay more on a Friday than on other days by the
amount of two days interest. The sellers of stocks should also require a higher
price for stocks sold on a Friday because of the two days extra delay before
being paid.

Finally, the results pointed out that over earlier periods, unadjusted returns on
Mondays are significantly negative, and returns on Fridays are positive; these
findings are similar to those of previous studies. Further results found that
taking into account interest earned during weekends and holidays, a calendar
time is relevant for interest bearing securities and therefore also for the
alternative of holding stocks.

Keim and Stambaugh (1984), for example, provide evidence of a double the
length of the period examined by French (1980). The returns for one-day
weekends (Saturday to Monday close) computed from historic New York Stock
Exchange (NYSE)1 data are more negative than returns for (the current) two-
day weekends. They also investigate the possible relation between the
weekend effect and firm size, and find that the smaller the firm the greater is the
tendency for average returns to be high on Friday.

Connolly (1989) was concerned about the foundation of econometric models for
the Day of the week effect. He cast doubt on the statistical significance of the
day-of-the-week effect per se by showing that appropriate adjustments for
sample size, heteroscedasticity, autocorrelation, and leptokurtosis greatly
reduce the significance of regression F and t-values. Thus, in this model the
error distribution may be conditionally heteroscedastic and non-normal.

1
During much of the 25-year period from 1928 through 1952, the NYSE was open on
Saturdays.

10
According to Connolly (1989) this is useful because the unconditional
leptokurtosis may be traced to non-normality in the conditional error distribution
and to time varying heteroscedasticity.

Hence, the evidence for a weekend anomaly is clearly dependent on the

estimation method and the sample period. However, he shows that when
transactions costs are taken into account, the probability that arbitrage profits
are available from weekend-oriented trading strategies seems very small.
Moreover, Chang et al. (1993), extend Connolly´s work by examining the
robustness of the day-of-the-week effect in international markets, and
emphasizes in the violation of the error terms in the OLS model, whereas the
day-of-the-week effect has largely disappeared within most countries.

Nevertheless, compared to studies of the US market, few studies have been

carried out on the day-of-the-week anomaly for non-US markets. There are
some for Japanese stocks and some on major European markets. Solnik &
Bousquet (1990) analyse the Paris Bourse, which exhibits a day-of-the week
effect with a strong and persistent negative return on Tuesdays. Dobois and
Louvet (1996) examine the day-of-the–week effect for eleven stock indices from
nine countries; Canada, United States, Japan, Hong Kong, Australia, Germany,
France, United Kingdom and Switzerland, during the 1969 – 1992 period. The
standard methodology is the moving average. They found negative returns on
Tuesdays for the Australian (1980-1992) and Japanese (1969 – 1988) indices.
Non-synchronous trading may be an explanation for the one-day lag. However,
some correlation analysis was made among daily returns with a one-day lag
between Western and Eastern countries, where the correlation between
Monday returns in the US market and Tuesday returns in the Japanese market
is higher than other days of the week.

In contrast, in Australia, there is a 14 hours difference between Sydney and

New York and that the Australian Stock Exchange opens 3 and half hours on
Tuesdays after U.S. markets close on Mondays. Therefore, one could
conjecture that the U.S. negative Monday returns potentially cause the negative
Tuesday returns in Australia as the average negative performances of the U.S.

11
markets on Mondays have immediate impact on the subsequent performance of
the Australian market on Tuesdays, but studies conclude that the day-of-the-
week effect in Australia is independent from the U.S. seasonal.

Oguzsoy and Guven (2003) investigate the existence of a day-of-the-week

effect in the Istanbul Stock exchange (ISE) for the years 1988 and 1999. They
observe that for most of the stocks among the 30 most heavily traded stocks of
ISE, the maximum return is on Friday whereas the minimum return is either on
Mondays or Tuesdays with return variances at their highest on Mondays.

Finally, Kiymaz and Berument (2003) include an analysis of a day-of-the-week

effect in the volatility of the major stock markets indices of Canada, Germany,
Japan, the United Kingdom, and the United States for the period of 1988 - 2002.
They employ a GARCH model in order to capture the conditional
heteroscedasticity and the day of the week effects.

In the financial field, the existence of a day-of-the-week effect seems to be a

wide spread and well accepted phenomenon. However, there is a question
remaining about the real possibility of enjoying abnormal returns, given the fact
that there is nothing that guarantees the permanence of these anomalies in the
future. Markiel (2003) argues that the anomalies are not dependable from
period to period. Even more, it is well-known that Wall Street traders make a
joke that the “January effect” is more likely to occur on the previous
Thanksgiving. Moreover, these non-random effects are very small relative to the
transactions costs involved in trying to exploit them.

2.1. UNCERTAINTY ASSOCIATED WITH THE FUTURE: THE ARCH MODEL

In the first part of the literature review, the day of the week in stock market
returns was the subject of concern. However, there is a second issue, which
has also been investigated, the time series behaviour of stock prices when
volatility is time-varying. This has been investigated by employing a range of

12
variants of the ARCH (Autoregressive, Conditional Heteroscedasticity) or
GARCH (generalised ARCH) class of models. It is important to ascertain
whether the higher return on a particular weekday is just a reward for higher risk
on that day.

French et al. (1987) were also concerned with statistical approaches to

investigate the relation between expected stock returns and volatility. They used
daily returns to compute estimates of monthly volatility and reported that
unexpected stock market returns are negatively related to the unexpected
changes in volatility. It has been recognized for quite some time that uncertainty
of speculative prices, as measured by the variances and covariances, are
changing through time. One of the most prominent tools that has emerged for
characterizing such changing variance is the ARCH model of Engle (1982). The
importance of this model is that it enables one to quantify the variations in
volatility across days of the week. As stated above, this is of interest because it
is important to know if the higher return on a particular weekday is just a reward
for higher risk on that day.

Since the introduction of the ARCH model, several hundred research papers
applying this modelling strategy to financial time series data have already
appeared. Engle (1982) proposes the following, ARCH(1), model to represent a
series with changing volatility.

rt = α + ε t
(1) σ t2 = a + bε t2−1 (2)
εt ~ N (0, σ t2 )

The assumption in (2) that volatility is a deterministic function of past returns is

restrictive. However, the ARCH model is attractive, because the return and
variance process are estimated jointly. Hence, it is a traditional econometric
model which assumes a constant one-period forecast variance and, in
particular, as far as the day-of-the-week is concerned, the variance of returns
tends to be higher on days following closures of the market.

13
The contribution of French and Roll (1986), one of the first papers to use daily
unconditional variances, remains significant among the low–order ARCH
models for daily index returns presented in French et al (1987). Next, Bollerslev
(1986) introduced a generalized autoregressive conditional heteroscedasticity
(GARCH) model. This model is capable of capturing the three most important
empirical features observed in stock return data: leptokurtosis, skewness and
volatility clustering.

σ t2 = a + bσ t2−1 + C1ε t2−1 + C 2 ε t2− 2 (3)

Engle, Lilien and Robins (1987) and Bollerslev, Engle and Wooldridge (1987)
propose generalizations of the ARCH model that allow the conditional mean
return to be a function of volatility and they refer to these as GARCH–in–mean
models. An example is given in (4) and (5) below.

( Rmt − Rf t ) = α + βσ t + ε t − θε t −1 (4)

( Rm t − Rf t ) = α + βσ t2 + ε t − θε t −1 (5)

Where Rmt – Rft is the daily excess holding period return of the S&P composite
portfolio, and σt2, the variance of the unexpected excess holding period return
εt, follows the process in (3) above. The scope of the GARCH class of models is
enormous and its fits the data rather better than other specifications for the
variance.

Other authors analyse the international market using this methodology. Alexis
and Xanthakis (1995), for example, provide evidence for a day-of-the-week
effect in the Greek stock market which shows particular characteristics of
uncertainty and risk. They test the day-of-the-week effect using a GARCH-M
model. The period examined is split in two subgroups, 1985 – 87 and 1988 –
94. Results show different features of stock returns and the presence of a day-
of-the-week effect. The first sub-period indicates high positive returns for
Monday and negative returns on Tuesday while the second sub-period shows

14
negative returns for Mondays, a reduction in the negative returns on Fridays,
and lower returns on Mondays.

Al-Loughani and Chappell (2001) carry out an analysis where the series are
highly leptokurtic relative to the normal distribution. This feature is used as a
justification for the use of a GARCH model to investigate the presence of a day-
of-the-week effect. They also test the independent and identically distributed IID
assumptions through the application of the Brock, Dechert and Scheinkman
(BDS) . If the residuals are not IID, then there is some structure in the data that
have not been explaining. Therefore, through the study this issue will likely
analyse.

