APPLICATION OF COMBINED ARMA-NEURAL NETWORK
MODELS TO PREDICT STOCK PRICES
Mohammad Umair Yaqub
Electrical Engineering Department
King Fahd University of Petroleum and
Minerals, KFUPM Box 767
Dhahran 31261, Saudi Arabia
+966535971232
umairy@kfupm.edu.sa
ABSTRACT
Predicting the prices of stocks is a very important field of research
nowadays. Many different types of models have already been
implemented. Two types of techniques include ARMA models and
Neural Networks. In this work, ARMA models, along with two
types of Neural Networks (Back Propagation) and Multi-Layer
Perceptron (MLP) have been used. Furthermore, the two neural
networks were combined with ARMA models (individually) in
order to generate the best forecasted prices. The two indexes used
for forecasting are the Dow Jones Industrial Average (DJI) and the
Saudi Stock Exchange Tadawul (TASI). 800 values were used to
predict the future 200 values. It was found out that for such large
number of predictions, MLP yields the best results, which are
drastically improved when combined with ARMA forecasting.
CCS Concepts
• Applied Computing ➝ Law, social and behavioral sciences➝
Economics • Computing methodologies ➝ Machine learning ➝
Machine learning approaches ➝ Neural networks
Keywords
Auto-Regressive Moving Average (ARMA); Back-Propagation
Neural Network (BPNN); Multi-Layer Perceptron Neural Network
(MLP-NN); stock forecasting; Artificial Intelligence (AI).
1. INTRODUCTION
The prediction of stock prices and trends is a very hot field of
research. Many people invest their money in various assets
nowadays and they would like to know the trends beforehand in
order to make the most informed decisions possible. A lot of work
has been done to predict the future values of various indices in the
stock market. Some of the techniques applied include financial
mathematics and regression method such as the Auto-Regressive
(AR) model, the Moving Average (MA) model and the combined
ARMA model.
A different set of techniques were employed including methods
from Artificial Intelligence (AI). AI techniques include Neural
Networks (NN), support vector machines (SVM), genetic
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, or republish, to post on servers or to
redistribute to lists, requires prior specific permission and/or a fee. Request
permissions from Permissions@acm.org.
MISNC, SI, DS '16, August 15-17, 2016, Union, NJ, USA
© 2016 ACM. ISBN 978-1-4503-4129-5/16/08…$15.00
DOI: http://dx.doi.org/10.1145/2955129.2955171
Mohammad Saad Al-Ahmadi
Accounting and MIS Department
King Fahd University of Petroleum and
Minerals, KFUPM Box 443
Dhahran 31261, Saudi Arabia
+966552055577
alahmadi@kfupm.edu.sa
algorithms and fuzzy logic. The most recent papers have tried to
improve these individual methods by combining ARMA models
with one or more of these AI techniques. In this research, we follow
the works of Shi and his colleagues [13], [14] as well as Kashei et
al [12].
An attempt is made to forecast the daily closing prices/values for
the Microsoft Corporation, Dow Jones Industrial Average (DJI),
and Saudi Stock Exchange Tadawul (TASI). ARMA models are
used to predict the stock prices and then Back-Propagation Neural
Network (BPNN) and Multi-Layer Perceptron Neural Network
(MLP-NN) is used to optimize the residual errors form the ARMA
forecast. The results of the individual models as well as the
combined techniques are then tabulated and a comparison is made.
The paper is split into the following sections. Section 2 is a brief
literature review about the work done in literature by other
researchers. Section 3 is a description of some of the theory behind
the techniques used. ARMA models are discussed and Neural
Networks are described in general with an emphasis on BPNN’s
and MLP-NN’s. In the next chapter, the overall process is described
and the details about the data are presented. Lastly, section 5
presents the various experiments performed along with the details
of the models and tabulates the results. This part is performed using
three sub-sections for each stock price dataset. Finally, we conclude
our project with a hint on some future directions possible.
2. LITERATURE REVIEW
Due to the nature and importance of the subject, various techniques
have been applied in research to predict the future prices of stocks.
Two such techniques are the Auto-Regressive Moving Average
(ARMA) models and the artificial intelligence technique called
Neural Networks (NN’s). In this section, we mention a couple of
works that use financial models for prediction, then we introduce
AI techniques and finally we focus on some work in literature that
combines the two techniques to obtain better results.
The work of Anaghi and Nourouzi [1] discusses in details how to
forecast prices using ARMA models. They analyze six different
transfer functions using equal number of poles and zeros at each
