Accepted Manuscript

A Modified Neural Network Model for Predicting the Crude Oil Price

Mohammad Reza Mahdiani

PII: S1822-8011(16)30012-1
DOI: 10.1016/j.intele.2017.02.001
Reference: INTELE 23

To appear in: Intellectual Economics

Received date: 15 March 2016

Accepted date: 20 February 2017

Please cite this article as: Mohammad Reza Mahdiani , A Modified Neural Network Model for Predicting
the Crude Oil Price, Intellectual Economics (2017), doi: 10.1016/j.intele.2017.02.001

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

A Modified Neural Network Model for Predicting the Crude Oil Price

Mohammad Reza Mahdiani1*

1
Faculty of Petroleum Engineering, Amirkabir University of Technology (Tehran Polytechnic), Hafez Avenue, Tehran, Iran

T
Abstract

IP
Predicting the crude oil price is very interesting and prominent for all governments and different

CR
companies. Here a neutral network model has been modified, in a way that a genetic algorithm

optimizes its parameters during its run. This model is able to predict the future price of crude oil

US
very accurately. Comparing the model of this study with the pure neural network in predicting

both daily and monthly crude oil prices confirmed its greater performance, especially in situation
AN
of small number of input data for training or the great changes of variables.
M

Keywords: artificial neural network, genetic algorithm, modeling

ED

1. Introduction

Predicting the crude oil price is very important in different economic, political and industrial
PT

areas, both for crude oil importer and exporter countries. Oil prices can shock the market and

many projects in different countries 1, 2.

CE

Finding a model for predicting different economic and engineering parameters is a very common
AC

method and this is not an exception in oil price prediction 3, 4, 5. There are various works that have

attempted to predict the oil price. The most important of them will be reviewed here. In 2007,

Chen3 compared the oil price with oil exchange rates to find a relationship between them for

*
Corresponding Author: [email protected]
 ACCEPTED MANUSCRIPT

forecasting oil prices. In 2008, Fan4 created a generalized pattern matching, based on the genetic

algorithm model. His model predicted the crude oil price based on the historical similarities of

points. In 2009, Ghaffari5 used some soft computing methods to predict the western Texas oil

price. He considered daily base data points in his study. In 2012, Jammazi6 used the wavelet

principal hybrid with neural network to predict the oil price. In 2013, Shin7 used a machine

T
learning method called semi-supervised learning (SSL) to predict the monthly oil price changes.

IP
His data belonged to the points up to 2008. In addition, in this year, Xiong8 used a multistep

CR
ahead forecasting in price prediction and claimed that his method’s computational load is low. In

2014, Yu9 combined compressed sensing based denoising (CSD) and artificial intelligence

US
methods to forecast oil prices. He claimed that his method was robust and effective. In this year,
AN
Godarzi10 used regression in the neural network to analyze the pattern of oil price changes.

As reviewed, there are different models for predicting oil prices. In most of them, the future price
M

is estimated using previous points (but how many points?). Here an artificial neural network

(ANN) is combined with the genetic algorithm (GA) to predict the future oil price. The details of
ED

this work will be explained in the next section.

PT
CE

2. Methods
AC

First of all a strong data bank is needed for building the model. These data were gained from the

WIT database. The data included the 219 points of daily oil prices from the beginning of 2014
11
until now (end of October of 2014). 95 points of monthly price begin from 2007 till now.

These data are shown in Figs. 1 and 2. As these figures show, the oil price change on a monthly

basis is much higher than a daily basis (it is natural); hence the model will have more problems
 ACCEPTED MANUSCRIPT

in monthly forecasting. Here, our goal is to use the data of these figures to forecast the oil price

in the next time series. The artificial neural network is a powerful tool for predicting the future

oil price based on past prices. Here, in ANN the output is the oil price and the input is some data

points preceding the output point. But how many points should be considered? If the number of

points is very small the model cannot predict it well because of not having sufficient data and if

T
the number of points is very large maybe some prices caused by some previous phenomena

IP
(which are no longer available) affected the prediction. Thus, here ANN is coupled with GA and

CR
thus GA optimizes the number of backward points and hidden layers. The properties of ANN and

GA are listed in Table I and Table II, respectively. In Table 1, ‘Levenberg-Marquart’ is the name

US
of a method; for more information see12. In Table II, the different listed properties are not
AN
complicated and there is no need to know them for this study. The interested reader can refer to
13
. GA optimizes the number of backwards points and hidden layers and ANN predicts the future
M

prices; whereas GA tries to reduce the error of its prediction. In this study, GA is changed, not to

just save the inputs but also to save the created network. Training is an important part in creating
ED

the model. For this purpose, the data are divided into two parts. One part is used for training the

model and another one for testing it. For daily prediction 22 points and for monthly prediction 10
PT

points are put aside for testing and the remaining are used for training. After running the model,
CE

GA concluded that 3 backward points and 14 hidden layers are suitable for accurate prediction.