Overall, results indicate that, during the 1980´s, this type of calendar anomaly
was clearly evident in the vast majority of developed markets, but it appears to
have faded away somewhat in the 1990´s. However, in the following chapters,
the CSE will be analyse including the problems on variance unchanged over
time and the possible nonlinear structure on the residuals.
Whereas the CSE is classified as an emerging market. Therefore it is plausible
that we may uncover some market inefficiencies that investors could exploit as
a trading strategy to benefit from these seasonal regularities.

15
 CHAPTER 3.

THE COLOMBIAN STOCK EXCHANGE

The Colombian capital market and, in particular, the Bogotá Stock Exchange,
were founded some 75 years ago in 1928. This coincides with the time that the
industrialization of Colombia really took off. In the years subsequent to 1928 the
Medellin Stock Exchange and the West Stock Exchange were founded and in
2001 all three exchanges were merged to create the Colombian Stock
Exchange (CSE).

Generally, in industrialized countries, the returns from investment on the

national stock exchange – and on other developed world stock exchanges -
results in a valuable flow of income to investors. However, in Colombia, this has
not always happened. Colombia, in recent years, has been dogged by
criminality (often related to drug dealing) and political crises. For example,
during the Samper Government (1994 – 1998) there was a political crisis
generated by rumors of the presidential campaign being partly funded by the
results of drug dealing. The last President (Pastrana, 1998 - 2002) was held to
be at least partly responsible for an economic recession and widespread
political violence in Colombia.

However, last year Colombia's economy expanded more than expected in the
fourth quarter as a reduction in kidnappings and murders bolstered consumer
confidence, helping fuel a surge in retail sales. The Gross Domestic Product,

16
the broadest measure of a country's output of goods and services, grew by
4.28% over the year, up from 2.46% in the previous quarter. Furthermore, the
improvements in national security, the low levels of FED rates and the
revaluation of the Colombian peso against the dollar, helped to boost the
domestic economy.

With the weak dollar policy by the U.S. in the short and medium term, with the
local consolidation of the external sector, the environment was propitious for the
peso-to-dollar exchange rate to continue appreciating during 2005. In fact, after
one month and a half of the Central Bank’s announced measure of intervening
directly in the foreign exchange market (through buying up to US$1 billion), and
during this period the FX rate had managed to remain relatively steady above
COP$2,550 at the end of 2004, the local-currency appreciation trend managed
to intensify in such a way that rapidly the Colombian peso strengthened to
reach levels of COP$2,350 per dollar towards the end of the month of March
2005.

Also, this strengthening of the economy has continued. In addition to a

generalized rise in investor appetite for emerging market assets and rising
remittances from Colombians living abroad, the weakening of the US$ in
international markets has served to bolster the Colombian currency, thereby
establishing a steady foreign capital inflow and increasing private sector
investment in both foreign and national markets.

In the international scenario, the economic situation in the U.S. played the
starring role, given that appreciation in world currencies against the dollar
intensified in November 2004. In the meantime, economic authorities of the U.S.
(Secretary of Treasury John Snow and President of the Federal Reserve Alan
Greenspan) reiterated that the closing of imbalances in the current and fiscal
accounts of the largest economy in the world would necessarily have to go
through further weakness in the U.S. currency.

Simultaneously, the debate on the need for flexibility in China’s foreign

exchange system regained strength, precisely with the purpose of allowing for a
correction, without major traumatisms, of the macroeconomic imbalances in the

17
U.S. Separately, it should be noted that the FED increased its reference rate for
a fourth consecutive time. In fact, with the new 25-bp(basic points) hike on
November 10, the accumulated increase in the year to date completed 100 bps,
rising from 1.0% in June to 2.0% in November.

In response to the weak dollar policy and the decline in oil prices (which, after
reaching historic highs in October), and considering the solid economic results
of important emerging economies like the Brazilian increased capital inflows to
the region, and country risk reached minimum levels during November.

In the case of Colombia, the spread over U.S. Treasury bonds narrowed to 331
bps, while at the same time local currencies and markets strengthened. In 2004,
the CSE was the best investment opportunity in Colombia, experiencing a
sharp increase in both value and size. In this context, the Index exhibited
several historically high values. In January it started at 2335.98 units and
increased to 4300 by the end of the year, an increase of 84%. The daily
average in November was COP$342 billions per day. The total value in 2004
was around COP$4.77 billions, a massive increase from the COP$1.89 billions
in 2003. These figures go a long way towards explaining why prestigious
international agencies, such as Bloomberg and The Economist, chose the CSE
as one of the best performers for stock exchange business around the world;
helped also by the healthy economy and stable political scenario. The
performance of the CSE over the period November 2003 to February 2005 is
illustrated in the chart below.

18
GRAPH 1. CSE FROM NOVEMBER 2003 TO FEBRUARY 2005

CSE Index

6000,00
5240,16

5000,00 4489,78
4276,38 4345,83

4000,00 3545,57
3742,59
3321,15 3374,25
3089,75 2966,55 3004,45 2953,48 3015,02
3000,00 2772,49
2224,59 2333,70

2000,00

1000,00

0,00
Nov-03 Dic-03 Ene-04 Feb-04 Mar-04 Abr-04 May-04 Jun-04 Jul-04 Ago-04 Sep-04 Oct-04 Nov-04 Dic-04 Ene-05 Feb-05

CSE Index

Source: BVC

However, despite the relatively long stock exchange tradition in Colombia, the
liquidity and the operations volume is not particularly high. This is due to high
concentration and a modest stock-exchange culture in the country. There are
40 registered brokerage houses and 124 stocks traded, consisting of just those
few those that are traded on a daily basis. But the trading volume has been
strongly increasing over the last years, surging to a daily average of COP$7
million in 2004, up from about COP$500,000 the previous year. Some of the
biggest increases in the index are connected to Grupo Empresarial Antioqueño
such as Bancolombia, SurAmericana, Nacional de Chocolates, Cementos
Argos, etc.

However, the players who have been most actively contributing to this growth
are private citizens. Around 36% of the investors are private individuals, second
are the pension funds with 20% of total participation, then the corporate sector
with 17%, the mutual funds with 13% and the rest 14% is controls for foreign
investors.

In contrast with the above, in the same year, on May 13, the CSE suffered the
largest daily fall, of 6.17%, in the exchange's history. The accumulated loss of
the previous two weeks (beginning April 27) in the CSE index was 21.5%. It was
a taste of what can happen within two years, a vision of (as it has been
described in Colombia) the end of the speculative surge. This will be a central

19
issue in our data analysis, and it may prove to be a significant breakpoint that
could indicate a structural change in the behavior of the CSE before and after
the stock crisis.

Analysts attributed the May 13 drop in the stock index to the rumors of an
interest rate increase in the US. This is undoubtedly a key factor in triggering
the bursting of the bubble. Capital flies fast and massively. May 13 made it
possible to see a little of what could happen: no one will want to buy more
treasury bills (Titulos de Tesoreria or TES), they will just want out of the
Colombian market. The crisis is illustrated by the figures in the following table.

TABLE 1. LARGEST FALL ON THE CSE

CSE Index 2004
Index Var.%
May-05 3.343,46
May-06 3.276,94 -1,99%
May-07 3.201,84 -2,29%
May-10 3.054,82 -4,59%
May-11 2.996,92 -1,90%
May-12 2.897,90 -3,30%
May-13 2.718,97 -6,17%
Source: BVC

Nevertheless, the nervousness in the CSE was stabilized the following day, and
the index increased by 7%. During the rest of the year the index showed a clear
recovery, closing at the end of December with wealthy, satisfied investors, who
once again have faith in the Colombian Stock Market.

In conclusion, Colombia has moved forward on a number of fronts in its efforts

to spur the growth of its securities market. There is currently a conducive
environment for market development, which is in turn the result of:

1. A stable macroeconomic environment

2. Consolidation of a substantial base of institutional investors — and with
it, a higher demand for capital market instruments

20
 3. Implementation of numerous reforms aimed at establishing an
appropriate regulatory and institutional framework, and
4. The launching of several initiatives aimed at creating an appropriate
infrastructure to adequately support the market dynamics.

In addition, the Securities Commission is aiming to improve transparency

standards through an accounting reform geared towards the adoption of
international accounting and auditing standards. All of these features bode
well for the future of the CSE.

21
 CHAPTER 4.

DATA AND METHODOLOGY

4.1. DATA

The data utilized in this study was formally requested from and supplied by the
CSE. It runs from 3 July 2001 (the day that the three regional Stock Exchanges
in Colombia merged to form the CSE), to 16 March 2005. The CSE is open from
Monday to Friday, and the data covers 906 trading days (after excluding the
huge number of bank holidays and national holidays when the CSE is closed).
The Daily rates of return for the CSE are expressed in local currency,
Colombian Pesos (COP). Throughout the study all statistics are calculated and
equations estimated using E-views 4.0 software.

