International Transaaction Journal of Engineerring, Management, … · 2020. 4. 13. · time series...
Transcript of International Transaaction Journal of Engineerring, Management, … · 2020. 4. 13. · time series...
*CorrespondinTransaction Jo1906-9642 CO
a Departme A R T I C LArticle historyReceived 11 JuReceived in reNovember 201Accepted 29 NAvailable onli2019 Keywords: Moving aveAutoregressArtificial neSupport vecFinancial mExtreme leamachine.
1. INTRFinanc
investmentGiven theiprediction.
©
ng author (Fahournal of EngineODEN: ITJEA8
InteMan
COMPCOMP
Beenish
nt of Manag
L E I N F O y: uly 2019
evised form 20 19 November 2019ine 09 Decembe
erages; sive models; eural networkctor machine
marketing; arning
RODUCTIcial marketst, institutionir importanc. These pr
©2020 Internatio
eem Aslam). Teering, ManagemPaper ID:11A04
ernationanagemen
PARATIVPUTING M
h Bashir a, F
gement Scienc A B S T
9 er
k; e;
Findue to prices rsoft cocomplenonlinea compcomputexchanunique dynamitraditioautoregartificialearningshows modelsfinanciamodelsOur reregardiinvestopredictiDiscipl©2020 IN
ON s are the banal investmce, much atredictions a
onal Transaction
Tel: +92-321-50ment, & Applied4J http://TUEN
al Transat, & App
http
VE ANALMODELS
Faheem As
ces, COMSA
R A C T nancial marthe nature reflect. Ecoomputing mex and dynaear models.parative anting modelsge coveringregarding t
ics of the onal modelgressive moal neural g machinesthat soft in tradingal time se, the artifici
esults are mng the car
ors to betteive ability olinary: ManNT TRANS J EN
arometer ofment acting ttention hasare then in
n Journal of Eng
063253. Email:d Sciences & TeNGR.COM/V11/
action Joplied Scie
p://TuEngr
LYSIS OFS FOR TR
lam a*
ATS Universi
rket forecasof the timenomist take
models assuamic in naThis resear
nalysis of s for the trag the dailythe target va
market nols, we inodels whernetworks,
s to have a computing
g signals preries data. ial neural nmore convireful selecter understaof the modenagement SNG MANAG SCI
f any econoas a part o
s been givenncorporated
gineering, Mana
: fahimparachaechnologies. Vol11A04J.pdf DO
urnal of ences & T
r.com
F TRADITRADING
ty Islamabad
sting alwaye series dataes this data ume that t
ature and carch paper setraditional
ading signay data fromariable it preot just the nclude moreas in sof
support vcomprehenmodels pe
rediction shHowever,
etwork becoincing as cion of marand marke
el. ciences (FinI TECH.
omy. It facilof mutual fn by researinto financ
agement, & App
[email protected] lume 11 No.4 ISOI: 10.14456/IT
EngineerTechnolo
PA
TIONAL G SIGNAL
d, PAKISTAN
ys remains a as well thas a linear time-series an be betteettles the deforecasting
als predictiom 1997 to edicts which
next periooving averft computinector mach
nsive analyserform betthowing non
amongst omes the becompare torket feature
et behavior
nancial Scie
litates internfunds and prchers and acial risk m
plied Sciences &
©2019 InternaSSN 2228-9860 TJEMAST.2020.7
ring, ogies
APER ID:
L AND SOLS PRED
N.
a challengihe informatiprocess whdata is n
er analyzedebate by cog models on of Pakist2018. Our h reflects thod future rages as ng models hines and sis. This comtter than trn-linear behall soft co
est predictivo existing es, which r and imp
ence).
rnational traportfolio macademia re
models and
& Technologies
ational eISSN
70 1
: 11A04J
OFT ICTION
ing issue ion stock
hereas the nonlinear, d through onducting and soft tan stock study is
he overall price. In well as we take extreme
mparison raditional havior of omputing ve model. literature can help rove the
ade, foreignanagement.egarding itsdeveloping
J
n . s g
2 Beenish Bashir, Faheem Aslam
trading strategies. (Moradi & Rafiei, 2019; Zhong & Enke, 2019). As the importance and predictability of financial markets are supported by previous studies
(Fama, 1965; Malkiel & Fama, 1970) but when it comes to the modeling techniques for forecasting there is no consensus. The dispute is much related to the nature of the data financial markets generate. The long-term investors are more concerned about the fundamentals of companies like price-earnings ratio, revenue, expenses, assets, liabilities, management policy and financial ratio (Lam, 2004). Whereas, short-term investors rely on price movements of stock, understanding market behavior through different market features (Murphy, 1986). Technicians avoid the analysis of all economic factors by focusing on pattern recognition of price considering this information enough for future price determination (Hu et al., 2015; Patel et al., 2015; Teixeira & De Oliveira, 2010; Żbikowski, 2015). Tsinaslanidis and Kugiumtzis (2014) use numerous technical indicators for market analysis which is time-series data in nature. Huang et al. (2019) also use different momentum and volatility indicators for bitcoin predictive analysis.
For short-term prediction, traditional forecasting models consider time series data as a linear process and apply the smoothening and autoregressive process to predict future price movement (Kumar & Murugan, 2013; Lin et al., 2012; Wang et al., 2012). Financial time series are complex, nonlinear, dynamic and chaotic. Soft computing models can capture this nonlinear behavior (Cheng & Wei, 2014; Huang & Tsai, 2009; Lee, 2009). Tay and Cao (2001) compare artificial neural networks (ANN) and support vector machines (SVM) and confirm their suitability for prediction of the stock market. Huang, et al. (2005) use SVM for predicting the directional movement of the NIKKIE 225 index. Kara et al. (2011) also employ SVM and ANN for predicting the Istanbul Stock Exchange.