The GA of this study was changed in a way to prefer the lowest hidden layer when the accuracy
AC

is similar. It is concluded that for the hidden layer of more than 14, there is no accuracy

improvement and thus 14 hidden layers are selected. Table 3 shows the statistical properties of

the model. This table shows that both daily and monthly prediction were well. Just the maximum

error of monthly oil prediction was slightly higher which was due to the small number of inputs,
 ACCEPTED MANUSCRIPT

but even in this case the amount of median relative error was small. The higher amount of mean

relative error was because of having a point with high error. The convergence of the genetic

algorithm is shown in Fig. 3. As this figure shows in both cases of daily and monthly prediction,

the number of iterations is small and thus the method is fast. From now on we call the new

method ANN-GA.

T
IP
AFTER CREATING THE MODEL IT WAS TESTED BY SOME DATA POINTS WHICH WERE NOT USED IN

CR
BUILDING THE MODEL. FIG. 4 SHOWS THE CROSS PLOT OF ESTIMATED PRICE VERSUS REAL ONE FOR

TEST AND TRAIN DATA POINTS IN BOTH MONTHLY AND DAILY PREDICTION. THIS FIGURE SHOWS

US
THAT THE DATA POINTS ARE WELL FITTED AND MEAN SQUARE ERROR IN BOTH DAILY AND

MONTHLY CASES IS WELL.

AN

3. Results
M

Here, to see the performance of the GA-ANN model, it is compared with the pure ANN model.
ED

In ANN, 10 hidden layers are proposed and two backward points will participate in forecasting
14
the new point. These values are considered as default in many references . After running the
PT

ANN, the following results are gained (Table 4). In next section the applicability of this method

in comparison with other methods will be discussed.

CE

4. Discussion
AC

Now, ANN will be compared with ANN-GA. Based on Table IV, maximum relative error of

ANN in daily prediction is slightly higher than ANN-GA, but in monthly prediction it is much

higher. This means that ANN cannot predict well when the data base is small or variables change

aggressively, but ANN-GA predicts much better. Also, the mean (average) relative error and
 ACCEPTED MANUSCRIPT

median relative error in monthly prediction by ANN is much greater than ANN-GA prediction.

The regression plot of ANN is shown in Fig. 5. Comparing this figure with Fig. 4 shows that the

points of ANN-GA have fitted much better than ANN. The test data in ANN had an r smaller

than 0.9 which has not occurred in any cases of ANN-GA. Another point of this figure is that in

training, the points have fitted well but their prediction was poor thus having a good ANN model

T
but available data does not ensure the good prediction in other points. This is while ANN-GA

IP
does not suffer from this problem.

CR
Now is the time to compare the method of this paper with the pure neural network. Fig. 6

US
compares these two with the actual price of crude oil. In this figure, the oil price is predicted

from the seventh of October until the end of this month using previous data points. As this figure
AN
shows, ANN-GA is much stronger than ANN in predicting the crude oil price. Fig. 7 compares

these for monthly prediction. This figure shows that similar to Fig. 6, ANN forecasting is poor in
M

comparison with ANN-GA.

ED

5. Conclusion and Policy Implications

PT

1. Optimizing the neural network (ANN) with the genetic algorithm (GA) greatly

improves the neural network estimation.

CE

2. When the changes in variables are small ANN works well, but as the changes
increase or the number of data points for training the model decrease, the efficiency
AC

of ANN decreases while ANN-GA works well in both situations.

3. For predicting the oil price considering that the three previous points are sufficient,
the created model will forecast well.
 ACCEPTED MANUSCRIPT

4. As it is desired to predict crude oil prices in the future, it is highly recommended to

use a larger database and the model be built by ANN-GA rather than pure ANN.

T
References

IP
1. D.C. Filis and G. Broadstock, Oil price shocks and stock market returns: New evidence from

CR
the United States and China. Journal of International Financial Markets, Institutions and

Money 33, 417–433 (2014).