The daily returns are the first differences of the natural logarithms daily closing
prices of the CSE index and are defined by:

Rt= ln ( Xt / Xt-1)

Where Xt and Xt-1 are the daily closing prices of the index at time t and time t-1
respectively.

However, it is important to mention a well-known economic data problem that is

related to the presence of calendar effects; the data mining or data-driving
problem. Sullivan et al (2001) draw attention to the danger of data mining. It is

22
the disadvantage of being able to test hypotheses independently of the data
that gave rise to them in the first instance. Therefore, in general practice, the
use of the same data set to formulate and test hypotheses introduces data-
mining biases. They also argue that none of the calendar effects were preceded
by a theoretical model predicting their existence. Conversely, when the
observational evidence supports a theory the confirmation is much stronger
when the evidence is new.

Some descriptive statistics for the data can be seen in table 2. It shows the
minimum return which corresponded at the fall on 13th may, 2004, a maximum
return that corresponded, the 28th December, 2001. Moreover, the series is
leptokurtic: i.e. they all have distributions with high kurtosis. 'High' is usually
taken to mean that the fourth central moment is greater than 3, the coefficient of
kurtosis for any normal distribution. The shape of such a distribution is typified
by a high degree of ‘pointedness’ and fat tails compared to a normal
distribution.

Table 2.
Summary Statistics for Daily Returns from July 2001 to
March 2005

Sample Period July 3, 2001 to March 16 ,2005

Observations 906
Mean Return 0,1695
Median Return 0,1499
Maximun Return 8,8981
Minimun Return -6,3733
Standard Deviation 1,1941
Skewness 0,4273
Kurtosis 11,6227

According to Adcock C, J (2000, p.18)2 fat tails are of particular interest in

finance because the presence of fat tails means, inevitably, that there is a high
probability of extreme returns, leading to larger profits or losses, than what
would be expected with a normal distribution.

23
If asset returns are normally distributed, then there is an exact linear regression
relationship between the observed return on an individual asset and the
observed return on the market. If returns are not normal then it is not always
clear that these linear regression type relationships hold and, if they do hold,
they may require modification and/or different procedures for estimation of
model coefficients.

GRAPH 2. MEAN RETURN FOR THE CSE

Mean Returns
Friday;
0,430467334
0,45
0,4
0,35
0,3 Monday;
0,25 0,199400399 Thursday;
0,2 0,167454533
Wednesday;
0,15
0,089896213
0,1
0,05
Tuesday; -
0
0,034844746
-0,05
Days of the week

Graph 2 also illustrates the mean stock return for each working day of the week.
Mean returns for Mondays are positive (0.1994), contrary to several previous
studies; i.e. Osborne (1962), Cross (1973), French (1980), Gibbons & Hess
(1981), Keim & Stambaugh (1984), Jaffe & Westerfield (1985) all find strong
evidence that Mondays’ average returns are negative and Fridays’ are positive.

However, in some other studies such Solnik & Bousqet (1990) in the French
Market , Dobois & Louvet (1996) in the stock markets of Japan and Australia,
and Jaffe & Westerfield (1985) in the stock markets of Australia and Japan. In
the CSE, the negative average return is observed to occur on Tuesdays with a
mean of –0.03484. Some authors have justified this observed behavior by the

2
Adcock, Christopher J, “Fat tails and the capital asset pricing model” from Dunis, Christian, ed.
Advances in quantitative asset management. Kluwer Academic, 2000. Chapter 2, p.17-39.

24
time difference ahead to New York. However, Jaffe & Westerfield (1985) tested
the hypothesis and it was rejected. Nevertheless, for the CSE, the time
difference with the NYSE is small; hence there is no justification to formally test
a correlation between those stock exchanges.

Moreover, in Appendix 1, histograms of daily percentage close-to-close returns

for each day separately are illustrated. The relative higher dispersion of
Tuesday returns can be seen at a glance. The frequency axis is informative for
both observing the dispersion on Tuesdays and the dominant positive returns
on Fridays.

Table 3. Number of days with more than 0%,1%,2%,3%,4%

increase/decrease in returns.
Criteria Mondays Tuesdays Wednesday Thursdays Friday All Days
Rt>0% 93 101 96 114 124 528
Rt>1% 30 29 31 32 38 160
Rt>2% 10 7 7 5 10 39
Rt>3% 4 0 1 2 3 10
Rt>4% 0 0 1 0 3 4
Rt<0% 64 87 91 72 64 378
Rt<-1% 12 34 23 19 12 100
Rt<-2% 0 13 7 8 3 31
Rt<-3% 2 4 1 3 0 10
Rt<-4% 1 3 0 2 0 6

Table 3 presents the attractive behaviour of the CSE. As can be observed from
the table, in 528 (378) of 906 observations the index increased (decreased)
between 0% and 1% in a day, and 160 (100) days it increased (decreased)
more than 1% in a day. Friday is the day with the highest number of positive
jumps and Tuesday with negative ones.

25
4.2. METHODOLOGY

The natural logarithms, ln (Xt), of the closing prices of the CSE index, Xt, for the
period 3rd July 2001 to 16th March 2005 are given in Fig. 3 below.

GRAPH 3. FULL PERIOD LN (XT)

8.8

8.4

8.0

7.6

7.2

6.8

6.4
250 500 750

Ln(Xt)

At a first glance the graph shows that ln (Xt) appears to is non-stationary. If the
data is non-stationary the mean and variance change over time and the data
has to be differenced one or more times to achieve stationarity. However, the
formal way to determine whether the series is stationary is by carrying out a unit
root test; the Augmented Dickey – Fuller (ADF) test, for example.

If the data are non-stationary, the transformation is the first difference of ln(Xt) is
the daily rate of return Rt=ln(Xt/Xt-1). This is depicted in Fig. 4 below.

26
 GRAPH 4 .FULL PERIOD Rt

.12

.08

.04

.00

-.04

-.08
250 500 750

Rt

Fig. 4 appears to illustrate a stationary series with the trend around the mean.
However, The formal test to check for stationarity is the unit root, (ADF) statistic.
Series with a unit root are non-stationary. A first order autoregressive AR(1)
process with a unit root is defined as a random walk. In equation 7 below, the Xt
are optional exogenous regressors which may contain a constant, or a constant
and trend: ρ and δ are parameters to be estimated, and the εt are assumed to
be white noise. If |ρ| ≥ 1, Rt is a non-stationary series and the variance of Rt
increases with time and approaches infinity. If |ρ| < 1 , Rt is a stationary series.
Thus, the hypothesis of stationarity can be evaluated by testing whether the
absolute value of ρ is strictly less than one.

Rt = ρRt −1 + X ' δ + ε t (7)

Dickey and Fuller (1979) show that under the null hypothesis of a unit root, this
statistic does not follow the conventional Student's t-distribution, and they derive
asymptotic results and simulate critical values for various test and sample sizes.

27
The econometrics software package, E-views 4.0, gives the critical values of
this test at the 1%, 5% and 10% levels. Note that, when the t-value statistic is
greater than the critical value, we accept the null at the chosen level of
significance.

Furthermore, as we commented earlier on, there is the possibility of a structural

change in the CSE on May 13th 2004, the day that the CSE had the largest fall
in the last 5 years. This could be tested by carrying out a Chow breakpoint test.
The Chow test fits the equation separately for each sub sample and shows
whether there are significant differences in the estimated equations, or not.

To carry out the test, we partition the data into two or more sub samples, in our
case, pre-fall and post- fall respectively. Each sub sample must contain more
observations than the number of coefficients in the equation in order for the
equation to be estimated. The Chow breakpoint test compares the sum of
squared residuals obtained by fitting a single equation to the entire sample with
the sum of squared residuals obtained when separate equations are fit to each
sub sample of the data. An F-statistic is used to test the null hypothesis of no
structural change. Hence, a significant F-statistic indicates a structural change
in the relationship.

4.2.1. Random Walk

The concept of an efficient market was developed in the finance literature by

authors such as Fama(1970) who introduced the concept of efficient market
hypothesis, which is closely related to the data following a random walk. This
hypothesis asserts that movements in daily rates of return will not follow any
patterns or trends and that past price movements cannot be used to predict
future price movements.

Firstly, in order to test the random walk hypothesis that states that the size and
direction of price changes are random with respect to the knowledge available
in any point of time, the variables ln(Xt) and Rt must be examined for the

28
presence of unit roots by calculating ADF statistics. However, as is well
documented in the literature, the natural logarithms of stock prices (or index
values) may well be expected to follow simple random walks.

