Content uploaded by Krishan K Pandey

Author content

All content in this area was uploaded by Krishan K Pandey on Jun 14, 2015

Content may be subject to copyright.

Energy Syst

DOI 10.1007/s12667-015-0151-y

ORIGINAL PAPER

Artiﬁcial intelligence methods for oil price forecasting:

a review and evaluation

Neha Sehgal1·Krishan K. Pandey2

Received: 19 September 2014 / Accepted: 8 May 2015

© Springer-Verlag Berlin Heidelberg 2015

Abstract Artiﬁcial intelligent methods are being extensively used for oil price fore-

casting as an alternate approach to conventional techniques. There has been a whole

spectrum of artiﬁcial intelligent techniques to overcome the difﬁculties of complexity

and irregularity in oil price series. The potential of AI as a design tool for oil price

forecasting has been reviewed in this study. The following price forecasting tech-

niques have been covered: (i) artiﬁcial neural network, (ii) support vector machine,

(iii) wavelet, (iv) genetic algorithm, and (v) hybrid systems. In order to investigate

the state of artiﬁcial intelligent models for oil price forecasting, thirty ﬁve research

papers (published during 2001 to 2013) had been reviewed in form of table (for ease

of comparison) based on the following parameters: (a) input variables, (b) input vari-

ables selection method, (c) data characteristics (d) forecasting accuracy and (e) model

architecture. This review reveals procedure of AI methods used in complex oil price

related studies. The review further extended above overview into discussions regard-

ing speciﬁc shortcomings that are associated with feature selection for designing input

vector, and then concluded with future insight on improving the current state-of-the-art

technology.

Keywords Neural networks ·Feature selection ·Support vector machine ·

Hybrid systems ·Oil price forecasting

BKrishan K. Pandey

krishan.pandey@gmail.com

Neha Sehgal

nmehra@jgu.edu.in

1Jindal Global Business School, O. P. Jindal Global University, Sonipat Narela Road,

Near Jagdishpur Village, Sonipat, Sonipat, Haryana 131001, NCR of Delhi, India

2College of Management and Economic Studies, University of Petroleum and Energy Studies,

Energy Acres, P.O. Bidholi, Via-Prem Nagar, Dehradun 248007, India

123

N. Sehgal, K. K. Pandey

List of symbols

R2Coefﬁcient of determination

A Annual

AC Analog complexity

ACF Auto-correlation function

ACIX Autoregressive conditional interval model with exogenous explana-

tory interval variable

AE Absolute error

AI Artiﬁcial intelligent

ALNN Adaptive linear neural network

AMIN AI framework of Amin-Naseri et al.

ANN Artiﬁcial Neural Network

APARCH Asymmetric power ARCH

AR Annualised return

ARIMA Autoregressive integrated moving average

BFGS Broyden–Fletcher–Goldfarb–Shanno–Quasi Newton

BiP Sig Bipolar sigmoid

BLR Bias learning rule

BNN Boltzmann Neural Network

BP Back-propagation

BPNN Back-Propagation Neural Network

BR Bayesian regulation

Br Brent crude oil market

BVaR Bayesian vector auto-regression

CA Correlation analysis

Ca-Var Conditionally autoregressive VaR

CC Cluster classiﬁer

CrI Crisis index

D Daily

DA Day ahead

Db Daubechies

DirS Direct strategy

DNN Decomposition based Neural Networks

DS Directional statistics

DT Delta test

Du Dubai oil market

ECM Error correction model

EGARCH Exponential GARCH

EM Expectation maximization

EMD Empirical mode decomposition

ENN Elman Neural Network

FBS Forward backward selection

FIGARCH Fractionally integrated GARCH

FIML Full information maximum likelihood

FLNN Functional Link Neural Network

123

Artiﬁcial intelligence methods for oil price forecasting...

FM Fuzzy model

FNN Fuzzy Neural Network

FP NYMEX future prices

GA Genetic Algorithm

GARCH Generalized autoregressive conditional heteroskedasticity

GB Geometric Brownian process

GD Gradient descent

GDX Gradient descent BEP

GPMGA Generalized Pattern Matching Genetic Algorithm

GRNN General Regression Neural Network

GSM Grey system model

GT Gamma test

HaT Harr a Trous

HM Hidden Markov Model

HQIC Hannan–Quinn info criterion

HR Hit rate

HTS Hyperbolic tangent sigmoid

HWBT Hull white with binomial tree

IBL Instance based learning

IGARCH Integrated GARCH

IGP Inverse Gaussian process

JC Judgemental criterion

KAB Genetic Programming framework of Kaboudan

L-RIM Linear relative inventory model

LD Log-differenced

Lgs Logistic

LM Levenberg–Marquardt Algorithm

LS Logarithmic sigmoid

LSE Least Square Error

M Monthly

MA Month ahead

MAE Mean Absolute Error

MAPE Mean absolute percentage error

MFA Manual feature extraction

MLP Multi-layered Feed Forward Neural Network

MoGNN Mixture of Gaussian NN

MRP Mean reverting process

MSE Mean Squared Error

NL-RIM Non-linear relative inventory model

NMSE Normalised Mean Squared Error

NN Neural networks

NORM Normalization

NRW Naïve random walk

NSR Noise-to-signal ratio

OLS Ordinary Least Square

OU Ornstein–Uhlenbeck Model

123

N. Sehgal, K. K. Pandey

PACF Partial autocorrelation function

PARCH Power ARCH

PCP Percentage of correct predictions

PGRP Persian Gulf region prices

PMI Partial mutual information

PR Prediction rate

PRMS Pattern modelling in recognition system approach

RBF Radial basis function

RecS Recursive strategy

RM Regression model

RMA Relative change of moving average

RMS Regime Markov switching stochastic volatility model

RMSE Root Mean Squared Error

RNN Recurrent Neural Network

RS Regime switching

RT Return transformation

RW Random walk

S-SVM Standard SVM

SA Step ahead

Sig Sigmoid

SM Stochastic model

SMAPE Symmetric MAPE

SMP Smoothing procedure

SNR Signal-to-noise ratio

SoMLP Self-organizing MLP

SP Spot prices

SR Scaling range

SSE Sum of Square Error

STEO EIA’s short-term energy outlook econometric model

SVM Support vector machine

SVR Support vector regression

TE Trial and error method

TGARCH Threshold GARCH

TM Text mining

TPA Time period ahead

TSig Tangent sigmoid

TSK Takagi–Sugano–Kang

VaR Value-at-risk model

VECM Vector error correction model

W Weekly

WA Week ahead

WANG AI framework of Wang et al.

WCI Without crisis index

WDE Wavelet decomposition ensemble

WNN Wavelet Neural Network

WSP Without smoothing procedure

123

Artiﬁcial intelligence methods for oil price forecasting...

WT Wavelet transform

WTI West Texas Intermediate Crude Oil Market

1 Introduction

Fossil fuels currently account for 87 % of primary energy demand and projected to

still make up 82 % of the global total by 2035. According to BP [1], oil remains the

world’s primary fuel, accounting to 33.1 % of global energy consumptions. Oil will

remain the energy type with the largest share for most of the projected periods and

will continue to play a foremost part in satisfying world energy needs [2]. Oil price

act as a key component dominating investment picture for years to come on. They act

as a key variable in evaluation of economic development, energy policy decisions and

stock markets [1]. A prior knowledge of oil prices ﬂuctuations helps oil producers to

make decisions about the increase or decrease in production levels accordingly.

Oil prices helps strategically in macroeconomic projections and macroeconomic

risk analysis for central and private banks. They are helpful in predicting recession

in business cycles [3]. They are helpful in planning regulatory policies regarding

taxes & standards. Businesses dependent on oil will be beneﬁted as they will be in

position to take measures to control manufacturing and sales of their products in

line with expected trend of forecast oil prices. Accurate forecasting helps Non-OPEC

countries to take effective measures so their growth remains robust and thus beneﬁted

consumers. Further, economic policies can be formulated in way to overcome recession

and unemployment.

