Identifying the dependency pattern of daily change (increase or decrease)

of Dhaka Stock Exchange index (DSEX) in Bangladesh using Markov chain

and logistic regression model
Md. Arif Hassan
Assistant Professor, Department of Business Administration
Daffodil International University
Email: hassan@daffodilvarsity.edu.bd

Dr. Sayedul Anam

Assistant Professor, Department of General Educational Development
Daffodil International University
Email: anam.ged@daffodilvarsity.edu.bd

Md. Ashik Abdulla

Lecturer [Contractual], Department of Business Administration
Daffodil International University
Email: ashik.bba@diu.edu.bd

Md. Kamruzzaman
Senior Lecturer, Department of Business Administration
Daffodil International University
Email: kamruzzaman.bba@diu.edu.bd

Abstract:
Bangladesh capital market is third largest market in South Asia region which has two stocks
exchange namely Dhaka stock Exchange (DSE) and Chittagong Stock Exchange (CSE)
regulated by Securities and Exchange Commission (SEC). DSE introducing DSE broad index
(“DSEX”) and DSE 30 index (“DSE30”) that effect from 2013. In this study, DSEX which
reflect 97% of the total equity market capitalization is considered and focused the effect of
change of index on market investment. In connection to the Markov chain of different order,
logistic regression model is conducted to visualize the dependency of change (increase or
decrease) of current index upon the change of previous two time period. It had been shown that
increase day of index of the previous two time period compared to the decrease day of index
of previous two time period influences the increase day of index of current time period. We
observed the dependency of increase-decrease index of spell for the occurrence of index in the
Dhaka stock exchange from 28/01/2013 to 30/04/2019 in the study period. The test result shows
that the occurrence of change of index follows a second order Markov chain and logistic
regression also tells that decrease index day of index followed by decreased and increase day
of index followed by increased is more likely for the index of Dhaka stock exchange.

Keywords: DSEX index, Markov chain, Logistic regression.

Introduction:

Dhaka Stock Exchange started formal trading of the Exchange in 1956 after obtaining the
Certificate of Commencement of Business though it was incorporated in 1954(Annual report
2018). It is the largest capital market in the country. DSE has achieved full membership of the
World Federation of Exchanges Limited (WFE) on June 06, 2017. DSE is also a primary
member of South Asian Federation of Exchanges (SAFE). It has achieved the ISO ISO
9001:2008 in 2016 and upgraded to ISO 9001:2015 for excellence in Quality Management
System (QMS).

DSEX (DSE Broad Index) developed by S&P Dow Jones Indices, a renowned Index Service
Provider were introduced in the Exchange on January 28, 2013. DSEX is the base index or
benchmark index now for Dhaka Stock Exchange which has replaced the old DSE general
Index (DGEN). The new index is based on the free-float method used by the world’s major
indices. A Free-float index is calculated on the readily available share on the market rather the
full number of active and inactive shares available in the market. This method actually excludes
the locked-in shares thus better represents the overall ongoing scenario of the market.

The DSE Broad index has a base value of 2951.91 on Jan 17th 2008 (base date) which was the
index value of the DSE General Index on this date and the base. To be eligible for DSEX index
the stocks need to reach a float-adjusted market capitalization above 100 million BDT.
However, if the index for a stock falls below 100 million BDT but no less than 70 million with
other criteria met the stocks will remain in this index.

DSEX demands that stocks will maintain an ADVT (average daily value traded) of 1 million BDT
for minimum six-month as of the rebalancing reference date. Moreover, it is required that each
of the eligible stocks are traded at least half of normal trading days each month for the three
months prior to the rebalancing reference date.