stage. The work of Chang and Ramakrishna [2] is slightly different
in that they compare the performances of ARMA, Generalized
Autoregressive Conditional Heteroskedasticity (GARCH), and NN
models. It is found that GARCH and NN models are more effective
than ARMA models alone. This is one of the motivation to combine
BPNN with ARMA in order to achieve higher accuracy.
As far as AI techniques are concerned, Wu and Lu have covered a
wide range of NN based algorithms in their study [3]. The
algorithms covered include many different types of neural networks
that are compared with respect to mean square errors as well as
computation times. Some of the neural networks covered include
Self-Organizing Polynomial (SOP) NN’s and many types of Back
Propagation Neural Networks (BPNN’s).
The idea of using only NN’s to predict the future stock price has
been explored in literature [4] - [6]. The work of Senanayake [4] is
concerned with using automated NN’s for forecasting. They
employ an implementation of BPNN to predict the price depending
on historical process as well as qualitative factors by questionnaires
filled by experts. The work of Chih-Ming Hsu [5] is a more modern
work that focuses on prediction based on two stages. In the first
stage, the prices are normalized between 0 and 1 and then feature
extraction is employed via the mean absolute percentage error
criteria. Finally, Khoa et al. [6] also propose a similar technique to
[5] but the features used in the second stage are the adjusted profit
margins.
The research of Gomes et al. [9] again utilizes the same concept of
hybridization but they use a specific Neural Network called DAN2
designed by the authors of [10] and [11]. Lastly, the work done in
[12] is focused on combine an ARIMA (AR Integrated MA) model
with probabilistic Neural Networks (MLP-NN’s). However, they
apply this to exchange rate prediction rather than stock prices.
Finally, this takes us to the work done in [13], which is done by the
same authors as our reference paper. IN this model, they apply the
ARMA to the price of Petro China and fit the errors using BPNN
model. Then they try to improve the accuracy by using the Hidden
Markov Model (HMM) algorithm.
3. BRIEF THEORETICAL BACKGROUND
In this research, three different models have been used for
forecasting. The first one is the Auto-Regressive Moving Average
(ARMA) Model whereas the second model involves artificial
Neural Networks (NN’s). Two different types of NN’s have been
used, the Back-Propagation NN (BPNN) and the Multi-Layer
Perceptron NN (MLP-NN). In the original papers [1], they used the
ARMA model to predict the prices and then employed BPNN to
optimize the errors. In our case, experiments are performed with
both BPNN and MLP. Prior to describing the complete
methodology of our system, we would like to begin by presenting
some theoretical background on these.
There is a constant c, which can be taken as zero for simplicity as
well as the expected value µ. In this research, MATLAB is used as
a simulation tool to obtain the best values for p and q based on the
minimum AIC criteria.
3.2 Neural Networks (NN’s)
NN’s were introduced to simulate the behavior of the human brain.
Usually, they are formed of three layers. The input layer, one or
more hidden layers and the output layer. The job of the hidden layer
is to optimize the relation between the outputs and the inputs using
a heuristic or iterative approach such as the steepest descent method.
In back-propagation, the hidden layer nodes or weights are
optimized using the gradient descent method. On the other hand, in
MLP, there is no feedback and the weights are updated based on an
activation function (in our case the logistic function is used). The
output of any NN is the sum of the weights multiplied by the
connected input nodes. MLP-NN is the most basic and earliest
neural network and most of the other architectures are derived from
it in literature. Therefore, the idea is to try the simplest NN in this
project and test for improvements against the BPNN.
4. PREDICTION MODEL
4.1 System Design and Data Description
The work done by Shi et al. uses the ARMA and BPNN models to
fit the data and choose the parameters that yield the least error [13]
and [14]. In the next step, both ARMA models and BPNN are used
to forecast the prices individually. In the combining step,
forecasting is done using the ARMA model and then BPNN is used
to fix the errors introduced by the first stage. On the other hand, the
work of [12] employs the same mechanism but for exchange rate
prediction and instead of BPNN, MLP is used. In this project, the
methodology of [13] and [14] is followed except that the last step
of Hidden Markov Models (HMM) is not applied. Instead, a new
model is tried where the BPNN is replaced by MLP. The
methodology is summarized in Figure 1. The ARMA and BPNN
modelling is done using MATLAB whereas the MLP regression is
performed using DTREG tool.