Our work is an addition to existing literature to settle the debate amongst traditional forecasting models and soft computing models for trading signals prediction. All the existing literature is related to price prediction, completely ignoring the trading signals forecasting that can enlighten investors about entry and exist decisions. Secondly, we also focused on the relevant feature selection to improve the predictive ability of models and to better understand market behavior. Lastly, our analysis confirms the nonlinear and dynamic nature of financial time series data negating the assumption of traditional forecasting models.
2. LITERATURE REVIEW In the past ten years, various time series methodologies have been developed for financial market
forecasting that helps improve investment decisions (Teixeira & De Oliveira, 2010). Traditional forecast models and soft computing models are two major approaches (Wang et al., 2011). In both modeling approaches, financial data is considered as a time series data which is actually numerical observations accumulated in sequential order over a period of time. (Brockwell & Davis, 2009; B. Wang et al., 2012). This organization of financial data enables model developers to utilize time series tools to understand market behavior (Oliveira & Meira, 2006). Further, these tools pave a way to do data mining tasks, such as classification, trend analysis, seasonal effect, cycles and extreme event detection (Cryer & Chan, 2008).
The traditional models involve averages, regression and autoregressive models based on the linearity of normally distributed variables (Lin et al., 2012; Wang et al., 2012). In reality, financial time series data is dynamic, chaotic, nonlinear and highly volatile which makes its prediction
*Corresponding author (Faheem Aslam). Tel: +92-321-5063253. Email: [email protected] ©2019 International Transaction Journal of Engineering, Management, & Applied Sciences & Technologies. Volume 11 No.4 ISSN 2228-9860 eISSN 1906-9642 CODEN: ITJEA8 Paper ID:11A04J http://TUENGR.COM/V11/11A04J.pdf DOI: 10.14456/ITJEMAST.2020.70
3
complex as compare to other non-financial time-series data (Vanstone & Finnie, 2009). Financial time series inherit these characteristics from economic factors, investor’s sentiments, political events and movement of other stock markets (Kara et al., 2011).
The nature of financial time series data calls for flexible and adaptive models for future price prediction (Huang & Tsai, 2009) and soft computing models serve the purpose (Lin et al., 2012). Soft computing models, such as artificial neural networks, support vector machine and extreme learning machine give better predictions with high accuracy (Lee, 2009). Most of the soft computing models can handle nonlinear relations through relevant market features with less statistical assumptions (Atsalakis & Valavanis, 2009). Weng et al. (2017) find that soft computing models are appropriate for developing rules when using with a rich knowledge database. Zhong and Enke (2017) conclude a trading strategy based on classification models yield high returns than the benchmark T-bill strategy. Liang et al. (2009) find that the non-parametric model outperforms parametric models in financial market forecasting. As soft computing models are self-adaptive and are more tolerant of imprecision (Cheng & Wei, 2014). Hence, we investigate whether these findings hold for trading signals prediction of Pakistan stock exchange or not.
3. METHOD 3.1 DATA
We use the Pakistan Stock Exchange (KSE-100 index) daily quotes from 7/02/1997 to 18/07/2018 with the total observations 5168. The data set is bifurcated into a training dataset and testing dataset. The training dataset has 3764 observations cover 70% of total observations and the testing set has 1404 observations include the remaining 30% of total observation.
3.2 TARGET VARIABLE “T” The target variable T (Equation 1) gives a holistic picture of stock prices by incorporating overall
dynamics in the following days. This cannot be merely done through price movements, therefore, the target variable is defined as the sum of all variation above an absolute value of target margin 𝑥% (Torgo, 2011) 𝑇 = ( 𝑣 ∈ 𝑉 : 𝑣 > 𝑥% ∨ 𝑣 < −𝑥%) (1)
where 𝑉 (Equation 2) is k percentage variation of current close price and the following k days price
average.
𝑉 = 𝑃 − 𝐶𝐶 (2)
and the daily average price is computed as Equation 3.
𝑃 = 𝐶 + 𝐻 + 𝐿3 (3)
𝐶 , 𝐻 𝑎𝑛𝑑 𝐿 are close, high and low prices for the day i respectively.
The target variable is T value and develops a model that predicts this value by using the feature’s
4 Been
informationusing the v
The this the averaThe only cvalues let a
3.3 INPUWe us
selected frothrough a r
3.4 METThis se
nish Bashir, Fa
n that givesvalues in Eq
hreshold 0.1age of last fcare should an investor
UT FEATUse fifteen feom the list orandom fore
Fig
Sr. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
THODOLOection expla
aheem Aslam
s three tradquation 4.