DGEN, calculated in a wrong method, has been the most talked-about issue among the retail
investors over the last five years of its existence. The data used to calculate the index were not
as per international standard. However, the fault got assurance with the jump of 764 points in
one day after the Grameen phone had been enlisted on DSE. Instantly the BSEC demand
eliminate these inconsistencies under the surveillance of internationally reputed firm Standard
& Poor’s. Hence, the new benchmark index DSEX comes. Initially the index starts with 97%
of the total equity market capitalization and to maintain the standard it periodically reviews the
enlisted company in the early January 2019 the index has included 15 new companies and
excluded other 17 old ones that ends up with having 283 companies for the index. The index
is based on DSE Bangladesh Index Methodology which is supervised by a DSE panel to ensure
the satisfactory performance of the index. Moreover, a number of precautionary steps has been
taken to protect the investors interest -Firm rules for guidelines, international standard
surveillance, circuit breakers for trade and disclosure requirements for both listed scripts and
IPOs etc. (Rashid, 2008).
A major challenge of stock market for sustainable development is the volatility of price index
and its investors decisions are primarily affected with these changes. However, the volatility
of the market is also associated with the inefficiency of the governance of the market, (Rahman
& Golam Moazzem, 2011). Moreover, the speculative behavior of many investors works
against the development of trust and a sense of fairness in capital markets. The charges of
market manipulations, insider trading, and outright scams as well as the speculative nature of
the market can be a fetal obstacle to capital formation, and efficient functioning of the capital
markets (Krishnamurti et. al., 2003).

Literature Review:

Shleifer (1986) assumes that the index is information free and thus shows a clear test of demand
curve slopes. He found stocks that are newly added to index earn significant positive abnormal
returns that do not evaporate after ten days. As new stocks are not perfect substitutes, the
opening price pressure created by index is not overturned. Naveen (2017) observed new firms
that are added to the index causes significant price increases in the run-up window. Moreover,
trading volume effects in the run up and also post change windows enhanced liquidity are not
reversed in the post change period. Lynch (1997) found a positive cumulative abnormal return
found of 3.807% over the period starting the day after the announcement and ending the day
before the effective date of the change.

Huffman (1987) observed exogenous shocks give rise to successively correlated transaction
volume as portfolio decisions made in one period influence transactions in consecutive period.
Reasonable circumstances are consistent with autocorrelations in trading volume even without
having a well developed theory for multi period trade volumes. Ajinkya and Jain (1989)
showed experimental distributions of daily trading volume prediction errors for individual
firms and for portfolios by using volume measures and expectation models. The forecasted
errors for raw volume measures are significantly positively skewed, with thin left tails and fat
right tails. on the other hand, natural log transformations of the volume measures found
approximately normally distributed. However, for more than one day prediction intervals, auto
correlation in daily trading volume found beneficial for identifying abnormal trading.

Past evidence has shown that stocks included in (excluded from) an index demonstrates a
significant positive (negative) abnormal returns on the announcement day, and that trading
volume is affected by the event. Bildik and Gulay (2008) examined the price and volume effects
on stocks associated with the changes in the value weighted index composition of two indices
of the ISE where the index funds and index derivatives do not exist. The results are consistent
with previous evidence that stocks included in (excluded from) the index intend a positive
(negative) abnormal returns in ISE. Volatility and volume affected significantly in this regard.

Harris and Gurel (1986) stated in his price pressure hypothesis that due to the excess demand
of fund managers the stock prices increase before the change date and then reverse when
passive sellers are paying attention by the price increases that drive the prices above
equilibrium level. Therefore, this theory forecast that a newly added (deleted) stock causes a
temporary increase (decrease) in the price. To figure out the recent volatility with the
corresponding standard errors used as the indicator variable, LeBaron (1990) used Rt for
weekly returns of the Standard and Poor 500 index for the period 1946-1985, giving about
2,000 observations. It showed that the small portfolios have a considerably higher average
return than the largest. Moreover the highest E/P portfolios are better than the lowest.

To predict the tendency of stock price Sulin Pang (2004) applied Logistic Regression model.
Whereas, Zheng Mei and Miao Jia (2007) used Logit model to forecast the stock trend of
Shanghai stock market and made contrast with ARMA model. After that Gong and Sun (2009)
applied Logistic Regression to predict next month’s stock price trend just through considering
current month financial data instead of analyzing and collecting long term financial data. Based
on the daily precipitation data for 30 years (1956-1985) at 14 stations in South Korea, Moon,
Ryoo and Kwon (1993) used Markov chain model on daily rainfall occurrence; it is found the
daily precipitation is mostly dependent on the previous day. Similarly Hossain (2012) used
Markov Chain to find the change influence previous two day state on today state about rainfall
of Dhaka station. Anam (2018) applied Markov chain to identify the flood occurrence.
However, there is no such research found using Markov model in determining the dependency
of daily changes of stock price. Therefore in this paper authors try to identify the influence of
previous days index change on today index change and their dependency pattern to make
decision for the investors using both Markov and Logistic regression model.