2 Econometric models

Forecasting of crude oil prices is an important task for better investment manage-

ment, macroeconomic policies and risk management. It is important to analyses the

probabilistic assumption of oil prices in terms of normality, linearity and serial cor-

relation [4]. To forecast crude oil prices, a variety of approaches have been proposed

by numerous authors employing time series [5–10], ﬁnancial models [11,12] and

structural models [13–18].

2.1 Time series model

Time series analysis is a method of forecasting that focuses on the historical behav-

iour of dependant variable. Oil prices are assumed to be normally distributed in many

studies but their departure from normal distribution was disregardeddue to misinterpre-

tation of Central Limit Theorem [4,19]. Crude oil prices are found to be non Gaussian.

Forecasting crude oil prices through fundamental method is a complex task due to

uncertainty, noisiness and non-stationary inbuilt in indicators that drive them. There-

fore, time series models provide an alternative to analyse and predict future movements

based on past behaviour of oil prices [20]. The price-forecasting models based on time-

series approach have been further classiﬁed into three subsets as shown in Fig. 1.

123

N. Sehgal, K. K. Pandey

Fig. 1 Classiﬁcation of time series oil price-forecasting models

The summary of time series forecasting models based on type of model being

employed and methods used by researchers have been presented in Table 1. Stochastic

models are inspired by ﬁnancial literature and are widely applied in forecasting of oil

prices. There are several stochastic models which have been employed for modelling

and forecasting of crude oil prices such as Random Walk [6,7,9,21,22], Mean Revert-

ing Processes [23], Brownian Motion Processes [24], Ornstein–Uhlenbeck Processes

[24], Inverse Gaussian Process [19] and Jump Diffusion Processes [25].

Regression type models are based on the relationship between oil price and

number of exogenous variables that are known or can be estimated. The most com-

mon approaches employing regression models are ARIMA models [12,26,27] and

GARCH-family models [6,8,26,28–30]. Arouri et al. [5] employs GARCH models

to forecast conditional volatility of spot and future oil prices with structural breaks

for better forecasting performance. Huang et al. [31] and Hou [28] presented superior

performance of non-parametric GARCH models relative to parametric GARCH mod-

els (in-sample and out-of-sample volatility forecasts). Researchers concluded that

non-linear dynamical approach is more appropriate for characterizing and predict-

ing crude oil prices than linear approach [32,33]. The parameters of forecasting

models for crude oil prices have been estimated by either Least Square Method

[12,23,34–38], Full Information Likelihood Method [16], Kalman Filter [23,24]or

under Bayesian Framework [39]. However, these numerous estimation algorithms have

failed to achieve high prediction accuracy. Stochastic models involving certain char-

acteristics of oil prices and regression models have been kept outside the scope of this

review.

A review of these econometric time series models for oil price forecasting has been

presented by Frey et al. [40].

Table 1provide summary of time series models for crude oil price forecasting.

In recent times, artiﬁcial intelligent models are extensively being used to capture

unknown or too complex structure in the time series. Researchers have used artiﬁ-

cial intelligent model based approach for oil price forecasting in more than 50 % of

the studies listed in Table 1. Out of thirty six studies listed in Table 1, twenty eight

studies have considered WTI crude oil spot prices as dependent variable in their stud-

ies. Section 3covers various artiﬁcial intelligent models ranging from single models

(e.g. neural networks, support vector regression, wavelets) to more complex hybrid

versions.

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 1 Summary of time series models for crude oil price forecasting

References Oil market Model type Methods

[9] WTI SM, RM RW, VaR, ECM

[19]WTI SM IGP

[81]WTI RM ARIMA

[23] WTI RM, SM OLS , MRP

[77]Br AI WNN

[24]FP SM GB,OU

[84] WTI, Br RM GARCH

[26] FP RM, AI ARIMA, GARCH, ANN

[85] FP AI GSM

[10] WTI AI FM

[59] WTI, Br AI SVM

[67] WTI AI SVM

[80] WTI, Br AI ANN

[74] WTI AI WNN

[7] WTI,Br AI,RM,SM WDE,RW,ARMA

[76] WTI AI ANN, WDE

[5] WTI, FP RM GARCH

[86] FP RM RS-EGARCH

[73] WTI, Br AI GA + ANN

[79] WTI AI FNN

[87]WTI AI RMS

[28] WTI RM GARCH

[88] FP RM GARCH

[89] WTI AI WDE

[30] WTI, Br RM GARCH

[56] WTI RM RS-GARCH

[22] WTI SM, RM VECM, NRW, ARIMA

[90] WTI RM CA-VaR

[69] FP AI GA + ANN

[61] WTI AI SVR

[63] WTI AI SVR

[75] WTI, Br, Du AI WNN

[78] WTI, Br AI FNN

[29] WTI, Br, Du RM CGARCH, FIGARCH, IGARCH

[6] WTI SM, RM RW, HM, ARIMA, GARCH,

EGARCH, TGARCH, PARCH,

CGARCH

[8] SP RM GARCH, EGARCH,APARCH,

FIGARCH

123

N. Sehgal, K. K. Pandey

2.2 Fundamental models

Fundamental models predict oil prices based on their relationship with economic,

ﬁnancial, social and political indicators that drive them. This study assumes ﬁnancial

and structural models as part of fundamental models. Crude oil prices have been

inﬂuenced by large number of factors which are complex, noisy, and uncertain [20].

There is no single indicator which can provide a comprehensive portrait of how prices

can be determined. Each indicator can give us a snapshot of present condition and

modelling of these signiﬁcant snapshots together provides a clear picture of direction

of oil prices. Similar to time series models, fundamental models can also be classiﬁed

into two major classes: Regression models and artiﬁcial intelligent based models

as shown in Table 2. This table enlists the analytical methods used by authors for

forecasting crude oil prices.

Globalization hypothesis holds that oil prices (WTI-Brent, WTI-Dubai, WTI-Maya

and Dubai-Maya) move together and exhibit greater conditional dependency [41];

therefore, most of the study listed in Table 2considers WTI spot crude oil prices as

benchmark price. It is evident from Table 2that around 50 % of the studies have incor-

porated artiﬁcial intelligent models for forecasting oil prices. As evident from Table 2,

researchers have preferred ordinary least square methods for parameter estimation in

a regression model. The different input variables, along with the class they belong to,

used by different researchers are presented in the next section.

2.2.1 Factors driving oil prices

Oil prices had shown upward trend in 1996 but prices declined drastically by end

of 1998. As a consequence of cuts in OPEC production targets, oil prices increased

again in late 2000. The impact of such extreme events is of prime importance as they

effect the direction of oil prices and thereby the objective of increasing the predication

accuracy of crude oil prices. It is important to identify the key indicators driving crude

oil prices (during the time-frame of happening of such extreme events) for designing

better structural forecasting models. In 2001, 9/11 attack led to increase in volatility of

oil prices that soared oil prices till 2004. Oil prices have been steadily rising for several

years and in July 2008 stood at a record high of $145/bbl due to low spare capacity.