𝑠𝑖𝑔𝑛𝑎 and -0.1 ar
four days avbe taken is
trade on mi
URES eatures (Tabof features est techniqu
gure 1: Var
TableSelected TAverage TAverage TAverage TATR tr. ADX (WeADX DipADX DinADX DXBollinger Bollinger Bollinger Bollinger SAR (ParRun MeanRun Stand
OGY ains both the
ding actions
𝑎𝑙 = 𝑠ℎ𝑜𝑙𝑑𝑏re heuristic verage prices that very inor variatio
ble 1) as theavailable in
ue (Figure 1
iable impor
e 1: List of STechnical IndTrue Range True Range HTrue Range Lo
elles Wilder’sp n
Bands Down Bands MovinBand Band Up
rabolic Stop ann dard Deviation
e traditional
s (sell, hold
𝑠𝑒𝑙𝑙 𝑖𝑓 𝑇 <𝑖𝑓 − 0.1𝑏𝑢𝑦 𝑖𝑓 𝑇 >and any oth
e that is 2.5%high values
on, incurring
e input of thn the Techni).
rtance accor
Selected Tedicators
High ow
s Directional M
n ng Average
nd Reverse)
n
l and soft co
d and buy),
−0.1𝑇 0.1> 0.1
her value ca% higher thas will generg higher ris
he soft comical trading
rding to the
echnical Ind
Movement Ind
omputing m
this transfo
n be used. Han the currenrate fewer sk (Torgo, 2
mputing modrule packag
random for
icators
dex)
models used f
ormation is
However, thent close (4xsignals and
2014).
del. These fge (Ulrich e
rest.
for predictio
carried out
(4)
he value 0.1x0.025=0.1)
very small
features areet al., 2018)
on analysis.
t
) l
e )
.
*Corresponding author (Faheem Aslam). Tel: +92-321-5063253. Email: [email protected] ©2019 International Transaction Journal of Engineering, Management, & Applied Sciences & Technologies. Volume 11 No.4 ISSN 2228-9860 eISSN 1906-9642 CODEN: ITJEA8 Paper ID:11A04J http://TUENGR.COM/V11/11A04J.pdf DOI: 10.14456/ITJEMAST.2020.70
5
3.4.1 TRADITIONAL MODELS
We first explain the simple moving average, centered moving average, right-aligned moving average, exponential weighted moving average then move to autoregressive moving average model.
Simple Moving Average (SMA) is the average of n prices (Equation 5) where each observation is given equal weight. If we take difference of the two of SMA values at times t and t-1, it gives Equation 6 and recursive form become as Equation 7 which improves the computational speed of SMA in real-life scenarios (Zakamulin, 2017) by reducing total n operations involve in Equation 5 to just three mathematical operations (Equation 8) irrespective of the length of averaging window length.
𝑆𝑀𝐴 (𝑛) = 1𝑛 𝑃 (5)
𝑆𝑀𝐴 (𝑛) − 𝑆𝑀𝐴 (𝑛) = 𝑃 − 𝑃𝑛 (6)
𝑆𝑀𝐴 (𝑛) = 𝑆𝑀𝐴 (𝑛) + 𝑃 − 𝑃𝑛 (7)
The Herfindahl index of SMA (Rhoades, 1993) equals ; hence defining the smoothness of
SMA(n) as ( ) = 𝑛. The increase in the average window length not only increases its smoothness
but also increases the average lag time of SMA which is a linear function of smoothness (Equation 9). 𝐿𝑎𝑔 𝑡𝑖𝑚𝑒(𝑆𝑀𝐴 ) = ∑ 𝑖∑ 𝑖 = 𝑛 − 12 (8)
𝐿𝑎𝑔 𝑡𝑖𝑚𝑒(𝑆𝑀𝐴 ) = 12 × 𝑆𝑚𝑜𝑜𝑡ℎ𝑛𝑒𝑠𝑠(𝑆𝑀𝐴 ) − 12 (9)
It is quite easy to find a trend and breakthroughs in a trend by looking on historical data
considering time series data of 𝑃 as a combination of trend 𝑇 and an irregular component called as noise 𝐼 . Then, an additive model can be written as Equation 10. Noise is a short-lived variation around the trend eliminated through centered moving average or other smoothening tools. Any average of time series data is computed using a fixed window length that rolls through time. This window length is called an averaging period. In case of centered moving average if n is the window length then it consists of a center and two halves of length k such that n = 2k + 1. Since centered moving average at time t (equation11) removes noise so the value of CMA is the value of trend in the given time series data. The window length n is selected keeping in mind the elimination of noise (Equation 12) 𝑃 = 𝑇 + 𝐼 (10)
𝐶𝑀𝐴 = 1𝑛 𝑃 (11)
6 Beenish Bashir, Faheem Aslam
Tt = 𝐶𝑀𝐴 (n) (12)
In real life, an analyst is interested in the forecasting of time series t + 1 given the historical data series t (Equation 13).
𝑅𝑀𝐴 = 1𝑛 𝑃 (13)
Mathematical comparison of centered moving average and right-aligned moving average (RMA)
at a given time t leads to the conclusion that 𝑅𝑀𝐴 is equal to 𝐶𝑀𝐴 (Equation 11). In fact, RMA is a lagged version of CMA given the same window length and shares the same properties of CMA. Particularly, RMA with longer window length is better at removing noise from the data series but this comes with longer lag time. The lag time is given as Equation 15 𝑅𝑀𝐴 (𝑛) = 𝐶𝑀𝐴 (𝑛) (14)
𝑙𝑎𝑔𝑡𝑖𝑚𝑒 = 𝑘 − 𝑛 − 12 (15)
However, the linearly weighted moving average has the drawback that it assigns the same weight to all observations ignoring the importance of recent observation in future prediction. This problem is addressed by exponential moving average (EMA), using the concept of exponential factor λ (Equation 16). The value of λ can be greater than zero and less than or equal to 1. 𝐸𝑀𝐴 (𝜆, 𝑛) = ∑ 𝜆 𝑃∑ 𝜆
(16)
Autoregressive Integrated Moving Average (ARIMA) is the most commonly used forecasting
model amongst traditional forecasting models. It is a generalized version of ARMA (autoregressive moving average) when used for differenced data rather than original data series. Orders of AR part (p), the difference (d) and MA (q) part specify the ARMA model and the model is said to be of order (p,d,q). However, AR and MA are different models for stationary time series and ARMA (and ARIMA) is a hybrid form of these two models for a better fit. The steps of building ARIMA models are explained as follows.