3.1 Auto-Regressive Moving Average

(ARMA)
The concept of ARMA filters was introduced in digital signal
processing DSP. The AR model assumes that a time-series is a
summation of the previous p number of terms itself weighted by
some parameter Φ. The errors are represented by e, which is
assumed to be white noise of Gaussian distribution. Equation 1
describes the AR model.
𝑝
𝑋𝑡 = 𝑐 + ∑𝑖=1 𝜑𝑖 𝑋𝑡−𝑖 + 𝑒𝑡    
The error terms can be estimated using the MA model as shown
below in equation 2. Usually, e is assumed to have normal
distribution and it is estimated using q terms.
𝑋𝑡 = 𝜇 + ∑𝑖=1 𝜃𝑖 𝑒𝑡−𝑖 + 𝑒𝑡 
𝑞

The combined model is given by equation 3. The total order of the

systems is defined as ARMA (p, q) where p represents the AR terms
and q represents the MA terms. The challenge is to develop a model
by selecting the best possible numbers of p and q as well as
estimating the parameters Φ and θ.
𝑝 𝑞
𝑋𝑡 = 𝑐 + 𝜇 + 𝑒𝑡 + ∑𝑖=1 𝜑𝑖 𝑋𝑡−𝑖 + ∑𝑖=1 𝜃𝑖 𝑒𝑡−𝑖 (3)
 Figure 1: Complete System Process including BPNN and
MLP-NN
As one of our contributions, BPNN was replaced by MLP to test
whether it can improve on the results, especially for large datasets
such as the one chosen for DJI and TASI. Furthermore, the
experiments done in the reference papers involve the closing prices
of the Microsoft Corporation and Petro China; 320 samples are
used where the future prices of the 20 samples are predicted using
all of the models and the results are compared. In this experiment,
we try to predict the close values of the Dow Jones Industrial
Average (DJI) and Saudi Stock Exchange (TASI). Furthermore,
larger datasets are used where the data of approximately previous
four years is employed. This comes down to 1000 values, where
800 values are used to train the models and an attempt is made to
predict the future 200 values. Figure 2: ARMA-Fitted Data versus original for DJI

4.2 Error Definitions

Let us define the errors used to evaluate the performances in this
project. For the Microsoft Corporation, since the forecasted values
are only 20, it is feasible to show the results in one table. However,
for the DJI and TASI, 200 points are forecasted and it is not
practical to tabulate the results. Therefore, the performances have
been summarized in terms of the mean absolute percentage error
(MAPE), mean square error (MSE), root mean square error (RMSE)
and normalized root mean square error (NRMSE - percentage). The
formulae are given in equations 4-7.
1 𝑥𝑘 −𝑦𝑘
𝑀𝐴𝑃𝐸 = 100 ∗ ∑𝑁
𝑘=1 (6)
𝑁 𝑥𝑘
Figure 3: ARMA-Fitted Data versus original for TASI

1
𝑀𝑆𝐸 = ∑𝑁
𝑘=1(𝑥𝑘 − 𝑦𝑘 )
2
(7)
𝑁

1
𝑅𝑀𝑆𝐸 = √ ∑𝑁
𝑘=1(𝑥𝑘 − 𝑦𝑘 )
2 (8)
𝑁

1 𝑁
√ ∑ (𝑥 −𝑦𝑘 )2
𝑁 𝑘=1 𝑘
𝑁𝑅𝑀𝑆𝐸 = (9)
𝑚𝑎𝑥(𝑥)−𝑚𝑖𝑛(𝑥)

5. EXPERIMENTS AND RESULTS

As mentioned in the methodology, the experiments were carried out
using ARMA, BPNN, MLP, combined ARMA-BPNN and ARMA-
MLP. Two datasets are used: DJI and TASI from May 2010 to May Figure 4: BPNN-Fitted Data versus original closing values for
2014. Before displaying the results, let us display the fitted plots DJI
for all three indices.

5.1 Data Fitting Results

5.1.1 ARMA fitting
Based on the minimum AIC criterion, the ARMA model was
developed for both datasets. It was discovered that the best models
turned out to be ARMA (1, 5) for the DJI and ARMA (3, 5) for
TASI. The results of fitting the training data for all three are
presented in Figures 3 and 4. As you can observe, the fitting was
quite accurate and an initial forecast was made using this model.
The errors were later corrected using the two NN models.