Later, it declined due to global economic crisis at the end of 2008 and then recovered

to around $75/bbl by 2010. In 2013, oil prices have set record by surpassing $100/bbl

for the ﬁrst time in that year (in money-of-the-day terms). According to EIA [42], oil

prices were rising sharply because global demand (especially in China) was surging

and production wasn’t adequate to keep up. This led to a boom in unconventional

oil production by major energy companies across globe and world’s oil supply kept

growing by mid-2014. On the other side, due to slowdown in Asia and Europe, the

oil demand began weakening. This combination of steadily rising supply and weaker-

than-expected demand together with OPEC in favour of letting prices continue to fall

pushed oil prices to drop around $55/bbl by end of 2014. This rise or decline in oil prices

stimulates for studying in detail the factors behind movements in oil prices. There are

large number of factors, which are complex, noisy, and uncertain inﬂuencing crude oil

prices [20]. Understanding complex oil price movements and indicators driving them

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 2 Summary of

fundamental models for crude

oil price forecasting

References Oil market Model type Methods

[57] WTI AI ANN

[43] WTI AI WNN

[35]WTI RM OLS

[36]WTI RM OLS

[37]WTI RM OLS

[38] WTI RM OLS, ECM

[46]WTI RM OLS

[17] WTI AI WNN

[34]WTI RM OLS

[16] WTI RM OLS, FIML, ECM

[12] WTI RM OLS, ARMA

[66]Br AI GRNN

[18] WTI AI ANN, TM

[64] WTI AI FNN

[14] PGRP AI ANN

[50] WTI RM VECM

[51]WTI RM ACIX

[55] WTI AI WDE

[52]FP AI WDE

[53] FP RM GARCH

[91] FP RM GARCH

[33] Br, WTI, Du RM ESTAR

[92] WTI RM CGARCH

[39] WTI RM BVaR

[93]FP RM ECM

[65] SP AI ANN

[60] WTI AI SVR

[62] NF AI SVR

[68]SP AI GA

[74]WTI;BrAI WNN

[72] WTI AI WNN

was the impetus for Energy Information Administration (EIA) to launch a monthly

report to assess the ﬁnancial, trading and physical market factors that inﬂuence oil

prices.

Beside geopolitical and economic events, any ﬂuctuation in demand or supply side

also creates imbalance in the market that critically impact oil market participants and

makes markets unpredictable. The demand-supply framework plays a crucial role to

an extent that it determines the directions of crude oil prices but has not been sole

indicator that drives oil prices. Researchers have considered enormous factors such

as GDP, inventories, emerging economies and stock market ﬂuctuations to study their

inﬂuence on oil prices (Table 3covers this in detail). There is no solitary indicator

123

N. Sehgal, K. K. Pandey

Tab le 3 Summary of factors

inﬂuencing oil prices Class Input variable Key

Supply OPEC production V1

Non-OPEC production V2

World production V3

Oil supply V4

Demand OECD consumption V5

China consumption V6

India consumption V7

Seasonal demand V8

Global demand V9

Non-OECD consumption V10

Inventories U.S. reﬁnery capacity V11

OPEC total liquid capacity V12

U.S. gasoline ending stocks V13

OECD stocks V14

U.S. ending stocks V15

U.S. petroleum imports from OPEC V16

U.S. petroleum imports from non-OPEC V17

U.S. crude oil imports from OPEC V18

U.S. crude oil import from non-OPEC V19

OECD industrial inventory level V20

Crude oil distillation capacity V21

Price Historical prices V22

Heating oil spot price V23

Gasoline oil spot price V24

Natural gas spot price V25

Propane spot price V26

NYMEX crude oil futures V27

NYMEX heating oil futures V28

Reserves OPEC reserves V29

OECD reserves V30

No. of well drilled V31

Economy GDP growth rate V32

U.S. dollar nominal effective exchange rate V33

Foreign exchange of GBP/USD V34

Foreign exchange of YEN/USD V35

Foreign exchange of Euro/USD V36

U.S. inﬂation rate V37

U.S. consumer price index V38

Population of developed countries V39

Population of less developed countries V40

S&P500 V41

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 3 continued Class Input variable Key

Gold prices V42

Producer price index V43

World events World event impact factor V44

OPEC quota tighten (April 99) V45

Sept 11, 2001 attack V46

Properties API density V47

Sulphur content V48

Country V49

Time V50

No. of weeks V51

(lags, future prices or macroeconomic variables) that can provide a complete picture

on how prices have been determined, however there are few key indicators that have

governed and ruled crude oil prices. The key indicators can give us snapshots of

ﬂuctuations in oil prices and modelling of these signiﬁcant snapshots only (as input

variables) can give a clear picture on directions of oil prices. The ﬂuctuations in these

factors cause complex, volatile, non-linear and chaotic tendency of crude oil prices,

therefore, it is important to ﬁnd the key strategic indicators that are ruling crude oil

prices from decades. Thus, it has become crucial to develop predictive models using

various inﬂuential factors that drive crude oil prices to understand the complex and

dynamic nature of oil prices.

This review has identiﬁed petroleum inventory level as a virtuous market indicator of

change in crude oil price. Inventory levels have been a measure of balance or imbalance

between production and demand [43]. Ye et al. proposed a linear forecasting model

using relative inventory level as input for oil prices [36] but later improved Linear-

RIL model to non-linear-RIL model due to dynamic relationship between them [37].

Pang et al. [43] proposed to consider both crude oil inventory level and petroleum

product inventory level as input factors for better forecasting performance. Weiqi

et al. [15] constructed a structural econometric model using relative inventory and

OPEC production as explanatory variables for short-run oil price forecasts. Other

than inventories level, variables like production, net imports and forward prices were

taken as independent variables to estimate spot prices by Considine and Heo [44]. The

evidence for the non-linear relationship between GDP growth and oil prices has been

examined by Hamilton [34] and Kim [45].

Zamani [38] examined a short term quarterly econometric forecasting model using

OECD stocks, Non-OECD demand and OPEC supply to forecast WTI crude oil prices.

Ye et al. [35] considered the possible substitution for large proportion of world demand

and inventory (in form of OECD demand) as input variable compared to U.S demand

alone. Dées et al. [46] assessed a structural econometric model to show the immediate

impact of OPEC quota decisions and capacity utilization on oil prices. According

to BP [1], emerging economies (especially Asia) accounted for major net growth

in energy consumption whereas OECD demand remains falling for a third time in

last 4 years. Ratti and Vespignani [47] has highlighted that intense increase in China

123

N. Sehgal, K. K. Pandey

and India liquidity led to increase in real oil price and production. Li and Lin [48]

has provided evidence indicating demand from China and India as leading driver in

the world oil pricing system since 2003. Zhang and Wang [49] indicated a greater

contribution (95.71 %) of crude oil future markets in price discovery function of

spot price. Alvarez-Ramirez et al. [13] used fourier analysis to examine the strong

relationship between U.S macro economy and crude oil prices. Dées et al. [16] has

examined the dynamic relationship between oil prices and OPEC capacity utilization.

Murat and Tokat [50] examined the forecasting power of crack spread futures under

vector error correction framework to predict spot oil markets. Yang et al. [51] studied

the inﬂuence of European debt crisis and ﬁnancial crisis on crude oil prices with the

proposed Autoregressive Conditional Interval Model with Exogenous Explanatory

Interval Variable (ACIX). He et al. [52] used error correction models to examine the

inﬂuence of Kilian economic index (as global activity indicator) on crude oil prices.

Bu [53] examined the relationship between speculative traders’ position and crude

oil future prices using estimates of GARCH model. Basher and Haug [54] proposed a

structural vector autocorrelation model to investigate the dynamic relationship between

emerging markets, stock prices and oil prices. Chai et al. [39] developed a oil price VaR

model based on path-analysis using core inﬂuential factors. Zhang et al. [55] proposed

an empirical mode decomposition-based event analysis method to estimate the impact

of extreme events on crude oil prices. Fong and See [56] suggested regime switching

models framework to study factors driving crude oil prices volatility. Ignoring the clear

evidence of presence of non-linearity or structural breaks in the oil price series can lead

to false imprint on predictability and persistence. Oil prices are shown more sensitive

to any change in oil supply by Chai et al. [39] but weekly dependent on exchange rates

by Reboredo [41]. The study of factors driving oil prices that are considered in various

stochastic or regression models of oil prices forecast is a major area of research and

has been kept outside the scope of this review.

Despite above mentioned attempts, oil price prediction has remained a difﬁcult

problem due to its complex non-linear and time-varying nature. In addition, recent stud-

ies lay emphasis on developing structural econometric models for forecasting crude

oil prices without focusing on ﬁnding the key drivers of oil prices. Most recently, a cat-

egory of artiﬁcial intelligent models have emerged and are being attempted to predict

oil prices. AI based framework for oil price prediction are discussed in detail in Sect. 3.