Auto Regression (AR) is a class of linear models where the dependent variable is regressed against its own lagged values. If 𝑦 is modeled via the AR process, it is written as Equation 17 similar to simple linear regression. It has an intercept like term (δ), regressors 𝑦 , and parameters ∅ an error term 𝜀 . The only special thing is that regressors are the dependent variable’s own lagged terms. If lag up to p is included in the model, the AR process is said to be of order p. 𝑦 = 𝛿 + ∅𝑦 + ∅ 𝑦 + ⋯ + ∅ 𝑦 + 𝜀 (17)
Moving Average (MA) is another class of linear models. In MA, the output or the variable of interest is modeled via its own imperfectly predicted values of current and previous times. It can be written in terms of error terms as Equation 18.
𝑦 = 𝜇 + 𝜃 𝜀 + 𝜃 𝜀 + ⋯ + 𝜃 𝜀 + 𝜀 (18)
*Corresponding author (Faheem Aslam). Tel: +92-321-5063253. Email: [email protected] ©2019 International Transaction Journal of Engineering, Management, & Applied Sciences & Technologies. Volume 11 No.4 ISSN 2228-9860 eISSN 1906-9642 CODEN: ITJEA8 Paper ID:11A04J http://TUENGR.COM/V11/11A04J.pdf DOI: 10.14456/ITJEMAST.2020.70
7
Again, it has a form similar to classic linear regression. The regressors are the imperfections (errors) in predicting previous terms. Here the model is specified with a positive sign for the parameters. It is not uncommon where we have a negative sign for the parameters. The model above include errors for q lags and said to have an order of q. In the case of differenced data ARIMA (p,d,q) become ARMA(p,q) having the mathematical form as Equation 19.
𝑦 = 𝛿 + ∅ 𝑦 + ∅ 𝜀 + 𝜀 (19)
3.4.2 SOFT COMPUTING MODELS
Artificial Neural Network (ANN) is a nonlinear regression technique and popularly used for stock market prediction (Zhiqiang et al., 2013). This field is tracked back to (McCulloch & Pitts, 1943) mathematical function perceived as a model of biological neural network. This model of the neural network comprises three components: weighted inputs that are like synapses in the human brain. Adder- The summation of all input signals corresponds to the neuron membrane which assembles all electrical charges. Activation function- determines if a neuron has an action potential for a specific set of inputs (Algorithm 1).
ALGORITHM 1 Artificial Neural Network
1. Start with small random weights 2. Input the data set 3. For forwarding phase: hidden layer compute activation function of each neuron then
calculate the activation function of output 4. For the backward phase: calculate error at the output and at the hidden layers then update
the weights of both layers 5. For recall apply the forward phase described in step 3
When it comes to the interpretation of ANN it’s more like a black box. Therefore, Support vector machines (SVMs) apply the concept of margin to solve the problem of ANN. SVM easily separates the data by mapping it into high dimensions (Boser et al., 1992). By aiming to maximize the size of margin that classifies the objects without any point lying inside.
ALGORITHM 2 Support Vector Machine
1. Begin with an input data set 2. Classify the data set 3. Apply SVM with a different kernel function 4. Specify hyperplane 5. Repeat step 3 if obtained accuracy is not obtained
The learning speed of ANN and SVM is slower than what is required because of slow
gradient-based learning algorithms. Whereas the Extreme Learning Machine (ELM) randomly selects hidden nodes and analytically finds the output weights (Algorithm 3).
8 Been
ALGO1. 2. 3. 4.
4. RESUSMA i
R package returns thepurple colrepresentedline. This i95% predic
Center(Svetunkov
nish Bashir, Fa
ORITHM 3With the trAssign ranCompute hCompute th
ULT AND is computedforecast (H
e model witlored line wd after a reds because thction interv
red movingv, 2017) w
aheem Aslam
3 Extreme L
raining set andom input whidden layerhe output w
DISCUSSd using sma
Hyndman et th the loweswith AIC v
d cutting linehe actual va
val.
Fg average iith the argu
Learning Ma
and activatioweights andr output mat
weights
SION a function () al., 2019). Bst value. Hevalue of -3e (Figure 2)alues are use
Figure 2:
igure 3: Ces computedument orde
achine
on functiond bias trix
in R packaBy default,ere the orde3151.602. T. The foreca
ed in the con
: Simple Mo
entered Movd in R deper equals to
n and hidden
ge smooth (SMA orderer of SMA The forecasast trajectorynstruction o
oving Avera
ving Averagloying func “null” to f
n neuron num
(Tukey) andr selection (his 12 givin
st is done y of SMA (1f point forec
age.
ge. ction cma (find order b
umber
d forecast fu(h) is based ong a good ffor the nex12) is not jucasts up to h
() in packabased on A
unction () inon AIC and
fit shown inxt 10 days
ust a straighth=12 with a
age SmoothAIC. In this
n d n s t a
h s
*CorrespondinTransaction Jo1906-9642 CO
analysis Cproperty ofreal data pa
The rialign “righblack line prediction a magnifieand the lighof the pred
We esmethod of is minimizhigher valu1990). Figuis EMA wh
ng author (Fahournal of EngineODEN: ITJEA8
CMA at ordf a centeredattern (Figu
ight-alignedht” in R pacand the righinterval (Fi
ed version oht shaded ar
diction interv
stimate expoleast square
zed (Čisar &ues of lambure 6 illustrahereas the g
eem Aslam). Teering, ManagemPaper ID:11A04
der 13 with d moving avure 3).
d moving avckage zoo (ht aligned mgure 4). Ov
of the forecarea is 95% pval.