US
2. W. Hamma, A. J. A. G., Effect of Oil Price Volatility on Tunisian Stock Market at Sector-
AN
level and Effectiveness of Hedging Strategy. Procedia Economics and Finance 13, 109–127

(2014).
M

3. Mahdiani, M. R. & Kooti, G., The Most Accurate Heuristic-Based Algorithms for Estimating
ED

the Oil Formation Volume Factor. Petroleum, doi:10.1016/j.petlm.2015.12.001 (2016).

4. Mahdiani, M. R. & Khamehchi, E., A New Method for Building Proxy Models Using
PT

Simulated Annealing. Middle-East Journal of Scientific Research 22 (3), 324-328 (2014).

CE

5. Mahdiani, M. R., Khamehchi, E., Soltan Mohammadi, R. & Azkayi, B., A New Proxy Model,

Based On Meta heuristic Algorithms For Estimating Gas Compressor Torque, presented at
AC

11th international industrial conference, Tehran, 2015 (unpublished).

6. S. Chen and H. Chen, Oil prices and real exchange rates. Energy Economics, 390–404

(2007).
 ACCEPTED MANUSCRIPT

7. Y. Fan, Q. Liang and Wei, A generalized pattern matching approach for multi-step prediction

of crude oil price. Energy Economics 30, 889–904 (2008).

8. Ghaffari, A. Z. a. S., A novel algorithm for prediction of crude oil price variation based on

soft computing. Energy Economics 31, 531–536 (2009).

T
9. R. Aloui and C. Jammazi, Crude oil price forecasting: Experimental evidence from wavelet

IP
decomposition and neural network modeling. Energy Economics 34, 828–841 (2012).

CR
10. H. Shin, T. Hou, K. Park, C. Park, S. Choi, Prediction of movement direction in crude oil

US
prices based on semi-supervised learning. Decision Support Systems 55, 348–358.
AN
11. Y. Tao Xiong, Y. Boa,Z. Hu, Beyond one-step-ahead forecasting: Evaluation of alternative

multi-step-ahead forecasting models for crude oil prices. Energy Economics 40, 405–415
M

(2013).
ED

12. L. Yu, Y. Zhao, L. Tang, A compressed sensing based AI learning paradigm for crude oil

price forecasting. Energy Economics 46, 236–245 (2014).

PT

13. A.A. Godarzi, R. M. A. A. T. A. a. T. J., Predicting oil price movements: A dynamic

CE

Artificial Neural Network approach. Energy Policy 68, 371–382 (2014).

14. NYMEX, Crude Oil, Available at http://www.nasdaq.com/markets/crude-

AC

oil.aspx?timeframe=7d (2014).

15. A. Abd-El-Hafiz and Adly, Utilizing neural networks in magnetic media modeling and field

computation: A review. Journal of Advanced Research 5, 615-627 (2014).

 ACCEPTED MANUSCRIPT

16. M. Mitchell , An introduction to genetic algorithms (complex adaptive systems) (A Bradford

Book, 1998).

17. Y. Zhang, Y. Y. G. X. Y. L. X., Cross-validation based weights and structure determination

of Chebyshev-polynomial neural networks for pattern classification. Pattern Recognition 47,

T
3414–3428 (2014).

IP
CR
US
AN
M
ED
PT
CE

Table 1. The artificial neural network parameters of the model of this study
AC

Parameter Value

Data division Random

Training Levenberg-Marquart

Performance Mean Square Error

 ACCEPTED MANUSCRIPT

Stop tolerance 1.00E-05

T
IP
CR
US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Table 2. The Genetic Algorithm Parameters of the Model of this Study

Property name Property value

Population type Integer vector

Population Size 20

T
Fitness Scaling Rank

IP
Selection Function Stochastic uniform

CR
Elite Count 4

Crossover Function

Crossover Fraction
US Scattered

0.8
AN
Mutation function Uniform
M

Mutation Probability 0.05

Migration Forward
ED

Max Generation 100

PT

Max Stall Generation 50

CE

Function Tolerance 1.00E-06

AC
 ACCEPTED MANUSCRIPT

Table 3. The statistic properties of the model of this paper (max, mean and median are relative errors (%)

and. R is square error)

Points Max Mean Median r

Train 194 1.58 0.58 0.54 0.99

T
Daily Test 22 4.61 0.81 0.69 0.98

IP
Train 85 17.15 3.25 1.32 0.97

CR
Monthly Test 10 4.86 1.84 1.55 0.99

US
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Table 4 The statistic properties of ANN (relative errors (%) and mean square error)