The factors inﬂuencing oil prices may be classiﬁed within the categories: C1—supply,

C2—demand, C3—inventories, C4—price, C5—reserves, C6—economy, C7—world

events, and C8—properties.

There are as many as ﬁfty one variables used in different AI related studies for

crude oil price forecasting. A detailed list of factors driving oil prices considered in

artiﬁcial intelligent models, along with the class to which they belong are presented

in Table 3. In order to build an effective model, careful attention should be paid

on selecting informative and inﬂuential inputs which cause changes in prices [57].

However, until recently, the input variables of oil price forecast have been selected

on judgemental criteria or trial and error procedures [58–63]. This review discovers

that most of the studies were concentrated on non-linearity, non-stationary and time

varying properties of oil prices but seldom focused on feature selection method for

selecting signiﬁcant inputs to improve forecasting accuracy. The study identiﬁes that

123

Artiﬁcial intelligence methods for oil price forecasting...

historical oil prices (either daily, weekly or yearly) are the most popular input variables

used by researchers. Abdullah [18] has used 22 input variables from the categories

as mentioned in Table 3to achieve high prediction accuracy but the variables were

selected on judgemental criteria. To handle optimal long-term oil price forecasting,

Azadeh et al. [64] developed a ﬂexible algorithm based on artiﬁcial neural networks

and fuzzy regression by using oil supply, crude oil distillation capacity, oil consumption

of Non-OECD, USA reﬁnery capacity and surplus capacity as economic indicators.

The study concluded ANN models outperform FR models in terms of mean absolute

percentage error (MAPE). Section 3discusses in detail the characteristics of different

types of artiﬁcial intelligent models that used factors listed in Table 3as input variables

for oil price forecasting.

3 Review of methodology

In this section, the studies are segregated on the basis of type of models, starting

from single models (such as neural networks, support vector regression, wavelets,

genetically evolved models) to more complex and hybrid models. Recently, hybrid

models are been extensively used for building oil price forecasting models as they

overcome the limitations of single models and provide better forecasting accuracy.

3.1 Neural network based models

In the past , neural networks were being used extensively for oil price forecasting.

Neural Networks can model richer dynamics and can approximate any continuous

function of inputs [57]. In this category, seven researchers have forecasted oil prices

using artiﬁcial neural networks as single model. There are many research studies that

suggest integration of neural networks with other traditional methods (such as support

vector regression, genetic algorithm, or wavelets) by mean of hybrid approach for

improving the prediction performance. These studies are discussed in detail in Sect.

3.5. In Table 4, information regarding the data, time scale, input variables for the

study, method of input variable selection and preprocessing techniques employed

are discussed. Researcher have utilized neural networks for all major oil markets. It

is evident from the Table 4that neural networks can handle large number of input

variables.

Abdullah and Zeng [18] integrated 22 quantitative input variables (sub-factors of

demand, supply, economy, inventory and population) together with the qualitative data

(collected from experts’ view and news) using neural networks to predict oil prices for

long and short term time period. The authors have utilized manual feature extraction

method for ﬁnding signiﬁcant input variables for the study. Most of the studies have

pre-processed the raw price data either by scaling range, normalization or cluster

classiﬁcation as shown in Table 4.

This review identiﬁed that most of studies have selected input variables based on

judgemental criteria or trail and error basis. In Table 5, forecasting performance of

various neural networks models has been compared.

123

N. Sehgal, K. K. Pandey

Tab le 4 Data characteristics, preprocessing technique, input variables and its selection method

References Oil

market

Time

scale

Input variable Variable

selection

Preprocessing

technique

[57] WTI D V22, V27 JC RMA, NORM

[26]NFDV22 JC–

[66] Br M V11, V32, V33, V12, V1, V13 TE CrI is formed

[14] PGRP M V47, V48, V49, V50 SA SR

[58]SP–V22 JCSR

[18] WTI M V1, V2, V5, V6, V7, V14, V15,

V16, V17, V18, V19, V29,

V30, V31, V34, V35, V36,

V32, V37, V38, V39, V40

MFA LD

[65] SP D V22, V23, V24, V25, V26 CA CC

Tab le 5 Forecasting performance comparison of neural network models

References Training

data (%)

Testing

data (%)

Forecast

horizon

Comparison with

other models

Level of accuracy

[57] 90 10 3 DA – RMSE: 0.53–0.78 ;

HR: 53–79

[26] 86 14 1 TPA ARIMA, GARCH MSE: 8.14; RMSE:

2.85; MAE: 2.04

[66]86141MAWCI-GRNN MSE:−1.48–9.84

[14] 70 30 – – MSE: 7.24–8.82

[58]––5DA– NMSE:−0.35; DS:

61; SNR: 25.37;

AR: 92

[18]8020– TEI@I,

EMD-FNN-ALNN

RMSE: 2.26;; NMSE:

0.009; DS: 94

[65]80201MARM MSE:−2.15–4.73

Neural networks showed superior results compared to benchmark TEI@I method-

ology and ARIMA-GARCH models as shown in Table 5. The models are validated

using Root Mean Square Error (RMSE), Hit Rate (HR), Mean Square Error (MSE),

Mean Absolute Error (MAE), Normalized Mean Square Error (NMSE), Annualized

Return (AR) and Directional Statistics (DS). Malliaris and Malliaris [65] studied ﬁve

inter-related energy products for forecasting one-month ahead prices using neural net-

works. The results thus obtained through neural network consistently led to a MSE

less than half than that of the regression predictions. Malliaris and Malliaris [65]used

correlation analysis to ﬁnd signiﬁcant input variables for their study. The model archi-

tecture of seven studies considered under this category is shown in Table 6. Mahdi

et al. [58] examined three different neural network models: Multi-layered Perceptron

(MLP), Functional Link Neural Network (FLNN) and Self-Organized MLP (SoMLP)

for the oil price series. Mahdi et al. [58] compared the prediction capability of SoMLP

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 6 Neural Networks model’s architecture

References NN type Learning

algorithm

Hidden

neurons

Activation

function

[57] MLP, RNN LM 1–10 Sig

[26] MLP GD 5 HTS, Id

[66]GRNN – TE–

[14] MLP BP 15 LSig

[58] MLP, FLNN, SoMLP BP, IA – –

[18] MLP BP TE Sig

[65]– – 20–

with MLP and FLNN for ten different data sets including oil prices. The experimental

results thus demonstrated that all neural networks performed better by using stationary

data and failed to generate proﬁts while using non-stationary data.

Haider et al. [57] presented a short term forecasting model to understand oil price

dynamics based on multi-layer feed forward neural network. Several transformation

methods were tested with original data and results showed that relative change of sim-

ple moving average is the best method amongst other methods. Moshiri and Foroutan

[26] examined chaos in daily crude oil future prices using BDS and Lyapunov test.

The results indicated that future prices follow complex non-linear dynamic process

and showed superiority of ANN model as compared to ARIMA and GARCH models.

Movagharnejad et al. [14] designed a neural network to predict the prices of seven

different crude oils in Persian Gulf region, provided that the benchmark light oil of

Saudi Arabia is known or predicted by another hybrid forecasting method.

Further, this review identiﬁes that number of hidden neuron varies across studies.

Haider et al. [57] has ﬁxed the number of neurons in hidden layer ranging from 1–10

while few authors [14,26,65] have ﬁxed a constant value based on judgemental criteria

for number of neurons in hidden layer. There is no rule of thumb applied by researchers

for ﬁnding the optimal number of neuron to be set for hidden layer. Alizadeh and

Maﬁnezhad [66] proposed a general regression neural network forecasting model for

Brent crude oil price with particular attention on ﬁnding number of features as input

data to achieve best performance.

Most of the studies have used sigmoid function as preferable activation function

as shown in Table 6. It is evident from the Table 6that the multi-layered perceptron

neural network with back propagation as the learning algorithm is the most popular

among researchers for price forecasting.