Fi
onential moe is used to f& Čisar, 20da. Whereaates EMA a
gray line is t
Figure 5
Tel: +92-321-50ment, & Applied4J http://TUEN
AIC valueverage that
verage is c(Shah, Zeilemoving avererall, the RMasted regionprediction in
igure 4: RM
oving averafind the opti11). Table
as most of thand actual dathe predicte
5: Magnified
063253. Email:d Sciences & TeNGR.COM/V11/
e, -3896.784is irrespect
computed byeis, & Grothrage is showMA fits weln where thenterval and
MA and Tra
age by own imal value oe 2, there ishe packagesataset, the b
ed interval a
d Graph of R
: fahimparachaechnologies. Vol11A04J.pdf DO
4 is computive of wind
y using funhendieck, 2wn by the oll with the p
e dark shadethe predicte
ading Signal
code with of λ for whics an increass use lambdblack line is and EMA is
RMA Forec
[email protected] lume 11 No.4 ISOI: 10.14456/IT
uted. Our redow length,
nction rollm2005). The arange line a
pattern of reaed area is 8ed blue line
ls Data.
the lambda ch the sum osing trend ina value as 0the actual dfollowing t
casted Regio
©2019 InternaSSN 2228-9860 TJEMAST.2020.7
esult reveal, CMA clos
mean () withactual data and the grayal data. Figu0% predicti
e lies within
a optimized of squared en SSE valu0.2 (Lucas &data and the the actual d
on.
ational eISSN
70 9
ls the basicsely follows
h argumentseries is in
y line is theure 5 showsion interval the bounds
at 0.1. Theerrors (SSE)ues with the& Saccucci,orange line
data set.
c s
t n e s l s
e ) e , e
10 Been
Three Parameter the data bynumber is and autocothe AR or(autocorrelfirst differePACF plotoff after la
Fi
Since stationary (Pfaff et alOnce we d
nish Bashir, Fa
major stepsestimation
y taking diffd. There ar
orrelation. Arder p. The lation) cut oence stationt of a simulag 1. So thre
igure 7: Gr
the ACF pat first diff., 2016). Thetermine th
aheem Aslam
Tab
Fs are follow(3) Diagno
ferences andre different After determ
lag at whioff is the ordnary as showated time seee potential
raphical Rep
plot of firstference. Wehe value of the orders, th
ble 2: The o
00000
Figure 6: EMwed to makeostics and pod keep takintools to det
mining d, wech PACF cder q of MAwn in Figur
eries. For the(p, q) speci
presentation
t differencee further invthe test statie next step i
optimized vaλ SS
0.1 0.040.2 0.050.5 0.050.7 0.050.9 0.05
MA and Acte an ARIMAotential impng the diffetect stationae can utilizecuts off is tA. The tradiures 7 and 8e series, theification wo
n of First Di
e is better vestigate it istic is 0.003is to find th
alue of LamSE 4432 035 146 157 157
tual Data SeA forecast mprovement. rence until
arity. Like vsample PA
the order oing signal da8. Next Figue ACF cuts oould be (1, 0
ifferenced s
than level with unit r39 so the see parameter
mbda
et. model: (1) MWe first chthe series b
visual obserACF (partial
f AR and tata was not ures 9 and off after lag
0), (0, 1) and
series of Tra
stationary goot test KP
eries is first rs. Common
Model speciheck the stabecomes starvation of thautocorrela
the lag at wlevel statio10 show th
g 1 and PACd (1,1).
ading Signa
graph so thPSS in R padifferenced
nly used tec
ification (2)ationarity oftionary thathe data plotation) to getwhich ACFonary but itshe ACF andCF also cuts
als.
he series isackage urcad stationary.hniques are
) f t t t
F s d s
s a . e
*CorrespondinTransaction Jo1906-9642 CO
least-squar2007) use adequacy t
The or
ng author (Fahournal of EngineODEN: ITJEA8
Figu
Figure 9
Figure 10:
re, maximumthe maxim
through diagriginal plot
eem Aslam). Teering, ManagemPaper ID:11A04
ure 8: Autoc
: Autocorre
Partial Aut
m likelihoodmum likelihgnostics. shows clear
Tel: +92-321-50ment, & Applied4J http://TUEN
correlation P
lation Plot o
tocorrelatio
d, method ohood metho
r nonstation
063253. Email:d Sciences & TeNGR.COM/V11/
Plot of Orig
of First Dif
on Plot of Fi
of moments.od. After d
narity. The A
: fahimparachaechnologies. Vol11A04J.pdf DO
ginal Tradin
fferenced Tr
irst Differen
R package developing t
ACF does n
[email protected] lume 11 No.4 ISOI: 10.14456/IT
ng Signals S
rading Signa
nced Tradin
forecast (Hythe model,
not cuts off,
©2019 InternaSSN 2228-9860 TJEMAST.2020.7
Series.
als Series.