5.1.2 BPP-NN fitting

In terms of fitting, BPNN performed very well as shown in Figures Figure 5: BPNN-Fitted Data versus original closing values for
4 and 5. However, as we will observe later, this model was not able TASI
to predict the future error values accurately. So although BPNN
provided a good fitting model, it was not very useful for the
purposes of prediction.
5.1.3 MLP-NN fitting
For this method, it was discovered that six hidden nodes give the
best results in terms of the MSE. Since the values are larger than
one (not normalized), this model should give the least RMSE and
NRMSE as well. As it can be observed, MLP-NN does not provide
a very good fitting model, but it is a better tool for prediction as it
was able to follow the trend better. This will be clear from the final
forecasting results in sections 5.2 and 5.3.

5.2 Forecasting Results for DJI

For the DJI as well as for the TASI, the forecasted range has 200
results. Therefore, a summary is provided below in Table 1 and
graphically presented in Figure 8. The results are summarized in Figure 8: Results of forecasting 200 stock values for DJI using
the following manner: Two subsections are used to detail the all models
models used for each index. The results are then tabulated as
follows: the model used is specified in the first column and each of
the four types of errors discussed in section IV part B are listed in Table 1: Summary of the Results for DJI
the next 4 columns of the same row. Model MAPE MSE RMSE NRMSE
It can clearly be observed for DJI that the best results are obtained ARMA 2.21 172990.00 415.92 7.17
when the MLPNN is used in conjunction with the ARMA model.
In terms of all of the errors, it seems that the ARMA-BPNN will MLP 1.46 55168.16 234.88 9.01
not work for DJI whereas MLP-NN alone improves the result ARMA- 2.24 173750.00 416.83 7.19
significantly. Overall, the best result is provided by the composite BPNN
ARMA-MLP technique and by observing Figure 8, the reason is ARMA- 0.64 18209.00 134.94 2.33
clear as the predicted values follow the actual very closely. This MLP
implies that the MLP-NN was the best stand-alone model in terms
of following the error trend.
5.3 Forecasting Results for TASI
Similar to what has been done for DJI, the 200 values have been
forecasted for TASI and the resulting errors have been summarized
in Table 2 and graphically displayed in Figure 9. Once again, the
pattern is similar. However, for this data the effects are even more
severe. The ARMA and ARMA-BPNN models fail to track the
progress of the curve and move in the opposite direction to the
general trend. On the other hand, MLP alone does follow the
direction of the trend but the error is still large. In contrast,
combining the both models yields the most successful results.