3.2 Support vector regression models

Oil prices are complex series with mixture of linear and non-linear characteristics

underlying data generating processes of different nature. He et al. [63] introduced

morphological component analysis to explore the complex nature underlying oil prices.

There are few authors who have tested for non-linearity [61,63,67] and normality

123

N. Sehgal, K. K. Pandey

Tab le 7 Data characteristics, preprocessing technique, input variables and its selection method

References Oil

market

Time

scale

Input

variable

Input

selection

Preprocessing

[59]WTI;BrWV22 JCSR

[60] WTI W V22, V27, V8, V51, V44, V9 JC SR

[61]WTIDV22 JCRT

[62] NF D V22, V33, V28, V41 JC CC

[63] WTI D V22 JC WT; RT

[67]WTIMV22 JCRT

assumptions [61,63] of oil price series. Support vector regression has an advantage

of reducing the problem of over-ﬁtting or local minima. Khashman [60] experimental

results proved SVM could be used with a high degree of precision in predicting oil

prices. Bao et al. [59] presented a comparative study of recursive and direct strategies of

multi-step ahead prediction for both WTI and Brent crude oil spot prices with support

vector regression. As compared to results obtained through benchmark ARMA and

Random Walk models for crude oil price prediction, He et al. [61] conﬁrmed the

superiority of the proposed slantlet denoising algorithm based on SVR model. Zhu

[62] formulated a two-stage structure for modelling oil future prices by partitioning

the whole input data space into mutually exclusive regions by K-mean clustering

algorithm and then corresponding SVM models. Table 7showed that each study have

preprocessed raw price data either by scaling range, return transformation or by cluster

classiﬁer.

Most of the authors have used WTI as benchmark oil price data in their studies.

An important research gap in selection of input variables through judgemental criteria

or by literature review is highlighted under this category. It is evident from Table 8

that most of the authors have compared SVM with linear models. Xie et al. [67] has

shown SVM model performed better than back-propagation neural networks. The

model proposed by Zhu [62] has shown better performance in terms of MSE, MAE

and MAPE compared to standard SVM model. Radial basis kernel function is the most

popular choice among researchers for a price forecast problem as seen from Table 9.

The values for epsilon, cost and gamma vary across different studies. He et al. used

gradient search method to set appropriate model parameters [61,63]. Xie et al. [67]

and He et al. [63] have used directional statistics as performance criterion to compare

their respective models with traditional stochastic or regression models.

3.3 Genetically evolved models

Kaboudan [68] performed short term monthly forecasting of crude oil price using

genetic programming (GP) and neural networks. The study presented that GP can

produce impressive one-month ahead forecast compared to that by Random Walk and

ANN. This GP based oil price forecasting framework by Kaboudan [68] is considered

as benchmark for comparison by Amin-Naseri [69].

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 8 Forecasting performance comparison of support vector regression models

References Training

data (%)

Testing

data (%)

Forecast

horizon

Comparison with

other models

Level of accuracy

[59] 71 29 4, 8, 12 WA DirS, RecS RMSE: 5.05–52.11;

MAPE: 10.23

[60] 50 50 1 WA – PR: 81.27 %

[61] 60 40 – ARMA, RW MSE: 4.65

[62] 95 5 – S-SVM MSE: 1.37; MAE:

0.95; MAPE: 1.15

[63] 60 40 – RW, ARMA MSE: 8.74; DS:

53.07 %

[67] 88 12 1 MA ARIMA, BPNN RMSE: 2.19; DS:

70.83 %

Tab le 9 Support vector regression model architecture

References Kernel

function

Epsilon Cost Gamma Model

parameters

[59]RB0.01 – – –

[60] RB – 2965820 0.001953 2−15 to 215

[61]RB6.64 ×10−16 24.25 27.86 GSM

[62]RB– – – [0,1]

[63]RB7.81 ×10−30.5−83.90 ×10−3GSM

[67]RB– – – –

Tab le 1 0 Data characteristics, preprocessing technique, input variables & its selection method

References Oil

market

Time

scale

Input variable Input

selection

Preprocessing

[68] SP M V22, V3, V5, V15 TE –

[70]WTI;BrD– TE–

Xiao et al. [70] combined transfer learning techniques with analog complexing and

genetic algorithm for crude oil price forecasting. However, there is no information as

to how the input variables are selected in studies mentioned in Table 10. According to

Table 11,Xiaoetal.[70] showed that genetically evolved models performed better in

comparison to neural networks and ARIMA family models based on MSE,

RMSE and directional statistics. There is lack of information with respect to pre-

processing technique and ﬁtness function used by authors as shown in Table 12.The

new combined models based on genetic algorithm with neural networks and that with

SVM are discussed in Sect. 3.5.

123

N. Sehgal, K. K. Pandey

Tab le 1 1 Forecasting

performance comparison of

genetically evolved models

References Forecast

horizon

Comparison with

other models

Level of accuracy

[68] 1–12 MA ANN; RW MSE: 0.24–1.85;

[70] 1 TPA ARIMA; ANN;

GPMGA; AC

RMSE: 1.0691;

DS: 79.02

Tab le 1 2 Genetically evolved

model’s architecture References Cross-over

probability

Mutation

probability

Fitness

function

[68] 0.02 0.06 –

[70] 0.9 0.05 –

3.4 Wavelet-based models

A wavelet-based prediction model is proposed to provide forecast over 1–4 months’

horizon and to compare with future oil price data by Yousuf [71]. He et al. [7]intro-

duced the wavelet decomposed ensemble model to analyse dynamic changing nature

of underlying oil market structure. This study found that hybrid version comprising of

wavelet with neural networks is more appealing to researchers as compared to single

wavelet based model.

3.5 Hybrid models

3.5.1 Genetic Algorithm and Neural Network

Amin-Naseri [69] proposed a hybrid artiﬁcial intelligence model combining local

approximation techniques with genetically evolved neural network. The author used

Hannan-Quinn info criterion (HQIC) as ﬁtness function and set number of hidden

neurons in range from 1–30. The performance of the model was evaluated with three

competing frameworks (STEO, KAB and WANG) for oil price forecasting. The pro-

posed model has performed well in terms on MSE, RMSE and directional statistics,

and has found to be effectively mapping the non-linearity and non-normality present

in crude oil price data. The proposed model has been considered as benchmark model

for comparison by Alexandridis and Livanis [72]. Fan et al. [73] presented General-

ized Pattern Matching based on Genetic Algorithm (GPMGA) to predict future prices.

GPMGA overcomes some limitations of Elman Networks and Pattern Modelling in

Recognition System (PRMS) approach for multi-step prediction of oil prices.

As evident from Table 13, authors have preprocessed raw price data to achieve high

level of accuracy. Authors have preferred autocorrelation function and partial auto-

correlation function to determine the optimal number of lags for model identiﬁcation

and estimation.

The review states that genetically evolved neural networks are superior to other

competitive models as mentioned in Table 14. The studies under this category have

utilize sigmoid function as activation function. The number of neurons in hidden layers

varies as seen from Table 15.

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 1 3 Data characteristics, preprocessing technique, input variables & its selection method

References Oil market Time

scale

Input

variable

Input

selection

Preprocessing

[69]FP MV22PACFCC

[73] WTI; Br D V22 GT; ACF Standardization

Tab le 1 4 Forecasting performance comparison of Genetic Algorithm and Neural Network models

References Training

data (%)

Testing

data (%)

Forecast

horizon

Comparison with

other models

Level of accuracy

[69] 95 5 – STEO; KAB; WANG MSE: 0.90–9.10;

RMSE: 0.95–3.02;

DS: 71–81 %

[73] 99 1 1 MA PRMS; ENN RMSE: 1.57–2.43

Tab le 1 5 Genetic Algorithm and Neural Networks model’s architecture

References Model

type

Learning

Algorithm

Hidden

neurons

Activation

function

Fitness

function

Cross-over

probability

Mutation

probability

[69] MLP LM; GD 1–30 LSig HQIC 0.9 0.01

[73] RNN BP TE TSig – 0.9 0.09

3.5.2 Wavelet and Neural Network

Mingming et al. [74] proposed a multiple wavelet recurrent neural network based

hybrid method for international crude oil prices. This model utilized wavelet analysis

to capture multi scale data characteristics, while designing an appropriate recurrent

neural network to predict oil prices at different time scales, followed by a standard

back-propagation neural network to combine these independent forecasts. He et al.