ng Signals S
Hyndman & Kwe check
, rather it sh
ational eISSN
70 11
Series
Khandakar,the model
hows a slow
, l
w
12 Been
decrease. Osome lags. from the plis stationaris the ordermodels, A
4.1 MODWe de
70:30 ratiothe testing (Hyndman
In Figuinterval is sautocorrelaand then 10of the test Figure 13 r
The mshows thatalgebraical(1,1,1) wit
nish Bashir, Fa
On the otheWe can per
lot of the orry. Since it tr of AR and
ARIMA(0,1
DEL BUILevelop threeo. The total ndataset has
n & Khandakure 11, the shown in daation betwee0 is out of thstatistic 0.0
represents a
model selectit all AIC vally lower AIth AIC va
aheem Aslam
r hand, the rform anothriginal data atook one difd MA part i1,1), ARIMA
DING ANDe ARIMA mnumber of othe remainikar, 2007) aforecasts fo
ark shaded aen forecast he significa0287 confira more or le
Figure 1
ion among talues are inIC is selectelue -3923.6
first differeher formal teand the ACFfferencing tois determineA(1,1,0) and
D DIAGNOmodels withobservationsing 1545 oband packageor the next 1area and 95%errors then t
ance boundsrming the ess normal d
Figure 11
12: ACF Plo
the above ten negative med rather tha65 and pre
ence data loest to test theF plot provio get stationed through td ARIMA(
OSTIC h orders as ps is 5151, th
bservations.e Metrics (H10 days are% predictiothe model c
s. We carry evidence of distribution.
1: Ten days
ot of Residu
ested traditimeans that tan the absolecision valu
ooks much se stationaritide a good inarity, here dthe plots of 1,1,1).
proposed eahe training dFor estimat
Hamner et aplotted as a
on interval acannot be imout a unit ro
f non-zero c
s Forecast.
ual ARIMA
ional modelthe likeliholute value oue 0.1333.
stationery. Tty of the diffndication thd=1. The va
f ACF and P
arlier. We bidata set has tion purpose
al., 2018) ara blue line, s a light sha
mproved. In oot test whecorrelation.
(1, 1, 1).
ls is based oood at the mf AIC whichRight-align
The ACF cufferenced dahat the diffealues of p anPACF that m
ifurcate the3606 obseres, R packa
re deployedwhere 80%
aded area. IfFigure 12, s
ere KPSS haThe foreca
on AIC valumaximum wh in our casned averag
uts off afterata. But hererenced datand q, whichmakes three
e data into arvations andage Forecast.
% predictionf there is nospike 2, 4, 7as the valueast errors in
ues. Table 3was > 1 and
e is ARMAges and the
r e a h e
a d t
n o 7 e n
3 d A e
*CorrespondinTransaction Jo1906-9642 CO
exponentia
The A
hidden andfunction. Tindicators parameters30% of theestimated ahighest preelm_train()sigmoid anprecision vthe highest
ng author (Fahournal of EngineODEN: ITJEA8
ally weighte
TeA
ANN is applid output. ANThe network
through ths are fixed fe data set. Tagainst threecision valu) and elm_pnd tanh are uvalue is 0.50t precisions
eem Aslam). Teering, ManagemPaper ID:11A04
ed average h
Figure 1
Tabl
Table 4chnique ANN
SVM
ELM
ied in packaNN model uk has 15 inhe random for this ANNThe highestee activatioue 0.5265 ipredict(). Thused and th000 with acscore with
Tel: +92-321-50ment, & Applied4J http://TUEN
has no repor
3: Histogram
le 3: CompaModeSMACMARMAEMA
ARMA(0ARMA(1ARMA(1
4: ComparisActivation fu
ReLUTanh
SigmoiPolynomRadial B
TanhSigmoiReLUTanh
age nnet of Rutilized in thnput neurons
forest (TabN. The netwt precision sn functionsin polynomhe number
he input layectivation funactivation f
063253. Email:d Sciences & TeNGR.COM/V11/
rted value o
am of Residu
arison of Trel A A A A 0,1,1)
,1,0) ,1,1)
son of Soft function U h id
mial Basis h id
U h
R. The feedhis study ems that correble 1). Thework is trainscore is 0.6s which are
mial activatioof hidden ner has 15 senction sigmfunction ReL
: fahimparachaechnologies. Vol11A04J.pdf DO
of AIC.
ual ARIMA
raditional MAIC
-3151.602 -3896.784
Null Null
-3680.29 -2611.42 -3923.65
ComputingPrecision
0.630.580.590.520.480.490.500.420.44
d-forward Amploys the siespondents te output laned on 70%
6311 with ac, polynomion functionnodes is 10,elected inputmoid. The soLU for KSE
[email protected] lume 11 No.4 ISOI: 10.14456/IT
A (1, 1, 1).
Model
g Model. n Score 11 36 45 65 30 50 00 85 10
ANN model igmoid, tanhto the 15 seayer has a
% data and tectivation fual, tanh and
n. ELM is athree activ
t technical ioft computinE-100 tradin
©2019 InternaSSN 2228-9860 TJEMAST.2020.7
has three lanh and ReLUelected inpupredicted
ested on theunction ReLd radial basapplied usin
vation functiindicators. Tng models
ng signal dat
ational eISSN
70 13
ayers: input,U activationut technicalsignal. All
e remainingLU. SVM issis with theng functionions ReLU,The highestANN givesta as shown
, n l l g s e n , t s n
14 Beenish Bashir, Faheem Aslam
in Table 4. The analysis of traditional econometric techniques only ARMA (1,1,1) is the best model but the precision score is quite low (Table 5). These results are in line with previous studies (Hsu, Lessmann, Sung, Ma, & Johnson, 2016; Rojas et al., 2008) which also assert the better predictive performance of soft computing models over traditional models.