If this is the case, the series Rt may be examined further by using ordinary least
squares to estimate the equation:

Rt =α + µt

Under the random walk hypothesis, the constant term should be insignificantly
different from zero and the resultant residuals should be IID.

4.2.2. Randomness

The existence of a random walk in a series necessarily excludes the possibility

of non–linear structure in the data; if the first difference of the series is
stationary, the next step is to examine the series to see if its elements are IID.
The BDS test can be utilised here since its null hypothesis is that the data under
examination is IID.

The BDS statistic was proposed by Brock, Dechert and Scheinkman(1987) and
can be applied to a series of estimated residuals to check whether they are IID.
The distribution of the statistic, which is defined by:

n{C m (ε ) − C1 (ε ) m }
W m (ε ) (6)
σ (ε )

is asymptotically N(0,1) under the null. W m(ε) is known as the BDS statistic and
σm(ε) is an estimate of the standard deviation under the null hypothesis of IID.
(Chappell,D(1997)).

29
The test is carried out using the E-views 4.1 econometric package with
embedding dimensions, m, from 2 to 6 and distance, ε, between 0.5 and 2 times
the standard deviation of the data. Chappell et all (1998) claim that this range of
values is suitable in as much as it avoids the situation where ε is too small and
no m-histories are within ε of each other, or too big and so that all m-histories
are within a distance ε of each other.

Consequently, if the residuals appear to be non-IID as indicated by the results

of the BDS test, then there is a possible inefficiency in the market and this may
be our starting point to test for the presence of the a daily calendar anomaly in
the CSE.

4.2.3. Univariate time series

The foregoing discussion suggests as a first step a test for whether the variable
Rt in either of the sub-series is IID or not. If one, or both, of the series is not IID,
we should fit a linear ARMA model to the particular sub-set of the data using
the well-known general-to–specific (GTS) methodology to decide the order of
the ARMA. The (GTS) methodology was introduced by D.F. Hendry (2001)3 as
“the concept of general-to-specific modelling: starting from a general dynamic
statistical model, which captures the essential characteristics of the underlying
data set, standard testing procedures are used to reduce its complexity by
eliminating statistically insignificant variables, checking the validity of the
reductions at every stage to ensure the congruence of the selected model”.

The correlogram may be used to obtain an approximation to the number of lags

that should be included in the model in order to get a good fit to the data.
Hence, the final model after such reductions contain only significant lagged
variables and the residuals should be saved and tested for IIDness by means of
the BDS test. As discussed above, embedding dimensions m from 2 to 6 and a

3
H.-M. Krolzig, D.F. Hendry,” Computer automation of general-to-specific model selection procedures” Journal of
Economic Dynamics and Control, Vol 25, pag. 831-866 (2001)

30
range of values for ε from 0.5 to 2 times of the standard deviation of the series
should be used for the BDS tests. If this model does not give IID residuals, then
we will proceed as follows.

Our next step is to fit the model again, but now also including the four dummy
variables for Tuesday to Friday, Monday’s return is captured by the constant in
the regression. Firstly, however, a simpler methodology, as summarised in the
following equation, is commonly used for testing for a day-of-the-week effect.
This model is as follows.

Rt= c + β2D2 + β3D3+ β4D4+ β5D5 +µt

The regression is run to test for differences among returns on the trading days,
where the four D’s are dummy variables (one for each weekday from Tuesday
to Friday) that take the value of 1 for the corresponding day-of-the-week and
zero otherwise e.g. D2=1 if it is Tuesday and 0 if not; and µt is the zero mean
stochastic disturbance term. β1 captures the mean return for Monday and β2, β3,
β4 and β5 are parameters used to estimated means returns for Tuesday to
Friday respectively. If the day-of-the-week effect is to exist, at least two of these
coefficients must be statistically significant and different from one another.

Under the null hypothesis of no day-of-the-week effect; β2= β3= β4= β5 = 0,

residuals should be IID random variables. This will be tested for in the empirical
results with standard Ordinary Least Squared (OLS) and using the ANOVA
process to test the significance and equality of mean returns. However, the daily
stock returns are very likely to violate conventional assumptions about the OLS
error terms. Chang et al. (1993), focussed their attention on the weakness of
this method to estimated day-of-the-week effects, hence the statistics should be
calculated under the assumption that regression errors are homoskedastic,
serially uncorrelated and normally distributed. Thereby, they demonstrate - as in
Connolly´s contributions - (1989), (1990) for the United States market, that in
foreign markets the usual daily stock returns error assumptions are violated.

31
If the results from this model indicate that the residuals are still not IID, then the
previous ARMA model including the (GTS) methodology should be repeated,
but should now also include the four dummy variables for Tuesday to Friday.
When the GTS methodology gives a reduced model in which each of the
remaining variables is statistically significant, the residuals should be saved and
test for IIDness by again using BDS statistics.

If the BDS statistics are still significant, it is appropriate at this stage to fit a non-
linear model. However, in order to fit an appropriate non-linear model, it should
be noted that there are various different models that may well explain the
expected non-linearity of the data generating process: Examples include
Threshold auto regression models (TAR), generalised autoregressive
conditional Heteroscedasticity (GARCH) models and the Smooth transition
autoregressive models.

Kiymaz and Berument (2003) identify some problems in the standard OLS
approach. Firstly, the model could have misleading inferences resulting from
autocorrelations. When autocorrelation of any order is found, the OLS estimates
are unbiased but they are inefficient and we cannot rely on the standard errors
because they are not correctly estimated. Nevertheless, they propose a solution
that addresses the autocorrelation problem: include lagged values of the return
variable in the equation. Secondly, there is a problem when the error variances
may not be constant over time; this is known as heteroscedasticity. The direct
consequences are on the variability of the errors, which may depend on the size
of (one of) the exogenous variables and the standard errors of the OLS
estimates will be biased downwards. The estimator is not efficient; one
estimator is defined to be more efficient than another if it has lower variance.
Gujarati (2003) argues that heteroscedasticity may well be a problem where
important explanatory variables are omitted from the model.

If the null Hypothesis of no ARCH effects in residuals is rejected, standard OLS

estimates are not invalidated, but it may well be the case that more efficient
estimators exist. In spite of that, If the null hypothesis of IID cannot be accepted

32
in the previous ARMA model, the implications is that the residuals may well
contain some non-linear structure.

Chappell and Padmore (1995) have drawn attention to the fact that there is
strong evidence for the presence of conditional heteroscedasticity in exchange
rate innovations. However, there are several different types of model for
conditional variances suggested in the literature review. See, for example,
Engle (1982), Bollerslev (1986), Engle, Lilien and Robins (1987) and Bollerslev,
Engle and Wooldridge (1987).

The importance of such models in the investigation of a day-of-the-week effect

is in the conditional variance of the CSE returns. Such models are also capable
of capturing the three most empirical features observed in stock returns data:
leptokurtosis, skewness and volatility clustering. Also, according to Connolly
(1989), the problem created by the fat tailed distributions described above, is
that test statistics based on non-robust standard error estimates cannot be
interpreted in the usual way. The GARCH class of models are capable of
dealing with non-normal error terms and make the interpretation of the t-
statistics more robust.

The GARCH (p,q) model that is applied to the study of the day-of-the-week
effect on the CSE returns and volatility is as follows:

Rt = β 1 + β 2 D2t + β 3 D3t + β 4 D4t + β 5 D5t + ε t

ε t / Ψt −1 ~ t.d .(0, ht , v) (8)

p α
ht = c + γ 2 D2t + γ 3 D3t + γ 4 D4t + γ 5 D5t + ∑ δ j ht − j + ∑ α j ε t2− j
j =1 j =1

Where Rt is the stock return considered to be linearly related to a vector of

explanatory dummy variables (Dt) and an error term (εt) which depends on past
information ( Ψt −1 ); ht is the conditional variance. The error terms are assumed

to follow a conditional student-t density (t.d.) with v degrees of freedom. The t-

distribution approaches a normal distribution with variance (ht) as 1/v

33
approaches zero. Choudhry (2000) comments that the error distribution may
well be conditionally heteroscedastic and non-normal. Furthermore, Connolly
(1989) finds that this feature very useful because the unconditional leptokurtosis
may be traced to non-normality in the conditional error distribution and to time
varying heteroscedasticity.