Figure 6: MLP-NN Fitted Data versus original closing values

for DJI

Figure 9: Results of forecasting 200 stock values for DJI using

all models
Table 2: Summary of the Results for TASI
Model MAPE MSE RMSE NRMSE
ARMA 16.12 2765000.00 1662.80 63.78
MLP 2.50 55168.16 234.88 22.15
Figure 7: MLP-NN Fitted Data versus original closing values ARMA- 16.35 2812900.00 1677.20 64.33
for DJI BPNN
ARMA- 0.58 4076.30 63.85 2.45
MLP
6. CONCLUSION
In conclusion, in this experiment ARMA, BPNN, MLP and
combined ARMA-BPNN and ARMA-MLPNN models were used
to predict stock prices/values. 800 historical values were used to
predict the future 200 values form May 2010 all the way to May
2014 for DJI and TASI indexes. By applying the individual ARMA
and MLP models, it was discovered that the best stand-alone model
was MLP for both datasets, as the MAPE was reduced form 2.21%
to 1.46% and 16.12% to 2.5% for both cases. Furthermore,
combining ARMA with BPNN resulted in no improvements.
However, when ARMA model was combined with MLP-NN, it
resulted in the best improvement. The MAPE went all the way
down to about 0.5% for both cases. In addition, the NRMSE was
also reduced all the way to about 2.5%.
In the original paper, combining ARMA with BPNN results in a
marked improvement. However, in our case, there is no
improvement when that particular model is used. This has to be
investigated and the reason may be that BPNN is only suitable for
prediction of small number of values. Moreover, he original authors
apply Hidden Markov Models to improve the predictions and this
avenue can be explored further. It is possible that ARMA-BPNN
may not improve the results but with HMMM, the accuracy may
increase.
7. ACKNOWLEDGMENTS
The authors would like to acknowledge the support provided by
King Abdulaziz City for Science and Technology (KACST) through
the Science & Technology Unit at King Fahd University of
Petroleum & Minerals (KFUPM) for funding this work through
project No. 09-INF769-04 as part of the National Science,
Technology and Innovation Plan.
8. REFERENCES
[1] Anaghi, M.F.; Norouzi, Y., "A model for stock price forecasting based
on ARMA systems," Advances in Computational Tools for
Engineering Applications (ACTEA), 2012 2nd International
Conference on , vol., no., pp.265,268, 12-15 Dec. 2012
[2] W. A. Chang and R. S. Ramakrishna, “A genetic algorithm for shortest
path routing problem and the sizing of populations,” IEEE
Transactions on Evolutionary Computation, vol. 6, no. 6, pp. 566–579,
2002. 2010.
[3] Hossain, A.; Nasser, M., "Comparison of GARCH and neural network
methods in financial time series prediction," Computer and
Information Technology, 2008. ICCIT 2008. 11th International
Conference on , vol., no., pp.729,734, 24-27 Dec. 2008
[4] Senanayake, A., "Automated neural-ware system for stock market
prediction," Cybernetics and Intelligent Systems, 2004 IEEE
Conference on , vol.2, no., pp.1166,1171, 1-3 Dec. 2004
doi: 10.1109/ICCIS.2004.1460755
[5] Chih-Ming Hsu, "Forecasting stock/futures prices by using neural
networks with feature selection," Information Technology and
Artificial Intelligence Conference (ITAIC), 2011 6th IEEE Joint
International , vol.1, no., pp.1,7, 20-22 Aug. 2011
doi: 10.1109/ITAIC.2011.6030137
[6] Khoa, N.L.D.; Sakakibara, K.; Nishikawa, I., "Stock Price Forecasting
using Back Propagation Neural Networks with Time and Profit Based
Adjusted Weight Factors," SICE-ICASE, 2006. International Joint
Conference , vol., no., pp.5484,5488, 18-21 Oct. 2006
doi: 10.1109/SICE.2006.315683
[7] Sallehuddin, R.; Shamsuddin, S.M.; Hashim, S.Z.M., "Hybridization
Model of Linear and Nonlinear Time Series Data for Forecasting,"
Modeling & Simulation, 2008. AICMS 08. Second Asia International
Conference on , vol., no., pp.597,602, 13-15 May 2008
doi: 10.1109/AMS.2008.142
[8] Abounoori, E.; Tazehabadi, A.G., "Forecasting Stock Price Using
Macroeconomic Variables: A Hybrid ARDL, ARIMA and Artificial
Neural Network," Information and Financial Engineering, 2009.
ICIFE 2009. International Conference on , vol., no., pp.149,153, 17-
20 April 2009
doi: 10.1109/ICIFE.2009.23
[9] Gomes, G.S.S.; Maia, A.L.S.; Ludermir, T.B.; de A.T.de Carvalho, F.;
Araujo, A. F R, "Hybrid model with dynamic architecture for
forecasting time series," Neural Networks, 2006. IJCNN '06.
International Joint Conference on , vol., no., pp.3742,3747, 0-0 0
doi: 10.1109/IJCNN.2006.247391
[10] M. Ghiassi e H. Saidane, A dynamic architecture for artificial neural
network, Journal of Neurocomputing, 63, 397–413, 2005
[11] M. Ghiassi, H. Saidane e D. K. Zimbra, A dynamic artificial neural
network model for forecasting time series events, Journal of
forecasting, 21, 341–362, 2005
[12] Khashei, M., Bijari, M., & Raissi Ardali, G. A. (2012). Hybridization
of autoregressive integrated moving average (ARIMA) with
probabilistic neural networks (PNNs). Computers & Industrial
Engineering, 63(1), 37-45. doi: 10.1016/j.cie.2012.01.017
[13] Shuzhen Shi; Wenlong Liu; Minglu Jin, "Stock price forecasting using
a hybrid ARMA and BP neural network and Markov model,"
Communication Technology (ICCT), 2012 IEEE 14th International
Conference on , vol., no., pp.981,985, 9-11 Nov. 2012
doi: 10.1109/ICCT.2012.6511341
[14] Shuzhen Shi; Wenlong Liu; Minglu Jin, " Stock Price Forecasting
Based on a Combined Model of ARMA and BP Neural Network and
Markov Model," International Journal of Information Processing and
Management(IJIPM), vol. 4, no. 3, 9-11 May. 2013