Methodology:
Discrete Markov Chain Model:

P. Billingsley (1961) showed that the probability model is based on the assumption that the
state of any day depends only on the state of the preceding day. The dependence relation is
assumed as the first order dependence in which the result of one experiment is dependent on
the result of the previous experiment such that these transition probabilities are constants. So
the statistical model for studying the effect of dependence on this conventional procedure is a
two state discrete time Markov Chain. In recent years we founded some attempts to model such
dependence by m order Markov chains.

A Markov chain of order m is a sequence of trials of the outcome if each trial depends only on
that. According to the sequence of a random variables {Xn}forms a Markov chain of order m,
if given a fixed m, for all possible values of the variables Xn( n=0,1,2…..) it is true that

P[ X n  j / X 0  i0 , X 1  i1 ,........., X nm ]  P[ X n  j / X nm  inm ]

Multiple Logistic Regression Model:

Consider a collection of p independent variables which will be denoted by the vector

X’= (X1, X2 , X3 …….,Xp). For the moment we will assume that each of these variables is at
least interval scaled. The specific form of the logistic regression model is given below;
 p
0    i xi
i 1
e
P( X )  p
; i  1,2,...., p
0    i xi
1  e i 1
Then the multiple logistic regression modelis as follows;

f ( X )   0  1 x1   2 x 2  ........   p x p
e f ( x )i
in which case P( X ) 
1  e f ( x)

Data Collection:

Secondary data have been used for the research. The data is collected from Dhaka Stock
Exchange website (https://www.dsebd.org/). We have collected daily change index from
January 2013 to May 2019. We have collected from 2013 because from this year DSEX
indexing was introduced which is more significant to measure the change of index.

Analysis and Discussion:

Markov Chain:

According to Basu (1971), the explanatory variables are measured in different kinds of scale
but they are categorized in dichotomous form considering long past behavior of this
meteorological factors in Bangladesh. The dependent and independent variables are as follows;
the dependent variable index (Y) is a dichotomous one, it takes on the value 1 with probability
P (say) if the index increase and it takes on the value 0 with probability 1-P if the index
decrease. The independent variables are;

X1 = Yesterday’s index
= 0 if index decrease
= 1 if index increase

X2 = Day before yesterday’s index

= 0 if index decrease
= 1 if index increase
The transition counts for the first order Markov model are obtained by considering today’s and
yesterday’s index of DSE where index is considered as increase index and decrease index. The
following table-1 shows the frequencies of the first order transitions considering today’s
increase index and decrease index followed by yesterday’s increase index and decrease index.

Table -1: Frequency for first order transition counts.

 Yesterday’s Today’s state of index Total
state of
Index Decrease (0) Increase (1)
Decrease (0) 420 308 728
Increase (1) 307 470 777
Total 727 778 1505

From the above table, we see that the highest proportion(0.605) belongs to transition of the
form increase index day to increase index day and lowest proportion (0.395) belongs to
transition from increase day to decrease day. Table-2 gives the maximum likelihood estimates
of transition probabilities for the first order Markov chain obtained directly by using transition
count by the formula;
nij
Pij  , ni   nij and Pij  P[ X t  j / X t 1  i]
ni j

Table-2:The maximum likelihood estimates for the first order model

Transition Probabilities P00 P01 P10 P11
Maximum Likelihood estimates 0.500014 0.499986 0.46689 0.53311

From the above table we see that being in increase index state given that the day was increase
at the previous time point is high (0.53311) and that of leaving decrease state is lowest
(0.46689).

In order to count the number of transitions for the second order chain, it is necessary to consider
the state of index at the three successive days. In other words we observe whether today is
decrease or increase given the state of index at the immediate past two days. Transition counts
for the second order chain are shown in the following table-3.