The coefficients β1 to β5 in equations (8) represent the size and the direction of
the effect of each working day of the week on stock returns. Similarly,
coefficients γ1 to γ2 represent the size and direction of the day-of-the-week
effect on volatility. Some of the previous studies have found that the coefficient
on the Monday dummy (β 1) should be negative and significantly different from
zero, and the coefficient on the Friday dummy (β5) should be positive. In
contrast, the volatility coefficient for the Monday dummy (γ1) is positive but the
Friday effect (γ5) is negative.

If the standardised residuals from this model are still not IID, our methodology
suggests that the GARCH model should be fitted again using all the lags in Rt
and the dummy variables and again following the (GTS) methodology. One
hopes that this model will result in a good fit with all the coefficients in the mean
and variance equations statistically significant. The standardised residual series
should be saved and BDS tests then carried out to check for any remaining
unexplained structure. The standardized residuals should become more and
more conservative, while the size distortion correction provided by the
transformation on the residuals improves and there is no clear dominant
measure of effective sample size. Thus, if these residuals turn out to be IID,
then this final model will be used to derive the equations for the day-of-week
effect.

34
 CHAPTER 5.

EMPIRICAL RESULTS

Table 4 reports the preliminary statistics for returns for each day of the week on
the CSE. The first column of table 4 reports the daily mean, standard deviation,
Skewness, and kurtosis. The second through sixth columns of table 4 show the
same measures for each day of the week.

TABLE 4. SUMMARY STATISTICS FOR DAILY CSE RETURNS

Statistics All Days Monday Tuesday Wednesday Thursday Friday

Observations 906 157 188 187 186 188
Mean 0,170 0,199 -0,0348 0,089 0,1773 0,4305
Std. Dev 1,1941 1,127 1,2035 1,067 1,1702 1,3368
Skewness 0,427 -0,182 -1,0051 -0,0046 -1,2727 3,0192
Kurtosis 11,623 5,829 5,8213 4,2774 9,3372 19,5118

The average return for the entire study period is 0,17 percent, the standard
deviation is 1,194, and the skewness is 0,427. The kurtosis is 11,623, which is
significantly larger than the value of 3 for a normal distribution. The Jarque-Bera
normality test rejects normality of returns for each of the five days; see
Appendix 1.

When the returns for each day are examined, the findings indicate that Friday
has the highest mean return with 0,4305, while Tuesday has the lowest mean
return with -0,034. The signs of the findings are in line with some of the day of
the week effect literature (Solnik & Bousqet 1990, Dobois & Louvet 1996, Jaffe
& Westerfield 1985).

35
Table 4 also reports standard deviation, skewness, and kurtosis for each day.
Friday has the highest standard deviation of 1,3368 and Wednesday has the
lowest with 1,067. Moreover, all sample distributions are negatively skewed,
indicating that they are non-symmetric. They also exhibit high levels of kurtosis,
indicating that these distributions have thicker tails than a normal distribution.
Thus, the CSE daily returns are clearly not normally distributed, as indicated by
the Jarque - Bera statistics for each of the five days.

The natural logs of the CSE data, Ln(Xt), and the rate of the return (Rt) were
plotted in the previous section. However, in order to ascertain the stationarity of
these series, ADF tests will be carried out.

The results of the ADF tests for the two series are given in Appendix 2. The null
hypothesis that the series Ln(Xt) has a unit root is accepted at the 1% and 5%
significance levels but is rejected at the 10% level. For the returns series, Rt, the
null hypothesis of a unit root is rejected at all significance levels. Hence, ln(Xt) is
a non-stationary, I(1), variable and Rt is stationary, (I(0)). So in what follows,
this study will analyse the rate of return, Rt.

However, on May 13th 2004 the CSE suffered the largest one-day fall in the last
5 years. The existence of a structural break on that date was tested by carrying
out a Chow breakpoint test. The results are given in Table 5 below.

TABLE 5.CHOW TEST FOR A STRUCTURAL CHANGE IN RT

Chow Breakpoint Test: 702, May 13th 2004

F-statistic 3.430468 Probability 0.002355
Log likelihood ratio 20.62265 Probability 0.002144

36
The null hypothesis for the test in Table 5 is that there is not a structural break.
The test result shows quite clearly that we must reject the null and conclude that
there was indeed a structural break on May 13th 2004. The consequence of this
is the necessary subdivision of the series into two, “pre-fall” and “post-fall”,
series in order to be able to construct an adequate model for this data.

Following the results of the break-point test, the data was split into two
subgroups corresponding to ‘before’ and ‘after’ the structural break. ADF tests
were carried out on the Rt data for the two sub-groups. The results are given in
Appendix 3 and they show quite clearly that Rt is stationary in both sub-periods.

Our investigation proceeds by examining whether the CSE is an efficient

market. If the market is efficient, the series ln(Xt) will follow a random walk and
the Rt will be a purely random IID series. This may be tested for by means of
BDS statistics. Under the null of IID, BDS statistics are distributed as N(0, 1).
Statistics were calculated for each sub-sample for a range of values of M (the
embedding dimension) and for critical distances ranging from 0.5 to 2 standard
deviations of the data. The results are given in Table 6.

TABLE 6. BDS TESTS FOR RT

6.1. Pre-fall period -BDS statistics for Rt

M
values 2 3 4 5 6
0,5 10,4133 12,7906 14,3016 16,9029 20,7837
1 9,8217 11,2709 12,1023 13,5370 14,7856
1,5 9,1697 9,8059 10,1873 10,7891 11,0292
2 8,5956 8,2364 7,9501 8,1540 7,6353
Note: all significant at the 5%

6.2. Post-fall period -BDS statistics for Rt

M
values 2 3 4 5 6
0,5 5,5952 6,4190 7,1364 8,7359 9,3053
1 4,9453 5,6232 5,7906 6,6529 6,9868
1,5 3,5637 3,4354 3,4054 4,2764 4,6067
2 2,7711 1,5774* 1,1499* 2,1371** 2,6354
Not significant at the 1% level* and 5% level** respectively

37
The overwhelming evidence for the pre-fall period indicates the Rt are not IID,
and it follows that the CSE was not an efficient market during this period. For
the post-fall period the evidence is mixed. We accept the null for embedding
dimensions 3 & 4 and distance 2 standard deviations at the 5% level and we
accept it at the 1% level for M = 5 and distance 2 standard deviations. We reject
the null for other combinations of embedding dimension and distance. Faced
with this evidence we will, on balance, accept the null for the post-fall period
and conclude that there is some evidence that the market was efficient during
this period. Our attention will now focus exclusively on the pre-fall period.

We proceed by fitting a linear ARMA model using General-to-Specific (GTS)

methodology that ultimately gives a reduced model in which each of the
remaining variables are statistically significant. We then test the residuals from
this model for to ascertain whether they are IID by means of BDS statistics. All
these results are given in Appendix 4. Firstly, the correlogram is used to
indicate the number of lags that we should include in the model. At a first
glance, the graph shows (see appendix 4.1) that just two lags should be
included in the model. However, we decided to fit a more general model which,
following GTS methodology, results in the ARMA (5,4) given in Appendix 4.2.
Furthermore, the BDS statistics (Appendix 4.3) on the saved residuals show
that the residual are still not IID. So our next step is to test for the presence of a
daily calendar anomaly in the CSE during the pre-fall period.

The previous ARMA model is now re-estimated with the four dummy variables
added. After applying the GTS procedure, the fitted model is ARMA(4, 4). The
results are given in Appendix 5.1. The residual are saved and tested for IIDness
and the results are given in Appendix 5.2. However, the BDS statistics are still
significant, and we again have to conclude that the residuals are not IID and the
model still contains some unexplained component. It is now clear that a non-
linear model is needed.

38
Appendices 6 and 7 report the results of tests for heteroscedasticity and ARCH
effects in the residuals of the lineal OLS model. It is clear that we must reject
the null hypotheses of homoscedastic disturbances and no ARCH effects. So
our next step will be to fit a GARCH model with the daily dummies in both the
mean equation and the variance equation. However, we will first estimate, for
comparison purposes, an OLS model containing just the daily dummies and
lagged Rt.

Table 7 below reports on the day of the week effects and stock market volatility
during the pre-fall period, using three different models: Standard OLS,
GARCH(1,1) and Modified GARCH(1,1) using GTS methodology. The first
column reports the results from the OLS estimation.

The results of the OLS estimation show that Tuesday has the lowest return
(-0.000814), while Friday has the highest return( 0.003552). This result is
consistent with the previous finding reported in table 4. We include a lag value
of order one on the return variable to the equation to minimise the possibility of
having autocorrelated errors. This improved the estimation with a Durbin-
Watson statistic closer to 2.