Table 5: Comparison of Best Traditional Model and Best Soft Computing Model Model Type Model Precision Score
Traditional Model ARMA(1,1,1) 0.1333 Soft Computing Model ANN 0.6311
5. CONCLUSION This study brings forward the unsettled debate in the literature about modeling techniques for
trading signals prediction. We cover a vast range of models from moving averages to autoregressive models and then soft computing models. Our detailed analysis finds the best of the traditional models and best of soft computing models but soft computing models perform better than traditional forecasting models for trading signals prediction. These findings are important for traders who can forecast trading signals on the basis of the soft computing model rather than the traditional model. Our analysis also confirms the non-linear behavior of time series financial data which can be better handled through the soft computing model.
6. AVAILABILITY OF DATA AND MATERIAL Data can be made available by contacting the corresponding authors
7. REFERENCES Atsalakis, G. S., & Valavanis, K. P. (2009). Surveying stock market forecasting techniques–Part II: Soft
computing methods. Expert systems with Applications, 36(3), 5932-5941.
Boser, B. E., Guyon, I. M., & Vapnik, V. N. (1992). A training algorithm for optimal margin classifiers. Paper presented at the Proceedings of the fifth annual workshop on Computational learning theory.
Brockwell, P. J., & Davis, R. A. (2009). Time Series: Theory and Methods, (Springer Series in Statistics).
Cheng, C.-H., & Wei, L.-Y. (2014). A novel time-series model based on empirical mode decomposition for forecasting TAIEX. Economic Modelling, 36, 136-141.
Čisar, P., & Čisar, S. M. (2011). Optimization methods of EWMA statistics. Acta Polytechnica Hungarica, 8(5), 73-87.
Cryer, J. D., & Chan, K.-S. (2008). Time series regression models. Time series analysis: with applications in R, 249-276.
Fama, E. F. (1965). The behavior of stock-market prices. The journal of Business, 38(1), 34-105.
Hamner, B., Frasco, M., & LeDell, E. (2018). Package ‘Metrics’.
Hsu, M.W., Lessmann, S., Sung, M.C., Ma, T., & Johnson, J.E. (2016). Bridging the divide in financial market forecasting: machine learners vs. financial economists. Expert systems with Applications, 61, 215-234.
Hu, Y., Feng, B., Zhang, X., Ngai, E., & Liu, M. (2015). Stock trading rule discovery with an evolutionary trend following model. Expert systems with Applications, 42(1), 212-222.
Huang, C.-L., & Tsai, C. Y. (2009). A hybrid SOFM-SVR with a filter-based feature selection for stock market forecasting. Expert Systems with Applications, 36(2), 1529-1539.
*Corresponding author (Faheem Aslam). Tel: +92-321-5063253. Email: [email protected] ©2019 International Transaction Journal of Engineering, Management, & Applied Sciences & Technologies. Volume 11 No.4 ISSN 2228-9860 eISSN 1906-9642 CODEN: ITJEA8 Paper ID:11A04J http://TUENGR.COM/V11/11A04J.pdf DOI: 10.14456/ITJEMAST.2020.70
15
Huang, J.Z., Huang, W., & Ni, J. (2019). Predicting bitcoin returns using high-dimensional technical indicators. The Journal of Finance and Data Science, 5(3), 140-155.
Huang, W., Nakamori, Y., & Wang, S.-Y. (2005). Forecasting stock market movement direction with support vector machine. Computers & operations research, 32(10), 2513-2522.
Hyndman, R. J., Athanasopoulos, G., Bergmeir, C., Caceres, G., Chhay, L., O'Hara-Wild, M., . . . Razbash, S. (2019). Package ‘forecast’. Online] https://cran. r-project. org/web/packages/forecast/forecast. pdf.
Hyndman, R. J., & Khandakar, Y. (2007). Automatic time series for forecasting: the forecast package for R: Monash University, Department of Econometrics and Business Statistics ….
Kara, Y., Boyacioglu, M. A., & Baykan, Ö. K. (2011). Predicting direction of stock price index movement using artificial neural networks and support vector machines: The sample of the Istanbul Stock Exchange. Expert systems with Applications, 38(5), 5311-5319.
Kumar, D. A., & Murugan, S. (2013). Performance analysis of Indian stock market index using neural network time series model. Paper presented at the 2013 International Conference on Pattern Recognition, Informatics and Mobile Engineering.
Lam, M. (2004). Neural network techniques for financial performance prediction: integrating fundamental and technical analysis. Decision Support Systems, 37(4), 567-581.
Lee, M.-C. (2009). Using support vector machine with a hybrid feature selection method to the stock trend prediction. Expert systems with Applications, 36(8), 10896-10904.
Lin, C.-S., Chiu, S.-H., & Lin, T.-Y. (2012). Empirical mode decomposition–based least squares support vector regression for foreign exchange rate forecasting. Economic Modelling, 29(6), 2583-2590.
Lucas, J. M., & Saccucci, M. S. (1990). Exponentially weighted moving average control schemes: properties and enhancements. Technometrics, 32(1), 1-12.
Malkiel, B. G., & Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. The journal of finance, 25(2), 383-417.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. The bulletin of mathematical biophysics, 5(4), 115-133.
Moradi, S., & Rafiei, F. M. (2019). A dynamic credit risk assessment model with data mining techniques: evidence from Iranian banks. Financial Innovation, 5(1), 15.
Murphy, J. J. (1986). Technical Analysis of the Futures Markets, New York Institute of Finance. Englewood Cliffs, NJ.
Oliveira, A. L., & Meira, S. R. (2006). Detecting novelties in time series through neural networks forecasting with robust confidence intervals. Neurocomputing, 70(1-3), 79-92.
