Predicting SET50 Stock Prices Using CARIMA

(Cross Correlation ARIMA)

Sornpon Wichaidit Surin Kittitornkun
Dept. of Computer Engineering Dept. of Computer Engineering
Faculty of Engineering Faculty of Engineering
King Mongkut’s Institute King Mongkut’s Institute
of Technology Ladkrabang, Thailand of Technology Ladkrabang, Thailand
Email: [email protected] Email: [email protected]

Abstract—Investing
in
stocks
is
one
of
the
most
popular
ap- This paper proposes a simple yet effective method for
proaches
for
money
investment.
This
paper
aims
to
predict
short- predicting short-term movements of a target stock based on
term
stock
prices
of
SET50
of
Stock
Exchange
of
Thailand
(SET).
ARIMA Model with another stock with highest lead/lag cor-
The
proposed
method
is
called
CARIMA
(Cross
Correlation

Autoregressive
Integrated
Moving
Average.
The
basic
idea
of
relation with the target one. The correlation coefficient can be
CARIMA
is
to
find
t he
m ost
h ighly
c orrelated
s tock
t o
predict
effectively incorporated with the original ARIMA thus called
the
target
one
in
addition
to
ARIMA
predicted
price.
The
results
Cross Correlation ARIMA (CARIMA). The performance of
of
CARIMA
model
yield
better
price
trends
(measured
by
10-day
CARIMA is evaluated based on SET (Stock Exchange of
correlation
coefficient)
w hile
% MAEs
( Mean
A bsolute
Errors)
Thailand) 50 dataset in terms of trend similarity and %Mean
are
quite
similar
with
those
of
ARIMA.
Absolute Error(MAE).
Keywords
-
Stock,
Prediction,
ARIMA,
Correlation,
Time

The rest of this paper is organized as follows. Section 2
Series
presents the existing related work. Section 3 explains the
theory of ARIMA and the proposed CARIMA method. Section
I.
I NTRODUCTION 4 describes dataset and results. Eventually, Section 5 provides
the conclusion including the future work.
Today,
Stock
Trading
is
a
very
popular
method
of
short-
term
investment.
It
is
known
that
the
time
series
of
stock
II. R ELATED W ORK
prices
fluctuate
g reatly
b ut
n ot
t otally
r andom.
T herefore,
the
There is much work based on time series technique for
investigation
of
factors
that
the
financial
data
has
long
attracted
analyzing and predicting trends of the stock market around the
researchers
from
various
different
areas,
such
as:
mathemati- world. All of these researches have been using various factors
cians,
economists
and
more
recently,
computer
scientists.
With
that have impacts on the stock market as input to predict the
the
advantages
of
information
technique,
a
huge
amount
of
trends.
stock
trading
data
can
be
collected
easily.
The
current
focus
is
C. Fonseka and L. Liyanage [5] developed an algorithm
to
design
a
method
of
extracting
useful
information
from
the
for predicting an individual stock from the Australian Stock
collected
data.
Therefore,
data
mining
has
drawn
considerable
Exchange(ASX) with correlation in 2008. Meanwhile, S. Chai-
attentions
from
the
young
generations
of
investor
in
order
to
gusin, et al [6], used a feedforward backpropagation neural
predict
the
changes
or
discover
the
patterns
of
the
stock
prices. network to predict the movement of SET index. The inputs of
There
are
two
analytical
methods
for
stock
trading:
1.Fun- the models consist of seven nodes including the Dow Jones
damental
[3];
2.Technical
Analysis
[4].
Fundamental
analy- index, Nikkei index, Hang Seng index, Gold prices, Minimum
sis
focuses
on
the
central
factors
of
a
company.
Such
as
Loan Rate (MLR), and the exchange rates of the Thai Baht and
financial
s tatements,
E arnings
p er
S hare
( EPS),
R eturn
on
the US dollar. In 2012, A. Srisawat [7] applied an association
Assets
(ROA),
Price-Earnings
Ratio
(P/E
Ratio),
Return
on
rule mining technique for discovering relationships between
Equity
(ROE),
price
and
book
value
(P/BV).
On
the
other
individual stocks from SET. Recently, W. Weiqing and Y. Lav
hand,
Technical
analyses
use
the
recent
historical
data
of
[8] used ARIMA Model to study volatile characteristics in
an
individual
stock,
including
opening,
highest,
lowest
and
the US dollar index itself in 2013. They used measurement
closing
prices
and
the
volume
of
stock,
to
predict
future
statistical model to fit the ARIMA Model to predict US Dollar
stock
price
movements.
There
are
many
indicators
that
can
be
index movement for one month.
created
from
the
data.
Such
as
Moving
Average
Convergence
As mentioned above, all of them used various data mining
Divergence
(MACD),
Average
Directional
Index
(ADX),
Ex- techniques to predict the stock or index. Very few of them
ponential
Moving
Average
(EMA),
Double
Exponential
Mov- have used relationships between stocks to help in the stock
ing
Average
(DEMA),
and
Relative
Strength
Index
(RSI). price prediction process. As a result, we are going to take
advantages of this gap to propose our method.
‹,(((
 Fig. 1. Unit Root Test of ADVANC