One of the assumptions made up until now is that the errors, ut in the regression
equations have a common variance σ2. Using the White test allows us to
discover whether errors are heteroscedastic. Under the null the errors are
homoscedastic. However, in Appendix 6, the null hypothesis is rejected at the
1% level of significance, thus the errors are heteroscedastic. Therefore, a
GARCH model is felt to be appropriate.

The second column of Table 7 reports the result of the GARCH(1,1) estimation.
In this estimation, we allow the time varying conditional variance to follow a
GARCH(1,1) specification and the model also reports the estimates of the
dummy terms that are included for each day. The results indicate that the
highest return is observed on Friday (0.002035) and the lowest return is on
Wednesday (-0.000349). Monday has the second highest rate of return

39
(0.000931), followed by Thursday (0.000476). Nevertheless, the entire set of
coefficients are all statistically insignificant.

Additionally, the lowest volatility is observed on Tuesday (-0.00006.89) and the

highest volatility on Monday (0.0000249) after controlling the persistence effect
p
with the lag values of the conditional variance ∑δ
j =1
j ht − j and the squared lag

α
values of the residual term ∑α ε
j =1
j
2
t− j . All the day-of-the-week dummy variables

in the conditional variance equation are statistically significant.

40
 TABLE 7. DAY OF THE WEEK EFFECT AND STOCK MARKET
VOLATILITY IN THE CSE
GARCH
OLS GARCH(1,1)
modified
Return
Monday(B1)
Equation 0.000384 0.000931 0.001183**
(0.001035) (0.001196) (0.000414)
Tuesday(B2) -0.000814 -0.001179 -0.001633*
(0.001394) (0.001268) (0.000737)
Wednesday(B3) 0.000208 -0.000349
(0.001399) (0.001357)
Thursday(B4) 0.000940 0.000476
(0.001402) (0.001284)
Friday(B5) 0.003552* 0.002035 0.001806*
(0.001391) (0.001603) (0.000993)
Rt-1 0.248985** 0.326817** 0.311568**
(0.038073) (0.036238) (0.036466)

Volatility
Monday(γ1)
Equation 2.49E-05** 2.34E-05**
(4.17E-06) (3.58E-06)
Tuesday(γ2) -6.89E-05** -4.87E-05**
(7.83E-06) (8.55E-06)
Wednesday(γ3) -2.35E-05** -3.18E-05**
(8.15E-06) (6.82E-06)
Thursday(γ4) -3.20E-05** -2.12E-05**
(4.61E-07) (6.47E-06)
Friday(γ5) 2.27E-05**
(7.39E-06)
ARCH(α) 0.118029** 0.108418**
(0.014467) (0.013170)
GARCH(δ) 0.871827** 0.890384**
(0.013088) (0.010862)

Durbin-Watson 1,9281 2,0637 2,03890

Prob(F-statistic) 0.000000 0.000004 0.000000
Notes: Standard errors are reported under the corresponding estimated coefficients.
** and * indicate the level of significance at the 1 percent and 5 percent level, respectively.

41
The sum of the coefficients of the GARCH equation without a constant term is
less than one, and both of them are positive and statistically significant. Hence,
we do not have either negative or explosive implied variances as suggested by
Bollerslev(1986) for the specification test. On the other hand, since the
summation of these two coefficients is close to one, it indicates that the volatility
is persistent. However, this second model does not reach the expectations in
terms of the return, the standard error remain almost the same.

The estimation results for the third specification of the model are reported in the
third column of Table 7. We apply GTS methodology to the model in column 3.
Following with the (GTS) methodology, the insignificant variables are deleted
one-by-one, until everything remaining is significant. This third model, which we
call Modified GARCH(1,1), also reports the estimates of the five dummy terms
that are included for each day.

When the Modified GARCH (1,1) is estimated, Tuesday has the lowest rate of
return ( -0.001633), and Friday has the highest rate of return ( 0.001806), both
statistically significant at the 5% level. Clearly, using the (GTS) methodology
and allowing time varying variance in the estimation process provides more
efficient estimates for the return. In contrast, this can be easily observed with
the lower standard errors for the estimated parameters of the return equation. In
sum, the third model specification increases the efficiency of the estimates
compared to the previous models.

The lowest volatility is observed on Tuesday (-0.0000487) and the highest

volatility is on Monday (0.0000234), a significant positive effect implies that
stock return volatility is increased, although the size of the coefficient (γ1) is very
small. A significant positive influence by Monday on volatilities provides some
evidence in favour of the information oriented theories of French and Roll(1986)
and Foster and Viswanathan (1990). In the case of Tuesday, a significant
negative effect is found. This indicates that volatility is reduced on Tuesdays, in
line with some of the findings in Solnik & Bousqet (1990), Dobois & Louvet
(1996), and Jaffe & Westerfield (1985).

42
For all five days with the exception of Friday, dummy variables in the
conditional variance equation are statistically significant. Consistent with the
second model, GARCH coefficients for these two parameters sum to less than
one, but quite close to one as reported in the former model. In addition, they are
both positive and statistically significant. Results from the GARCH model, OLS
and mean returns based on the day of the week are not identical, but are
similar.

In Appendix 7.1, the autocorrelation Q statistics are reported for the OLS
specification at 30 and 35-day lags. The coefficients are statistically significant
at the 10% level. Nevertheless, in the GARCH models, we cannot reject the null
hypothesis that the residuals are uncorrelated. On the other hand, the ARCH-
LM test in Appendix 7.2 does not indicate the presence of a significant ARCH
effect in any of the GARCH models. These findings indicate that the
standardised residual terms have constant variance and do not exhibit
autocorrelation.

In the final step, the results provide ample evidence of a day-of-the-week effect
in stock market volatility. Table 9 reports the results of applying the BDS test to
the standardised residuals in order to determine whether any remaining non-
linear structure was present.

TABLE 9.
Pre-fall period -BDS test on GARCH (1,1)
standardised residuals

M
values 2 3 4 5 6
0,5 2.073702* 2.773529 3.036908 3.622360 3.861715
1 1.802251* 2.622949* 3.093071* 4.101888 4.521770
1,5 0.998249* 1.507987* 1.944140* 2.844155 3.001099
2 0.250216* 0.182273* 0.645599* 1.512018* 1.297233*
Note: *Not significant

43
Under the null hypothesis there are strong indications that for the pre-fall series
all non-linear structure has been removed and the residuals are IID. This result
indicates that the Modified GARCH(1,1) model can adequately describe the
daily return process of the CSE stock price index in the pre-fall period.

44
 CHAPTER 6.

CONCLUSIONS

The day of the week findings appear to conflict with the Efficient Market
Hypothesis. Hence, any predictable pattern in stock returns and variances may
provide investors with returns in excess of the stock market average. This study
indicates that there is a substantial day-of-the-week effect in the Colombia
Stock Exchange between June 2001 and March 2005 based on our statistical
analysis of the CSE stock index.

In 2004, the CSE was the best investment opportunity in Colombia,

experiencing a sharp increase in both value and size. Moreover, the Index
exhibited several historically high values. The total value in 2004 was around
COP$4.77 billions, a massive increase from the COP$1.89 billions in 2003.

Empirical findings show that the day-of-the-week effect is interrupted by a

structural change in the CSE on May 13th 2004, the day that the CSE had the
largest fall in the last 5 years. So the series was split and analysed in two sub-
groups to achieve more accurate results. However, we found some evidence
that for the post-fall period the market was efficient, whilst the pre-fall was
successfully explained by a (non-linear) GARCH model.

Three different models were considered. The first, which assumes the
constancy of the residual term’s variance, is the standard OLS. However, all

45
sample distributions are negatively skewed, indicating that they are non-
symmetric. They also exhibit high levels of kurtosis, indicating that these
distributions have thicker tails than a normal distribution. Thus, the CSE daily
returns are clearly not normally distributed. The findings based on this model
indicate a flawed evidence of the day-of-the-week effect in the return equation
with Friday with highest return and unique significant dummy variable. In the
second model, a GARCH (1,1), volatility changes over time, and the capability
of this model to deal with non-normal error terms makes the interpretation of the
t-statistics more robust. The results are similar to those of the first model with
most of the coefficients statistically insignificant.

Finally, the third model employed is the modified GARCH(1,1), in which we

apply general-to–specific (GTS) methodology, which captures the essential
characteristics of the underlying data set but eliminates all statistically
insignificant variables one-by-one.

Under the null hypothesis of IID residuals, there are strong indications that for
the pre-fall series all non-linear structure has been removed and the residuals
are IID. This result indicates that the Modified GARCH(1,1) model can
adequately describe the day-of-the-week effect in the CSE stock price index in
the pre-fall period.