[75] proposed an ensemble approach incorporating wavelet and feed-forward neural

network for estimating VaR in crude oil market to further improve modelling accu-

racy and reliability of three oil markets: WTI, Brent and Dubai. Jinliang et al. [17]

decomposed crude oil price time series into several trend and random component. For

higher prediction accuracy, the trend component of oil prices is predicted with Boltz-

mann neural network and the random component is predicted with Gaussian kernel

density function as shown in Table 16. Jammazi and Aloui [76] examined a short

term forecasting of monthly WTI prices with different input-hidden nodes combina-

tions and three types of activation functions. The results highlighted combination of

Harr A Trous wavelet function with back propagation neural network as a promising

forecasting tool. Pang et al. [43] proposed to predict monthly oil prices using OECD

inventory level as independent variable, and used wavelet theory based feed forward

neural network to model the non-linear relationship between oil prices and inventory.

123

N. Sehgal, K. K. Pandey

Tab le 1 6 Wavelet and Neural

networks model’s architecture References Model type Learning

algorithm

Hidden

neuron

Activation

function

[74] RNN; MLP BP TE LSig

[75] MLP LM 6 LSig

[17]BNN– ––

[76] MLP BP TE BiP Sig

[43]MLPGD 8–

[77] RBF – 4 –

[72]WNN– 1–

Tab le 1 7 Data characteristics, preprocessing technique, input variables and its selection method

References Oil

market

Time

scale

Input variable Input

selection

Preprocessing Wavelet function

[74] WTI, Br A V22, V42 JC WT Db

[75] WTI, Br, Du W V22 JC RT, WT HaT; Db; Coiﬂet

[17] WTI M V22, V42 JC WT Db

[76] WTI M V22 JC WT HaT

[43] WTI M V22, V20,

V45, V46

JC WT Morlet

[77] Br M V22 JC WT, SR Db

[72] WTI M V22, V43, V3 CA WT –

The proposed model achieved lower RMSE, MAPE and MAE in comparison to both

linear and non-linear relative inventory models.

Qunli et al. [77] decomposed the original price sequence successfully using discrete

wavelet transform as input layer of radial basis function neural network. Alexandridis

and Livanis [72] used wavelet neural network to forecast monthly WTI crude oil spot

prices using price lags, world crude oil production and the producer price index for

petroleum as explanatory variables.

Out of seven articles listed in Tables 17,18 only one author has emphasized on

selecting input variables based on correlation analysis. Correlation analysis is a mea-

sure of linear relationship between variables but macroeconomic variables exhibit

non-linear relationship with oil prices. Therefore, there is a requirement to develop

a method for identifying signiﬁcant input indicators based on non-linear relationship

that exists between variables. It can be observed that Daubechies has been adopted by

most researchers as a wavelet function. The prediction accuracy of combined models

is evaluated with linear ARMA family, non-linear neural networks, STEO, WANG

and AMIN models.

3.5.3 Fuzzy Neural Network

Panella et al. [78] favoured the quality of forecasting accuracy based on neurofuzzy

approach (adaptive neuro-fuzzy inference system) in comparison to other linear and

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 1 8 Forecasting performance comparison of Wavelet and Neural Network models

References Training

data (%)

Testing

data (%)

Forecast horizon Comparison with

other models

Level of accuracy

[74] 70 30 4–8; 8–16; 16–32 YA – MSE: 3.88–4.06

[75] 36 24 – ARMA-GARCH MSE:

0.0059–0.0131

[17]– – – – –

[76] 80 20 19 MA MLP MSE: 3.89; HR:

73 %; R2: 0.997

[43] 82 18 1 MA, 2MA, 3MA L-RIM, NL-RIM RMSE:1.486;

MAE:1.073;

MAPE:2.263

[77] 68 32 – – SSE: 6.16 ×10−5

[72] 56 50 1MA, 3MA, 6MA WANG, AMIN, STEO MSE: 2.05; MAE:

1.02; Max AE:

7.36

Tab le 1 9 Data characteristics, preprocessing technique, input variables and its selection method

References Oil

market

Time

scale

Input variable Input

selection

Preprocessing

[78]WTI;BrDV22 JCLT;CC

[79]WTIDV22 JCSMP

[64] WTI A V11, V4, V21, V10 JC –

neural network models. Ghaffari and Zare [79] presented a method based on soft

computing approaches to forecast WTI crude oil spot prices for one-month ahead

forecast horizon. Azadeh et al. [64] used oil supply, surplus capacity, Non-OECD

consumption, U.S. reﬁnery capacity and crude oil distillation capacity as input variable

for designing a ﬂexible ANN-FR algorithm to model noisy and complex oil prices.

Further, the ANOVA and Duncan multiple range test are used to test the signiﬁcance

of the forecast obtained from ANN and FR models. Table 19 indicates that the input

variables has been selected based on judgemental criterion and no prior selection

method has been applied.

Ghaffari [79] explored the possibility of smoothing procedure as preprocessing

tool to explore the pattern of oil prices while Panella et al. [78] incorporated both

log transformation and cluster classiﬁcation. As clear from Table 20, the prediction

accuracy of proposed neuro-fuzzy model by Panella et al. [78] has been compared

with linear and non-linear models on the basis of noise-to-signal ratios.

Ghaffari [79] has compared the results of their respective models with and without

smoothing procedure and observed that prediction quality in terms on percentage

of correct predictions has been improved by smoothing oil price data. It is evident

from Table 21 that researchers preferred to develop fuzzy model by adopting Takagi–

Sugeno–Kang as fuzzy inference system and Gaussian as membership function. The

123

N. Sehgal, K. K. Pandey

Tab le 2 0 Forecasting performance comparison of Fuzzy Neural Network models

References Training

data (%)

Testing

data (%)

Forecast

horizon

Comparison with

other models

Level of accuracy

[78] 67 33 1 SA LSE; RBF; MoGNN NSR: −46 to −24

[79] 80 20 30 DA WSP PCP: 68.18–70.09

[64] 80 20 – ANN, FR MAPE: 0.035

Tab le 2 1 Fuzzy Neural Networks model’s architecture

References Model

type

Learning

Algorithm

Hidden

neurons

Fuzzy inference

system

Membership

function

[78] MLP LSE + BP TE TSK Gaussian

[79] MLP LSE + GD TE TSK Gaussian

[64] MLP BFGS; BR;

BLR; GDX;

LM

TE – Gaussian

multi-layered perceptron neural network is the most popular among researchers for

oil price forecasting under this category of neuro-fuzzy approach. Azadeh [64] model

architecture includes MLP along with ﬁve variant of learning algorithm to improve

the forecasting performance and achieved MAPE as low as 0.035.

3.5.4 Decomposition based Neural Network

Yu et al. [80] proposed a “decomposition-and-ensemble” strategy using EMD-based

NN ensemble learning model to predict oil prices. Empirical mode decomposition is

proposed to decompose oil price data into eleven intrinsic mode functions. Xiong et al.

[81] evaluated the performance of EMD-based feed-forward neural network frame-

work incorporating slope-based method for oil price forecasting with three leading

strategies: direct, iterative and multiple-input multiple-output (MIMO).

Xiong et al. [81] used Symmetric MAPE (SMAPE) as a forecasting performance

criterion to evaluate the performance of EMD based neural network. In Table 22,

partial mutual information is used to examine the relationship between historic prices

and oil prices together with forward backward selection and delta test for model

identiﬁcation and estimation. Multi-layered perceptron neural network is the most

widely used neural network architecture by researchers for hybrid models as evident

from Tables 23,24.