Table-3: Frequency for second order transition counts.

State of index in the Today’s of index Total
state
Immediate past two Decrease Increase (1)
daya (0)
0-0 288 172 460
0-1 115 190 305
1-0 173 135 308
1-1 191 278 469
Total 767 775 1542

In the above table, it is to be noted that among 1542 days 62.6% remain in the decrease index
for three consecutive days whereas 59.3% of the day remain increase index day state. The rest
of the states have changed the index status at least once in the three successive days. The highest
proportion (0.626) belongs to transition type decrease index all consecutive day and the lowest
proportion (0.374) belongs transition of day before yesterday decrease to yesterday decrease to
today increase index day state. The maximum likelihood estimates of transition probabilities
of a second order Markov chain obtained by the formula;

nijk
Pijk  , nij   nijk and Pijk  P[ X t  k / X t 1  j, X t 2  i]
nij k

The estimates are shown in Table-4.

Table-4:The maximum likelihood estimates for the second order model

Transition P000 P001 P010 P011 P100 P101 P110 P111
Probabilities
Maximum 0.602 0.398 0.431 0.6043 0.5979 0.402 0.453 0.64
Likelihood
estimates

From the above table we see that the probability (0.64) of being in increase index state given
that the index was in increase index state at the previous time point is high and that of being in
the increase state given the past two states were decrease is lowest (0.398)

Significance test of the order of Markov Chain:

Significance test of hypothesis for first order, i.e.,
H0 :Zero order Markov chain ( Pij  Pj )
H1 : First order Markov chain ( Pij  Pj )
The chi-squared test statistics is

1 1 nij nj
 2  2 nij [log e  log e ]
i 0 j 0 ni n..

Where S1-1(S-1)2=1 degrees of freedom. The calculated value of the above test is 52.844, which
is greater than  02.05,1 (3.84) . Hence we can conclude that alternative hypothesis is accepted that
means first order Markov chain is significant for that case.

Again significance test of hypothesis for second order, i.e.,

H0 :First order Markov chain ( Pijk  Pjk )
H1 :Second order Markov chain ( Pijk  Pjk )
The chi-squared test statistics is

1 1 1 nijk n jk
 2  2 nijk [log e  log e ]
i 0 j 0 k 0 nij n. j
Where S2-1(S-1)2=2 degrees of freedom. The calculated value of the above test is 8.156, which
is greater than  02.05, 2 (5.99) . Hence we can conclude that alternative hypothesis is accepted that
means second order Markov chain is significant for that case, I.J. Good (1976).

Significance Test for Logistic Regression Parameters and Identification of

Dependence of Index Change:
The logistic regression analysis of daily index changes is presented in table-5. The total number
of observation is 1510. Observation 0 means that the index decrease and the observation 1
means that the index increase. Out of total 1510 observation, 779 observations are the index
increase and 731 observations are index decrease. The proportion of index increase is 51.6%.
Conventionally loglikelihood is used as the measure of how well the model fits the data. Here
loglikelihood is 1021.1216, which shows that the model fits the data well.

Table-5: Logistic Regression output

Logistic regression Number of obs = 1510

LR chi2(2) = 49.54
Prob > chi2 = 0.0000
Log likelihood = -1021.1216 Pseudo R2 = 0.0237

TodayState Coef. Std. Err. z P>|z| [95% Conf. Interval]

YesterdayState .7293907 .1064736 6.85 0.000 .5207063 .938075

DaybeforeYesterday .0092372 .1064953 0.09 0.931 -.1994897 .2179641
_cons -.3155703 .0873828 -3.61 0.000 -.4868375 -.1443031

From the above table-5 we that the value of z-statistics is 0.0000 that indicates the model is
very much significant. Here one independent variable yesterday state is very significant with
today state rather the day before yesterday state. The probability that today is decrease index is
approximately 0.7293 times as likely if yesterday is decrease index compare to that day is
increase index. The probability that today is today decrease index is approximately 0.0092372
times as likely if day before yesterday decrease compare to that day is increase index.