Tuesday has the lowest rate of return ( -0.001633), and Friday has the highest
(0.001806), both statistically significant at the 5% level. Clearly, using the (GTS)
methodology and allowing time varying variance in the estimation process
provides more efficient estimates for the returns compared to the previous
models. The findings show that the day-of-the-week effect is present in both the
volatility and the mean equations.

Tuesday has the lowest significant day-of-the-week effect for the returns and
Friday has the highest. These findings are in line with some of the findings in
Solnik & Bousqet (1990), Dobois & Louvet (1996), and Jaffe & Westerfield
(1985). Oguzsoy and Guven (2003) also observe that for most of the stocks

46
among the 30 most heavily traded stocks in the Istanbul Stock Exchange, the
maximum return is on Friday whereas the minimum return is either on Monday
or Tuesday, with return variances at their highest on Mondays.

For Tuesday, non-synchronous trading may be an explanation for the one-day

lag. However, some correlation analysis was made among daily returns with a
one-day lag between Western and Eastern countries, where the correlation
between Monday returns in the US market and Tuesday returns in the
Japanese market is higher than other days of the week. In contrast, no
significant differences appear between US market and Australian market.
However, in the CSE, this theory is inconsistent; there are only a few hours of
time difference with the New York Stock Exchange.

Moreover, the significant day of the week effect in volatility is maybe in line with
the information availability theory. In the CSE variance equations, Monday has
the highest variance, with 0.0000234, according with French and Roll (1986)
and Foster and Viswanathan (1990), who point out that the stock return
variance should be highest on Monday, when the informed trader has the
maximum information advantage. The variance should decline through the
week with the arrival of publicly available information. Chaudhry (2000) also
points out, that in a situation when private information is available throughout
the week while public information is available only during weekdays, traders are
more sensitive to changes in order flow at the beginning of the week.
Consequently, the variance of prices changes would be higher at the beginning
of the week than at the end of the trading week.

The non-synchronous trading time zones and the information availability theory
are some possible explanations for the day-of-the-week effect on the Colombian
Stock Exchange, but clearly indicate the necessity for further research in stock
markets; e.g. distinction between trading and non-trading periods, settlement
periods, size of the companies, measurement error, etc.

47
In the financial field, the existence of a day-of-the-week effect seems to be a
widespread and well-accepted phenomenon. However, there is a question
remaining about the real possibility of enjoying abnormal returns, given the fact
that there is nothing that guarantees the permanence of these anomalies in the
future.

Overall, results indicate that, during the 1980´s, this type of calendar anomaly
was clearly evident in the vast majority of developed markets, but it appears to
have faded away somewhat in the 1990´s. Moreover, Connolly (1989) shows
that when transactions costs are taken into account, the probability that
arbitrage profits are available from weekend-oriented trading strategies seems
very small.

The Colombia Stock Exchange is classified as an emerging market. Therefore it

is plausible that we may uncover some market inefficiencies that investors could
exploit as a trading strategy to benefit from these seasonal regularities. As a
final conclusion it is important to remark that accurate results will suggest the
possibility of a trading strategy in order to take advantage of the market
inefficiency, buying and selling strategies formulated accordingly to increase
returns due to better timing. (For example, buy on Tuesday, sell on Friday).

48
 BIBLIOGRAPHY

• Adcock, Christopher J (2000), “Fat tails and the capital asset

pricing model” from Dunis, Christian, ed. Advances in quantitative asset
management. Kluwer Academic, Chapter 2

• Alexakis, P. and Xanthakis, M. (1995), “Day of the Week Effect on

the Greek Stock Market”, Applied Financial Economics, Vol. 5, pp. 43-50.

• Al – Loughani and Chappell, (2001) “Modelling the Day – of – the

– week effect in the Kuwait stock exchange: a Nonlinear GARCH
representation”. Applied Financial Economics, 11, 353 – 359.

• Bollerslev, T. (1986), “Generalized autoregressive conditional

heteroscedasticity”, Journal of Econometrica, 31, 307-327, 1986.

• Bollerslev, Tim, Robert F. Engle, and Jeffrey M. Wooldridge

(1987), “A capital asset pricing model with time varying covariances”,
Journal of Political Economy 96, 116-131.

• Brock, W.A., Dechert, W.D. and Scheinkman, J.A. (1987) “A test

for independence based on the correlations dimensions”, Social Systems
Research Unit, University of Wisconsin

• Chang, E.C., J.M. Pinegar, and R. Ravichandran, “International

evidence on the robustness of the day-of-the week effect”, Journal of
Financial and Quantitative Analysis, 28(4), 497-512, 1993.

• Chappell, D., Padmore, J., Pidgeon, J., (1998) “A note on ERM

Membership and the efficient on the London Stock Exchange” Applied
Economics Letters 5,19-23.

49
• Chappell, D. and Padmore, J.(1995), “Changes in Volatility of the
Sterling-Deutschmark Exchange Rate: the Effect of ERM Membership”,
Applied Economics Letters, Vol. 2, 291-294

• Chappell, D and Eldridge, R.M.(1996), “Non-linear characteristics

of the Sterling/European currency unit exchange rate:1984 – 1992”, The
European Journal of Finance, 3,159 – 182.

• Chappell, D (1997)” The BDS test for randomness”,Encyclopedia

of Statistical sciences; Update volume 3.

• Connolly, R. (1989), “An examination of the robustness of the weekend

effect”, Journal of Financial and quantitative Analysis, 24(2), 133-169.

• Chappell, D (1997)” The BDS test for randomness”,Encyclopedia

of Statistical sciences; Update volume 3.

• Choudhry, T.(2000), 2Day of the week effect in emerging Asian

stock markets: Evidence from the GARCH model”, Applied Financial
Economics, 10, 235-242.

• Cross, F., (1973), “The Behavior of Stock Prices on Fridays and

Mondays”, Financial Analysis Journal 29, 583-588.

• Dickey D. A.& Fuller W. A. (1979); “Distribution of the estimator for

autoregressive time series with a unit root”. Journal of the American
Statistical Association 74: 427 31

• Dubois and P. Louvet, (1996), “The Day-of-the-week Effect: The

International Evidence,” Journal of Banking and Finance 20, 1463-1484

• Engle, R. F. (1982). "Autoregressive conditional heteroscedasticity

with estimates of the variance of United Kingdom inflation",
Econometrica, 50, 987-1006

• Engle, R.F., David Lilien y Russell Robins (1987), “Estimating

Time Varying Risk Premia in the Term Structure, The arch-m Model”,
Ecomometrica 55, 391-407.

• Fama, Eugene F,(1970) “Efficient Capital Markets: A review of

theory & Empirical work,” Journal of Finance , 383 – 41

50
• Fama, Eugene F,(1991) “Efficient Capital Markets II”, Journal of
Finance 46, 1575 – 1617.

• Fama, Eugene and Kenneth French. (1992). “the Cross-Section of

Expected Stock Returns.” Journal of Finance 47:2, 427-465

• Fama, Eugene and Kenneth French. (1995). “Size and Book-to-

Market Factors in Earnings and Returns” Journal of Finance 50:1, 131-
155

• Fama, Eugene and Kenneth French. (1996). “Multifactor

Explanations of Asset Pricing Anomalies.” Journal of Finance 51:1, 55-84

• French, D. W., (1980), “Stock Returns and the Weekend Effect”,

Journal of Financial Economics 8, no.1, 55-69.

• French, K.R., G.W. Schwert and R.F. Stambaugh (1987),

“Expected stock returns and volatility”, Journal of Financial Economics
19:3−29.

• French, K. and R. Roll, (1986),” Stock return variances: The arrival

of information and the reaction of traders”, Journal of Financial
Economics 17, 5-26

• Foster, F.D., and S. Viswanathan , (1990), “A Theory of Interday

Variations in Volumes, Variances and Trading Costs in Securities
Markets”, Review of Financial Studies , 4, 595-624

• Gibbons, M. R. and P. J. Hess, (1981), “Day of the Week Effects

and Asset Returns”, Journal of Business 54, no. 4, 579-596.

• Gujarati Damodar ( 2003), Basic econometrics, Fourth Edition

• Jaffe, J. & Westerfield, R. (1985), “The weekend in common stock

returns: The international evidence”, Journal of Finance 40, 433–454.

• Kiymaz and Berument,(2003) “ The day of the week effect on

stock market volatility and volume : international evidence” Review of
Financial economics 12, 363 – 80

• Keim, D. B. & Stambaugh, R. F. (1983), “A further investigation of

the weekend effect in stock returns”,Journal of Finance 39, 819–835.