Points Max Mean Median r

Train 194 4.62 0.81 0.69 9.4

Daily Test 22 3.68 1.06 0.66 8.8

T
Train 85 22.65 5.39 5.05 9.7

IP
Monthly Test 10 13.84 6.8 6.58 0.86

CR
US
AN
M
ED
PT
CE
AC
AC
Crude Oil Price (US$/bbl)
CE

20
40
60
80

0
100
120

Jan 01, 2014

Jan 14, 2014
PT
Jan 27, 2014
Feb 07, 2014
Feb 20, 2014
Mar 05, 2014
ED
Mar 18, 2014
Mar 31, 2014
Apr 11, 2014
Apr 25, 2014
M May 08, 2014
May 21, 2014
Jun 03, 2014
Jun 16, 2014
Jun 27, 2014
Jul 10, 2014
AN
Jul 23, 2014
Aug 05, 2014
Aug 18, 2014
Aug 29, 2014
ACCEPTED MANUSCRIPT

Sep 11, 2014

US
Sep 24, 2014
Oct 07, 2014
Oct 19, 2014
Oct 29, 2014
Figure 1 Daily Crude oil price from 1 January of 2014 to 31 October of this year
CR
IP
T
AC
Crude Oil Price (US/bbl)

20
40
60
80

0
100
120
140
160
CE
Dec 01, 2006
Apr 01, 2007
Aug 01, 2007
Dec 01, 2007
PT
Apr 01, 2008
Aug 01, 2008
Dec 01, 2008
ED
Apr 01, 2009
Aug 01, 2009
M Dec 01, 2009
Apr 01, 2010
Aug 01, 2010
Dec 01, 2010
Apr 01, 2011
Aug 01, 2011
AN
Dec 01, 2011
Apr 01, 2012
Aug 01, 2012
Dec 01, 2012
ACCEPTED MANUSCRIPT

Apr 01, 2013

US
Aug 01, 2013
Dec 01, 2013
Apr 01, 2014
CR
Aug 01, 2014
Figure 2 Monthly Crude oil price from 1 December of 2006 to 1 October 2014

IP
T
 ACCEPTED MANUSCRIPT

6 Daily 7 Monthly
Oil Price (US$/bbl)

OIl Price (US$/bbl)

5 6
5
4
4
3
3
2 2

T
1 1

IP
0
0
1 3 5 7 9 11 13 15 17
1 3 5 7 9 11 13 15 17 19 21 23
Iteration
Iteration

CR
(a)

US (b)

Figure 3 Convergence of the genetic algorithm for tuninng the neural network; a:for daily b: for monthly
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Daily Train: R=0.98786 Daily test: R=0.98418

98
105
97
100

Estimated
96
Estimated

95 95

T
90 94

93
85

IP
92
85 90 95 100 105 92 94 96 98

CR
Exact Exact

(a) (b)

140
Monthly Train: R=0.96868
US Monthly test: R=0.99039
AN
90
120
Estimated
Estimated

100 80
M

80
70

60
ED

60
60 80 100 120 140 60 70 80 90
Exact Exact
PT

(c) (d)

Figure 4 Regression plot of the model of this paper

CE
AC
 ACCEPTED MANUSCRIPT

Monthly Train: R=0.94182 Monthly test: R=0.88363

140
90
120
Estimated 85

Estimated
100 80

75
80
70
60 65

T
60
60 80 100 120 140
60 70 80 90

IP
Exact Exact

(a) (b)

CR
Daily Train: R=0.96648 Daily Test: R=0.85699
98
105

100
US 97

96
Estimated
Estimated

95
95
AN
90
94
85
93
80
92
M

80 90 100 92 94 96 98
Exact Exact
ED

(c) (d)

Figure 5 Regression plot of the neural network model

PT
CE
AC
 ACCEPTED MANUSCRIPT

100

Crude Oil Price (US$/bbl)

90
80
70
60
50
40
30

T
20
10

IP
0

CR
Exact ANN-GA ANN

US
Figure 6 Daily prediction of different methods
AN
M
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

120

100
Oil Price (US$/bbl)
80

60

40

T
20

IP
0

CR
Exact ANN-GA ANN

US
Figure 7 Monthly prediction of different methods
AN
M
ED
PT
CE
AC

19