Patel, J., Shah, S., Thakkar, P., & Kotecha, K. (2015). Predicting stock and stock price index movement using trend deterministic data preparation and machine learning techniques. Expert systems with Applications, 42(1), 259-268.
Pfaff, B., Zivot, E., Stigler, M., & Pfaff, M. B. (2016). Package ‘urca’. Unit root and cointegration tests for time series data. R package version, 1.2-6.
Rhoades, S. A. (1993). The herfindahl-hirschman index. Fed. Res. Bull., 79, 188.
Rojas, I., Valenzuela, O., Rojas, F., Guillén, A., Herrera, L. J., Pomares, H., . . .Pasadas, M. (2008). Soft-computing techniques and ARMA model for time series prediction. Neurocomputing, 71(4-6), 519-537.
Shah, A., Zeileis, A., & Grothendieck, G. (2005). zoo Quick Reference. Package vignette.
16 Been
Svetunkov,
Tay, F. E., Om
Teixeira, Lana688
Torgo, L. (2
Torgo, L. pre
Tsinaslanidand
Tukey, J. E
Ulrich, J., U
Vanstone, syst
Wang, B., fore
Wang, J.-Zpro
Weng, B., Ausin
Zakamulin,Rul
Żbikowski,featApp
Zhiqiang, Gopt
Zhong, X.,Exp
Zhong, X.,mac
Trademarkthis only
nish Bashir, Fa
, I. (2017). S
& Cao, L. (2mega, 29(4),
. A., & De Oalysis and 85-6890.
2011). Data
(2014). Andictive mod
dis, P. E., & d dynamic tim
Exploratory d
Ulrich, M. J
B., & Finntems using a
Huang, H.ecasting. Ne
Z., Wang, Jopagation ne
Ahmed, M. ng disparate
, V. (2017). les: Springer
, K. (2015). ture selectioplications, 4
G., Huaiqingimized by P
& Enke, Dpert systems
& Enke, Dchine learnin
Beenish BasShe was a visMarkets, Fin
Dr. FaheemIslamabad. HKorea. His re
ks Disclaimearticle are th
y. Use of them
aheem Aslam
Statistical m
2001). Appl309-317.
Oliveira, A. Lnearest ne
a mining with
n infrastrucdels in r. arX
Kugiumtzisme warping
data analysi
., & RUnit,
nie, G. (200artificial neu
, & Wang, eurocomputi
J.-J., Zhang,ural network
A., & Megae data source
Market Timr.
Using voluon for the
42(4), 1797-
g, W., & QuPSO. Soft Co
D. (2017). F with Applic
D. (2019). Png algorithmshir is a PhD ssiting research
nancial Risk Ma
Aslam is an AHe earned his Mesearch intere
er: All produhe property om does not im
odels underl
lication of su
L. I. (2010).ighbor clas
h R: learnin
ture for peXiv preprint a
s, D. (2014)g. Expert sys
s 1977 Read
S. (2018). P
09). An empural network
X. (2012).ing, 83, 136-
, Z.-G., & k. Expert sys
ahed, F. M. es. Expert sy
ming with Mo
ume weightepurpose o
1805.
uan, L. (201omputing, 17
orecasting dcations, 67,
redicting thms. Financiascholar at Deph student at Uanagement, M
Assistant ProfMaster's and Pests are Financ
uct names incf their respecmply any en
lying functio
upport vecto
A method fssification.
ng with case
erformance arXiv:1412.
). A predictiostems with Ap
ding. MA Ad
Package ‘TT
pirical methks. Expert sy
A novel t-145.
Guo, S.-P. ystems with A
(2017). Stocystems with A
oving Averag
ed support vf creating
13). Financia7(5), 805-81
daily stock 126-139.
he daily retual Innovationpartment of MUniversity of CMachine Learn
fessor at DepaPh.D. Degrees cial Analytics
cluding tradective owners,ndorsement o
ons of'smoo
or machines
for automaticExpert sys
studies: Ch
estimation 0436.
on scheme upplications,
ddison-Wesle
R’.
hodology foystems with A
ext mining
(2011). FoApplications
ck market onApplications
ges: The An
vector machistock tradin
al time serie8.
market retu
urn directionn, 5(1), 4.
Management SCôte d'Azur, Fring and Data S
artment of Mafrom Hanyang, Data Mining,
emarks™ or r, using for id
or affiliation.
th package f
in financial
c stock tradistems with
apman and H
and experi
using percep 41(15), 684
ey.
or developinApplications
approach t
recasting sts, 38(11), 14
ne-day aheas, 79, 153-16
atomy and P
ines with wang strategy
es forecastin
urn using dim
n of the stoc
ciences, COMrance. Her reseScience.
nagement Scig University Bu, Artificial Inte
registered® tentification a
for R.
time series
ing combininApplicatio
Hall/CRC.
imental com
ptually impo48-6860.
ng stockmars, 36(3), 666
to financial
tock indices4346-14355.
ad movemen63.
Performance
alk forwardy. Expert sy
ng using LP
mensionality
ck market u
MSATS Universiearch interest
iences, COMSAusiness Schoolelligence and
trademarks mand educatio
forecasting.
ng technicalns, 37(10),
mparison of
ortant points
rket trading68-6680.
time series
s with back
nt prediction
e of Trading
d testing andystems with
P and SVM
y reduction.
using hybrid
ity Islamabadt are Financia
ATS Universityl, Seoul, SouthData Science
mentioned innal purposes
.
l ,
f
s
g
s
k
n
g
d h
M
.
d
. l
y h .
n s