51
• Lakonishok, J and Levi, M, (1982), “Weekend effects in stock
returns: A Note. Journal of finance, 37, 883-89

• Malkiel, Burton. (2003). “Efficient Market Hypothesis and Its

Critics.” Journal of Economic Perspectives. Vol. 17, No.1, 59-82

• Osborne, M. F. M. (1962), ‘Periodic structure in the Brownian

motion of stock returns’, Operations Research 10, 345–379.

• Oguzsoy and Guven, (2003), “Stock returns and the day-of-the-

week effect in Istanbul Stock Exchange”. Applied Economics, 35, 959 –
971.

• Rogalski, R. J., (1984), A Further Investigation of the Weekend

Effect in Stock Returns.Discussion, Journal of Finance 39, no.3 835-837.

• Solnik, B. and L. Bousquet, (1990), “Day of the Week Effect on

Paris Bourse”, Journal of Banking and Finance 14, 461-468.

• Sullivan, R., Timmermann, A. & White, H. (2001), “Dangers of

data-driven inference: The case of calendar effects in stock returns”,
Journal of Econometrics 105, 249–286.

52
 APPENDICES

APPENDIX 1

25
Monday Series: MONDAYS
Sample 1 157
20 Observations 157

Mean 0.199400
15 Median 0.117515
Maximum 3.480829
Minimum -4.700497
10 Std. Dev. 1.126712
Skewness -0.182471
Kurtosis 5.829102
5
Jarque-Bera 53.22956
Probability 0.000000
0
-3.75 -2.50 -1.25 0.00 1.25 2.50

28
Tuesday Series: TUESDAY
24 Sample 1 188
Observations 188
20
Mean -0.034845
16 Median 0.060260
Maximum 2.859596
Minimum -4.864521
12
Std. Dev. 1.203504
Skewness -1.005162
8
Kurtosis 5.821351
4 Jarque-Bera 94.01117
Probability 0.000000
0
-5.00 -3.75 -2.50 -1.25 0.00 1.25 2.50

53
30
Wednesday Series: WEDNESDAY
Sample 1 187
25
Observations 187

20 Mean 0.089896
Median 0.055682
15 Maximum 4.003336
Minimum -3.359876
Std. Dev. 1.067035
10 Skewness -0.004610
Kurtosis 4.277471
5
Jarque-Bera 12.71613
Probability 0.001733
0
-2.50 -1.25 0.00 1.25 2.50 3.75

50
Thursday Series: THURSDAY
Sample 1 185
40 Observations 185

Mean 0.177334
30 Median 0.270717
Maximum 3.780647
Minimum -6.373320
20 Std. Dev. 1.170266
Skewness -1.272779
Kurtosis 9.337220
10
Jarque-Bera 359.5185
Probability 0.000000
0
-6 -4 -2 0 2 4

60
Friday Series: FRIDAY
Sample 1 187
50
Observations 187

40 Mean 0.430523
Median 0.238758
30 Maximum 8.898115
Minimum -2.522364
Std. Dev. 1.336873
20 Skewness 3.019217
Kurtosis 19.51184
10
Jarque-Bera 2408.432
Probability 0.000000
0
-2 0 2 4 6 8

54
APPENDIX 2

2.1. Unit root test for variable Ln (Xt)

t-Statistic Prob.*
Augmented Dickey-Fuller test statistic -3.247281 0.0760
Test critical values: 1% level -3.968211
5% level -3.414782
10% level -3.129555
*MacKinnon (1996) one-sided p-values.

2.2. Unit root test for Rt

t-Statistic Prob.*
Augmented Dickey-Fuller test statistic -23.81719 0.0000
Test critical values: 1% level -3.968211
5% level -3.414782
10% level -3.129555
*MacKinnon (1996) one-sided p-values.

APPENDIX 3

3.1. Unit root test for Rt Pre–fall

t-Statistic Prob.*
Augmented Dickey-Fuller test statistic -19.76802 0.0000
Test critical values: 1% level -3.971217
5% level -3.416250
10% level -3.130425
*MacKinnon (1996) one-sided p-values.

3.2. Unit root test for Rt Post–fall

t-Statistic Prob.*
Augmented Dickey-Fuller test statistic -12.92353 0.0000
Test critical values: 1% level -4.003226
5% level -3.431789
10% level -3.139601
*MacKinnon (1996) one-sided p-values.

55
APPENDIX 4, ARMA model Rt variable

4.1. Correlogram Rt pre- fall

4.2. Rt pre- fall, ARMA(5,4)

Dependent Variable: RT
Sample(adjusted): 7 702
Variable Coefficient Std. Error t-Statistic Prob.
C 0.003009 0.001133 2.655938 0.0081
RT(-1) 0.223987 0.036272 6.175205 0.0000
RT(-2) -0.533678 0.029217 -18.26607 0.0000
RT(-4) -0.822902 0.026993 -30.48524 0.0000
RT(-5) 0.233104 0.038076 6.122136 0.0000
MA(2) 0.621838 0.024348 25.53972 0.0000
MA(3) 0.133768 0.027328 4.894965 0.0000
MA(4) 0.863936 0.024807 34.82648 0.0000
R-squared 0.086425 Mean dependent var 0.001541
Log likelihood 2137.277 F-statistic 9.297876
Durbin-Watson stat 1.897375 Prob(F-statistic) 0.000000

4.3.Pre-fall period -BDS test (ARMA (5,4)) residuals

M
values 2 3 4 5 6
0,5 9.179427 12.01744 13.67564 16.56853 21.01244
1 7.364463 9.180985 10.20432 11.79021 13.00783
1,5 6.511607 7.468624 8.095228 9.181802 9.468418
2 5.762196 5.973314 6.186023 6.997851 6.547275
Note: all significant at the 5%

56
APPENDIX 5

5.1 .ARMA(4,4) model Rt variable and dummy variables

Dependent Variable: RT
Sample(adjusted): 6 702
Variable Coefficient Std. Error t-Statistic Prob.
C 0.003002 0.000806 3.723666 0.0002
D2 -0.003413 0.000690 -4.948837 0.0000
D3 -0.003012 0.000565 -5.334700 0.0000
D5 0.002251 0.000450 4.997426 0.0000
RT(-1) -0.355702 0.077934 -4.564147 0.0000
RT(-3) 0.413716 0.072097 5.738344 0.0000
RT(-4) -0.411976 0.104781 -3.931775 0.0001
MA(1) 0.613081 0.092963 6.594875 0.0000
MA(2) 0.233085 0.045006 5.178960 0.0000
MA(3) -0.342225 0.072664 -4.709700 0.0000
MA(4) 0.297492 0.109317 2.721373 0.0067
R-squared 0.109847 Mean dependent var 0.001541
Log likelihood 2149.899 F-statistic 8.465391
Durbin-Watson stat 1.905867 Prob(F-statistic) 0.000000

5.2.Pre-fall period -BDS test (ARMA (4,4)) residuals

M
values 2 3 4 5 6
0,5 7.956710 9.843892 11.55930 13.88170 16.00100
1 7.155499 9.048985 10.39730 12.14669 13.34134
1,5 6.729682 7.760945 8.437321 9.651089 9.941559
2 6.440150 6.462428 6.674425 7.462068 6.936580
Note: all significant at the 5%

APPENDIX 6. White Heteroskedasticity Test on the OLS model

White Heteroskedasticity Test:

F-statistic 29.88064 Probability 0.000000
Obs*R-squared 143.8737 Probability 0.000000

57
APPENDIX 7

7.1 Autocorrelation Q statistics

GARCH( Modified
Lags OLS
1,1) GARCH(1,1)
5 6,3010 3,6618 4,3905
(0.278) (0.599) (0.495)
10 10.815 7,5722 9,4989
(0.372) (0.671) (0.485)
15 17.496 14.892 16.305
(0.290) (0.459) (0.362)
20 24.677 21.356 22.938
(0.214) (0.376) (0.292)
25 33.026 25.484 26.813
(0.130) (0.436) (0.365)
30 43.632* 37.715 38.741
(0.051) (0.157) (0.132)
35 46.929* 42.290 43.214
(0.086) (0.185) (0.134)
Note: p values in parenthesis,* statistically significant
at the 10 percent level

7.2.ARCH LM
GARCH( Modified
Lags OLS
1,1) GARCH(1,1)
5 22,5857** 0.1893 4,3905
(0.0000) (0.9666) (0.495)
10 11,288** 0.2566 9,4989
(0.0000) (0.9897) (0.485)
15 7,4186** 0.2518 16.305
(0.0000) (0.9983) (0.362)
20 5,5015** 0.2471 22.938
(0.0000) (0.9997) (0.292)
25 4,5146** 0.2516 26.813
(0.0000) (0.9999) (0.365)
Note: p values in parenthesis,** statistically significant
at the 1 percent level

58