3.5.5 Support vector and Genetic Algorithm

Guo [82] improved traditional SVR forecast precision by using genetic algorithm opti-

mized parameter of SVR in accordance with the training data. The model is found to

be effective in mapping the complexities of oil price series. Gabralla [83] investigated

123

Artiﬁcial intelligence methods for oil price forecasting...

Tab le 2 2 Data characteristics, preprocessing technique, input variables and its selection method

References Oil

market

Time

scale

Input

variable

Input

selection

Preprocessing

[80]WTI;BrDV22JC –

[81] WTI W V22 PMI; DT; FBS SR

Tab le 2 3 Forecasting performance comparison of decomposition based Neural Network models

References Training

set (%)

Testing

set (%)

Forecast

horizon

Comparison with

other models

Level of accuracy

[80] 72 28 1 DA ARIMA; MLP RMSE: 0.273–0.225;

DS: 86.99–87.81

[81] 67 33 4 WA MLP SMAPE: 2.28–8.15;

MAE: 0.81–1.28;

DS:54–87

Tab le 2 4 Decomposition based Neural Networks model’s architecture

References Model

type

Learning

Algorithm

Hidden

neurons

Activation

function

Decomposition

method

No. of

IMF’s

[80] MLP; ALNN BP – Lgs; Lin EMD 11

[81] MLP LM 15 – EMD –

performance of two different algorithms for feature selection together with several

machine learning methods (IBL, KStar and SMOreg) for oil price prediction.

4 Limitations

There is no solitary indicator driving crude oil prices. The output is based on how

much information is contained in the set of input variables selected for the study.

There are many studies that had examined the relationship between oil prices and

macroeconomic variables but there seem no consensus on the extent to which these

macroeconomic variables drive oil prices. Existing studies of predicting oil prices have

accounted for non-linearity, non-stationary and time-varying structure of the oil prices

but have seldom focus on selecting signiﬁcant features with high predicting power.

The empirical literature is very far from any consensus about selecting the appropriate

features/indicators that explains the characteristics of oil market. In most of the studies,

the design of input vector for oil price forecasting model is carried out on judgemental

criteria or trial and error procedures. Little attention is paid on selecting inﬂuential

factors and more on assessing new techniques for oil price forecasting.

The effect of input variables is considered to constantly driving oil prices in different

studies. There is shift in the inﬂuence of input variables subject to happening of any

123

N. Sehgal, K. K. Pandey

geopolitical and economic events in a given time-period but there is no literature

available that highlights this point. Predicted oil prices are dependent on short term

macroeconomic indicators whose effects are subject to structural changes. Few studies

have been carried out using artiﬁcial intelligent models to forecast complex oil price

series as compared to its application in diverse ﬁelds. Recently, researchers suggest

integration of neural networks with traditional methods as support vector regression,

genetic algorithms or wavelets by mean of hybrid approach to overcome limitations

of single models.

5 Concluding remark

Off late, artiﬁcial intelligent models are being extensively used to capture unknown

or too complex structures in time series. This review focussed on artiﬁcial intelligent

based oil price forecasting models and attempted to provide in-depth review based

on the following parameters: (i) type of model, (ii) input variables, (ii) input variable

selection method, (iv) data characteristics, (v) forecasting performance, and (vi) model

architecture. It enlisted the numerous key indicators used as input variables in artiﬁcial

intelligent based oil price forecasting models and attempted to highlight a serious issue

of selecting input variables based on judgemental or trial and error basis.

This review concludes that there is no single indicators driving oil prices and there

is need to identify the relevant input variables for oil price predictions. Multi-layered

perceptron neural network is most widely used by researchers for price forecasting.

Recently, researchers suggests integration of neural networks with traditional meth-

ods as support vector regression, genetic algorithms or wavelets by mean of hybrid

approach to overcome limitations of single models. It also highlighted that effect of

factors is constraint to happening of geopolitical and economic events. The research

gap highlighted to develop a robust feature selection method that can account for

non-linearity and time-varying structure of oil prices.

References

1. British Petroleum. Statistical review of world energy 2013. British Petroleum (2013)

2. International Energy Agency. World Energy Outlook, 2012. OECD/IEA (2012)

3. Benning, C., Pichersky, E.: Harnessing plant biomass for biofuels and biomaterials. Plant J. 54(4),

533–535 (2008)

4. Andreou, E., Pittis, N., Spanos, A.: On modelling speculative prices: the empirical literature. J. Econ.

Surv. 15(2), 187–220 (2001)

5. Arouri, M.E.H., Lahiani, A., Lévy, A., Nguyen, D.K.: Forecasting the conditional volatility of oil

spot and futures prices with structural breaks and long memory models. Energy Econ. 34(1), 283–293

(2012)

6. Bing, X., Ouenniche, J.: A data envelopment analysis-based framework for the relative performance

evaluation of competing crude oil prices’ volatility forecasting models. Energy Econ. 34(2), 576–583

(2012)

7. He, K., Yu, L., Lai, K.K.: Crude oil price analysis and forecasting using wavelet decomposed ensemble

model. Energy 46(1):564–574 (2012)

8. Mohammadi, H., Lixian, S.: International evidence on crude oil price dynamics: applications of

ARIMA-GARCH models. Energy Econ. 32(5), 1001–1008 (2010)

123

Artiﬁcial intelligence methods for oil price forecasting...

9. Tian, Z., Swanson, N.R.: Predictive evaluationof econometric forecasting models in commodity futures

markets. Stud. Nonlinear Dyn. Econom. 2(4), 159–177 (1998)

10. Zhang, X., Wu, Q., Zhang, J.: Crude oil price forecasting using fuzzy time series. In: 2010 3rd Inter-

national Symposium on Knowledge Acquisition and Modeling (KAM), pp. 213–216. IEEE (2010)

11. Haidar, I., Kulkarni, S., Pan, H.: Forecasting model for crude oil prices based on artiﬁcial neural

networks. In: International Conference on Intelligent Sensors, Sensor Networks and Information

Processing, 2008 (ISSNIP 2008), pp. 103–108 (2008)

12. Chinn, MD., LeBlanc, M., Coibion, O.: The predictive content of energy futures: an update on petro-

leum, natural gas, heating oil and gasoline. Technical report, National Bureau of Economic Research

(2005)

13. Alvarez-Ramirez, J., Rodriguez, E., Martina, E., Ibarra-Valdez, C.: Cyclical behavior of crude oil

markets and economic recessions in the period 1986–2010. Technol. Forecast. Soc. Change 79(1),

47–58 (2012)

14. Movagharnejad, K., Mehdizadeh, B., Banihashemi, M., Kordkheili, M.S.: Forecasting the differences

between various commercial oil prices in the persian gulf region by neural network. Energy 36(7),

3979–3984 (2011)

15. Weiqi, L., Linwei, M., Yaping, D., Pei, L.: An econometric modeling approach to short-term crude oil

price forecasting. In: Control Conference (CCC), 2011 30th Chinese, pp. 1582–1585. IEEE (2011)

16. Dées, S., Gasteuil, A., Kaufmann, R., Mann, M.: Assessing the factors behind oil price changes (2008)

17. Jinliang, Z., Mingming, T., Mingxin, T.: Effects simulation of international gold prices on crude oil

prices based on WBNNK model. In: ISECS International Colloquium on Computing, Communication,

Control, and Management, 2009 (CCCM 2009), vol. 4, pp. 459–463 (2009)

18. Abdullah, S.N. Zeng, X.: Machine learning approach for crude oil price prediction with artiﬁcial neural

networks-quantitative (ann-q) model. In: The 2010 International Joint Conference on Neural Networks

(IJCNN), pp. 1–8 (2010)

19. Krichene, N.: Crude oil prices: trends and forecast. IMF Working Papers, pp. 1–23 (2008)

20. de Souza e Silva, E.G., Legey, L.F.L., de Souza e Silva, E.A.: Forecasting oil price trends using wavelets

and hidden markov models. Energy Econ. 32(6), 1507–1519 (2010)