III. O UR P ROPOSED M ETHOD

The method used to develop ARIMA model for stock price

prediction with combining correlation of the other stock (the
stock that has lead/lag correlation to the focused stock) can
be explained in subsections below

A. Unit Root Test

We make use of the Augmented DickeyFuller test(ADF) as
the unit root test for the dataset whether it is stationary or not.
So we test it on the closing price of an individual stock. The
result of this test shown in Fig. 1 that the data is stationary.

B. ARIMA model
Autoregressive integrated moving average (ARIMA) model
is generalisation of an autoregressive moving average
(ARMA) model. ARIMA include differencing I to ARMA
Fig. 2. ADVANC: ARIMA Predicted Price, Xt (Top) vs Actual Price, Xt
AR(autoregressive)+I(integrated)+MA(moving average). The (Bottom)
equation of ARIMA Model is.


p 
q TABLE I

Xt = m + Θi−p Δd Xt−i − i−q t−i (1) ARIMA M ODEL S ELECTION OF ADVANC BASED ON A KAIKE
I NFORMATION C RITERION (AIC) AND BAYESIAN I NFORMATION
t=1 t=1
C RITERION (BIC)
Where m is a constant, Θ and  are parameter of autore- ARIMA AIC SC
gressive and moving average parts,  is error(white noise), i p=1,d=0,q=0 4594.27 4608.31
is lead/lag time variable (days), Δ is the difference, X  is a p=0,d=1,q=0 4582.58 4587.26
closing price of an individual stock, t is a day variable. p , p=0,d=0,q=1 6846.23 6860.27
d and q are orders of autoregressive, difference and moving p=1,d=1,q=0 4584.58 4593.94
average parts, respectively. p=1,d=0,q=1 4596.27 4614.99
p=0,d=1,q=1 4584.58 4593.94
The ARIMA predicted stock price Xt is parameterized as p=1,d=1,q=1 4586.58 4600.62

Xt = ARIM A(p, d, q) (2)

where AR: p is order of the autoregressive part, I: d is the C. Cross Correlation with Lead/Lag i days
degree of first differencing involved, and M A: q is order of
the moving average part. We assume the closing price is the one that shows the
To select the best model of ARIMA(p,q,d), Akaike Informa- realest price movement behaviour rather than High Low Open
tion Criterion (AIC) and Bayesian Information Criterion (BIC) price. Cross Correlation analysis of closing prices is a measure
can be evaluated to estimate the quality as listed in Table of the interrelationship between two stocks prices with a
1. In this case, ARIMA(0,1,0) model is selected to predict function of lead/lag time of one company relative to another.
ADVANC because of the smallest values of AIC and BIC. The The correlation coefficient can be calculated with respect to
model returns the smallest AIC of 4582.58 and relatively small lead/lag time i days as defined in Eq.(3).
BIC of 4587.26. in Fig. 2 show that there is no correlation 796
between predicted ADVANC and the actual price movement 
ρxt yt (i) = [(Xt − uX )(Yt−i − uY )]/δX δY (3)
of ADVANC. t=1
Fig. 3. Cross Correlation Coefficient ρXt Yt (i) with Lead/Lag i between
BAY and BH Fig. 4. ADVANC: CARIMA Predicted Price X̂t with Lead, i = 10 from
INTUCH (Top) vs Actual Price, Xt (Bottom)