Conclusion:
This research is mainly focused with two main statistical procedures; one is determination of
Markov Chain model both occurrence of daily index increase and decrease as well as the other
is logistic regression procedure, which describes the dependence of one binary dependent
variable on the other categorical independent variable. The Markov chain shown that decrease
of index of previous two time period compared to the increase of index of previous two time
period influence positively the decrease of index today. In logistic regression, we consider the
today index decrease is influenced by the decrease of index of previous two days.

References:
1. Ahmed, N., & Bassiouny, A. (2018). The Effects of Index Changes on Stock Trading:
Evidence from the EGX. Review of Economics & Finance, 11, 55-66.
2. Ajinkya, B. B., & Jain, P. C. (1989). The Behavior of Daily Stock Market Trading
Volume. Journal of Accounting and Economics, 11, 331-359.
3. Anam, S., Parvez, M., Khan, A. R., & Uddin, M. S. (2018). Developing Stochastic
Linear Programming Model to Optimize Agricultural Production under Uncertain
Flood Influence. Research Journal of Pharmaceutical Biological And Chemical
Sciences, 9(5), 1358-1369.
4. Anam, S., Khan, A. R., & Uddin, M. S. (2017). Developing Stochastic Linear
Programing Model for Production: Evidence from Bangladesh. International Journal
of Economic Perspectives, 11(4), 851-866.
5. Bildik, R., & Gülay, G.(2008). The effects of changes in index composition on stock
prices and volume: Evidence from the Istanbul stock exchange. International Review
of Financial Analysis, 17, 178–197.
6. Billingsley, P. (1961). Statistical methods in Markov chains. The Annals of
Mathematical Statistics, 12-40.
7. Basu, A. N. (1971). Fitting of a Markov chain model for daily rainfall data at
Calcutta. Indian Journal of Meteorology and Geophysics, 22, 67-74.
8. Gong, J., & Sun. S. (2009). A New Approach of Stock Price Trend Prediction Based
on Logistic Regression Model. International Conference on New Trends in
Information and Service.
9. Good, I.J. (1955) The likelihood ratio test for Markov chain. Biometrika, 42, 531-533.
10. Harris, L., & Gurel, E. (1986, September). Price and volume effects associated with
the changes in the S and P 500 list: New evidence for the existence of price pressure.
Journal of Finance, 41, 815−829.
11. Hossain, M. M., & Anam, S. (2012). Identifying the dependency pattern of daily rainfall
of Dhaka station in Bangladesh using Markov chain and logistic regression
model. Agricultural Sciences, 3(03), 385.
12. Huffman, G. W. (1987). A dynamic equilibrium model of asset prices and transaction
volume. Journal of Political Economy, 95, 138-159.
13. LeBaron, B. (1990). Forecasting improvements using a volatility index. Working
paper, Economics Department, University of Wisconsin.
14. Lynch, A. W., & Mendenhall, R. R. (1997). New evidence on stock price effects
associated with changes in the S&P 500 index. The Journal of Business, 70(3), 351-
383.
15. Mei, Z. & Jia, M. (2007). The application of Logit Model to predicting Shanghai
stock market. Statistics & Decision, 3, 102-104.
16. Moon, S., Ryoo, S., & Kwon, J. (1994). A Markov chain model for daily precipitation
occurrence in South Korea. International Journal of Climatology, 14, 1009-1016.
17. Pang, S. (2004). An Application of Logistic Model in Stock Forecasting. 8th
International Conference on Control, Automation, Robotics and Vision Kunming,
China, 1491-1496.
18. Rahman, M. T., &Moazzem, K. G. (2011). Capital market of Bangladesh: volatility in
the Dhaka stock exchange (DSE) and role of regulators. International Journal of
Business and Management, 6(7), 86-93.
19. https://www.dsebd.org/
20. Shleifer, A. (1986). Do demand curves for stocks slope down?. The Journal of
Finance, 41(3), 579-590.
21. Rashid, M. (2008, May). The Potential of the Bangladesh Capital Market. In Forum-A
monthly Publication of the Daily Star (Vol. 3, No. 5).

22. Krishnamurti, C., Sequeira, J. M., &Fangjian, F. (2003). Stock exchange governance
and market quality. Journal of banking & finance, 27(9), 1859-1878.