where t is the day variable that starts from January 1st, 2012
correlated stock, t is day variable and i is the lead time variable
to March 31st, 2015, i is the lead/lag time variable in days, δX
in days.
is a standard deviation of stock X, δY is a standard deviation
of stock Y , uX is a mean of stock X and uY is a mean of IV. E XPERIMENT AND R ESULTS
stock Y . In this section, there are 2 main activities. Firstly, we explain
We choose the highly correlated coefficient whose the details of SET50 dataset. Secondly, we present and discuss
|ρXt Yt (i)| ≥ 0.8 at time lag t − i. It means that stock Yt is the results.
leading stock Xt by i days. As a result, stock Yt−i can improve
ARIMA Model to predict stock Xt price movement. The A. SET50 Dataset
example of cross correlation between ADVANC and INTUCH The SET50 dataset is collected from Stock Exchange Thai-
shown in Fig. 3. land (SET). It consists of the 50 most valuable companies
(stocks) in Thailand. We use their closing prices to find the
D. CARIMA: Cross Correlation ARIMA correlation coefficient of each individual stock to the others
After we get the results of ARIMA price Xt predicted up to with the lead and lag time. The dataset starts from January
10 days from April 1-18, 2015. The highly correlated stock Y 1st, 2012 (t = 1) to March 31, 2015 (t = 796).
whose |ρXt Yt (i)| ≥ 0.8 that leads stock X for i days is picked. B. Results and Discussion
By shifting back stock Y to day t − i, the Rate of Change
The stock names, 10-day correlation coefficients, and
(ROC) from Yt−i is applied with respect to Xt . Finally, the
%MAEs between CARIMA X̂t and actual prices Xt vs.
CARIMA predicted X̂t is a product of ρXt Yt (i), Xt , and ROC
ARIMA Xt and actual prices Xt are listed in Table 2. NA in
of Yt−i as shown in Eq.(4).
ρXt Xt column means Not Available because the ARIMA result
Yt−i+1 − Yt−i is a straight line resulting divided by zero. The price trends of
X̂t = ρXt Yt (i) × Xt × (1 + ) (4) CARIMA are closer to the actual prices than those of ARIMA
Yt−i
alone (measured by 10-day correlation coefficients), although
where ρXt Yt (i) is the correlation coefficient between stock the %MAEs (Mean Absolute Errors) are quite similar. %MAEs
x and y, X is the ARIMA predicted price, Y is the most of CARIMA is not better than %MAEs of ARIMA.
 TABLE II %MAEs are similar to those of ARIMA. For future work, we
10-DAY C ORRELATION C OEFFICIENTS AND %MAE OF CARIMA, X̂t intend to apply two or more stocks as predictors to CARIMA
AND ACTUAL S TOCK P RICES ,Xt VS . ARIMA,Xt AND ACTUAL S TOCK
P RICES ,Xt , (NA: N OT AVAILABLE ) for each individual SET50 stock. It is expected that they can
improve the prediction performance even more.
Stock ρX̂t X  ρXt Xt %MAEX̂t X  %MAEXt Xt
t t R EFERENCES
JAS 0.58 NA 2.29 2.13
[1] P. S.P. Cowpertwait and A. W. Metcalfe, Introductory Time Series with
ADVANC 0.40 NA 2.85 2.51 R, Springer, 2009, pp.137-155
RATCH 0.35 NA 0.84 0.51 [2] C. M.Conover and D. R. Peterson, The Lead-Lag Relationship between
GLOBAL 0.33 -0.39 1.53 1.32 the Option and Stock Markets prior to Substantial Earnings Surprises
IVL 0.31 NA 2.07 2.01 and the Effect of Securities Regulation. Journal of Financial and Strategic
THCOM 0.30 NA 3.96 2.23 Decisions, Spring 1999, Vol. 12 No. 1
BANPU 0.27 NA 3.65 3.20 [3] B. Graham , J. Zweig and W. Premchaiswadi, The Intelligent Investor:
The Definitive Book on Value Investing. A Book of Practical Counsel,
TCAP 0.15 NA 1.86 1.46
Collins Business Essentials, 2006
BCP 0.14 NA 4.61 4.78 [4] S. Nison, Beyond Candlesticks: New Japanese Charting Techniques
EGCO 0.10 -0.87 1.53 1.51 Revealed, Wiley Finance, 2009
KKP 0.08 -0.15 0.89 0.74 [5] C. Fonseka, and L. Liyanage, A Data mining algorithm to analyse stock
PTTEP 0.07 -0.56 6.82 6.36 market data using lagged correlation, 4th International Conference on
CENTEL 0.04 NA 3.94 4.10 Information and Automation for Sustainability, 2008, pp.163 - 166
TTW 0.03 -0.88 1.38 1.41 [6] S. Chaigusin, C. Chirathamjaree, and J. Clayden, A Data mining
algorithm to analyse stock market data using lagged correlation, Pro-
BAY -0.09 0.32 4.62 5.17 ceedings of the International Conference on Computational Intelligence
BH -0.11 -0.75 3.67 1.74 for Modelling Control and Automation, 2008, pp. 670 - 673
GLOW -0.17 -0.74 2.64 2.46 [7] A. Srisawat, Discovery Stock Trading Patterns: A Case Study of Thai
TRUE -0.20 0.80 3.84 4.07 Stock Market, International Journal of Intelligent Information Processing,
BDMS -0.22 -0.39 2.62 2.42 2012, Vol. 3 Issue 1, pp.1-9
CPN -0.33 NA 3.20 3.61 [8] W. Weiqing and Y. Lav, A Study of the USDX Based on ARIMA Model
A Correlation analysis between the USDX and the Shanghai index.
SCB -0.38 -0.51 1.77 1.33
Consumer Electronics, Communications and Networks (CECNet), 3rd,
PS -0.48 -0.71 2.62 2.31 2013
[9] J. A. Ryan, J. M. Ulrich, and W. Thielen, quantmod: Quanti-
tative Financial Modelling Framework, Version: 0.4-5, https://cran.r-
project.org/web/packages/quantmod/index.html, 2015
Because the main purpose of ARIMA is to predict price [10] R.J. Hyndman, forecast: Forecasting Functions for Time
Series and Linear Models, Version: 6.1, https://cran.r-
movement as close to the actual price as possible. But project.org/web/packages/forecast/index.html, 2015
CARIMA incorporates the cross correlation coefficient with
ARIMA in order to improve the correlation of the price move-
ments. For example, CARIMA predicted prices of ADVANC
predicted by INTUCH can be plotted in Fig. 4 and compared
with those of ARIMA in Fig. 2. We can see that those of
CARIMA are more correlated while the predicted price from
ARIMA is a straight line thus not correlated to the actual price.
Although %MAE are slightly higher than ARIMA. CARIMA
can yield better price trends as ρX̂t X  are better than ρXt Xt .
t
Since the dataset is not big enough, only 50 individual
stocks, it is hard to find the leading/lagging stocks with
high correlation coefficients to make the better prediction.
Moreover, most of the stocks have the highest/ least correlation
coefficients in lag i = 0 that can be predicable that stock always
move up and down together in the same day. Its not reasonable
to use that to help ARIMA to predict individual stock price
movement.

V. C ONCLUSION
This paper aims to incorporate cross correlation coefficient
with ARIMA to predict short-term (daily) SET50 stock price
movement. The proposed method is called CARIMA to predict
stock A using the most highly correlated stock B within 10-day
lead. The empirical results obtained and compared CARIMA
predicted prices X̂t and ARIMA predicted Xt against the ac-
tual prices Xt . In terms of performance evaluation, CARIMA
prices are more correlated to the actual prices and their