21. Chernenko, S.V., Schwarz, K.B., Wright, J.H.: The information content of forward and futures prices:

market expectations and the price of risk. Board of Governors of the Federal Reserve System (2004)

22. He, A.W.W., Kwok, J.T.K., Wan, A.T.K.: An empirical model of daily highs and lows of west texas

intermediate crude oil prices. Energy Econ. 32(6), 1499–1506 (2010)

23. Pindyck, R.S.: The long-run evolution of energy prices. The Energy Journal. 20, 1–27 (1999)

24. Eduardo, S., Smith, J.E.: Short-term variations and long-term dynamics in commodity prices. Manag.

Sci. 46(7), 893–911 (2000)

25. Mazraati, Mohammad, Jazayeri, S.M.: Oil price movements and production agreements. OPEC Rev.

28(3), 207–226 (2004)

26. Moshiri, S., Foroutan, F.: Forecasting nonlinear crude oil futures prices. Energy J. 27(4), 81–96 (2006)

27. Radziukyniene, I., Boyko, N., Pardalos, P.:Model-based forecasting. Wiley encyclopedia of operations

research and management science, vol. 5, pp. 3305–3312 (2011)

28. Hou, A., Suardi, S.: A nonparametric GARCH model of crude oil price return volatility. Energy Econ.

34(2), 618–626 (2012)

29. Sang, H.K., Kang, S.M., Yoon, S.M.: Forecasting volatility of crude oil markets. Energy Econ. 31(1),

119–125 (2009)

30. Wei, Y., Wang, Y., Huang, D.: Forecasting crude oil market volatility: further evidence using GARCH-

class models. Energy Econ. 32(6), 1477–1484 (2010)

31. Huang, B.-N., Yang, C.W., Hwang, M.J.: The dynamics of a nonlinear relationship between crude

oil spot and futures prices: a multivariate threshold regression approach. Energy Econ. 31(1), 91–98

(2009)

32. Lixia, L.: Nonlinear test and forecasting of petroleum futures prices time series. Energy Procedia 5,

754–758 (2011)

33. Kisswani, K.M., Nusair, S.A.: Non-linearities in the dynamics of oil prices. Energy Econ. 36, 341–353

(2013)

34. James, H.D.: Nonlinearities and the macroeconomic effects of oil prices. Macroecon. Dyn. 15(S3),

364–378 (2011)

35. Ye, M., Zyren, J., Shore, J.: Forecasting crude oil spot price using OECD petroleum inventory levels.

Int. Adv. Econ. Res. 8(4), 324–333 (2002)

123

N. Sehgal, K. K. Pandey

36. Ye, M., Zyren, J., Shore, J.: A monthly crude oil spot price forecasting model using relative inventories.

Int. J. Forecast. 21(3), 491–501 (2005)

37. Ye, M., Zyren, J., Shore, J.: Forecasting short-run crude oil price using high-and low-inventory vari-

ables. Energy Policy 34(17), 2736–2743 (2006)

38. Zamani, M.: An econometrics forecasting model of short term oil spot price. In: 6th IAEE European

Conference, Citeseer (2004)

39. Chai, J., Guo, J.E., Meng, L., Wang, S.Y.: Exploring the core factors and its dynamic effects on oil

price: An application on path analysis and BVAR-TVP model. Energy Policy 39(12), 8022–8036

(2011)

40. Frey, G., Manera, M., Markandya, A., Scarpa, E.: Econometric models for oil price forecasting: a critical

survey. In: CESifo Forum, vol. 10, pp. 29–44. Institute for Economic Research at the University of

Munich (2009)

41. Juan, R.C.: How do crude oil prices co-move?: A copula approach. Energy Econ. 33(5), 948–955

(2011)

42. International Energy Agency. Oil Market Report, 2014. OECD/IEA (2014)

43. Pang, Y., Wei, X., Lean, Y., Ma, J., Lai, K.K., Wang, S., Xu, S.: Forecasting the crude oil spot price by

wavelet neural networks using OECD petroleum inventory levels. New Math. Nat. Comput. 07(02),

281–297 (2011)

44. Timothy, C.J., Heo, E.: Price and inventory dynamics in petroleum product markets. Energy Econ.

22(5), 527–548 (2000)

45. Kim, D.H.: What is an oil shock? Panel data evidence. Empir. Econ. 43(1), 121–143 (2012)

46. Dées, S., Karadeloglou, P., Kaufmann, R.K., Sanchez, M.: Modelling the world oil market: assessment

of a quarterly econometric model. Energy Policy 35(1), 178–191 (2007)

47. Ratti, R.A., Vespignani, J.L.: Crude oil prices and liquidity, the BRIC and G3 countries. Energy Econ.

39, 28–38 (2013)

48. Li, H., Lin, S.X.: Do emerging markets matter in the world oil pricing system? Evidence of imported

crude by China and India. Energy Policy 39(8), 4624–4630 (2011)

49. Zhang, Y.-J., Wang, Z.-Y.: Investigating the price discovery and risk transfer functions in the

crude oil and gasoline futures markets: some empirical evidence. Appl. Energy 104, 220–228

(2013)

50. Murat, A., Tokat, E.: Forecasting oil price movements with crack spread futures. Energy Econ. 31(1),

85–90 (2009)

51. Yang, W., Han, A., Cai, K., Wang, S.: ACIX model with interval dummy variables and its application

in forecasting interval-valued crude oil prices. Procedia Comput. Sci. 9, 1273–1282 (2012)

52. He, Y., Wang, S., Lai, K.K.: Global economic activity and crude oil prices: a cointegration analysis.

Energy Econ. 32(4), 868–876 (2010)

53. Hui, B.: Price dynamics and speculators in crude oil futures market. Syst. Eng. Procedia 2, 114–121

(2011)

54. Syed Abul, B., Haug, A.A., Sadorsky, P.: Oil prices, exchange rates and emerging stock markets.

Energy Econ. 34(1), 227–240 (2012)

55. Zhang, X., Yu, L., Wang, S., Lai, K.K.: Estimating the impact of extreme events on crude oil price: an

EMD-based event analysis method. Energy Econ. 31(5), 768–778 (2009)

56. Fong, W.M., See, K.H.: A markov switching model of the conditional volatility of crude oil futures

prices. Energy Econ. 24(1), 71–95 (2002)

57. Haidar, I., Kulkarni, S., Pan, H.: Forecasting model for crude oil prices based on artiﬁcial neural

networks. In: International Conference on Intelligent Sensors, Sensor Networks and Information

Processing, 2008 (ISSNIP 2008), pp. 103–108. IEEE (2008)

58. Mahdi, A.A., Hussain, A.J., Al-Jumeily, D.: Adaptive neural network model using the immune sys-

tem for ﬁnancial time series forecasting. In: International Conference on Computational Intelligence,

Modelling and Simulation, 2009 (CSSim ’09), pp. 104–109 (2009)

59. Bao, Y., Yang, Y., Xiong, T., Zhang, J.: A comparative study of multi-step-ahead prediction for crude oil

price with support vector regression. In: 2011 Fourth International Joint Conference on Computational

Sciences and Optimization (CSO), pp. 598–602 (2011)

60. Khashman, A., Nwulu, N.I.: Intelligent prediction of crude oil price using support vector machines. In:

2011 IEEE 9th International Symposium on Applied Machine Intelligence and Informatics (SAMI),

pp. 165–169 (2011)

123

Artiﬁcial intelligence methods for oil price forecasting...

61. He, K., Lai, K.K., Yen, J.: Crude oil price prediction using slantlet denoising based hybrid models. In:

International Joint Conference on Computational Sciences and Optimization, 2009 (CSO 2009), vol.

2, pp. 12–16 (2009)

62. Zhu, J.-R.: A new model for oil futures price forecasting based on cluster analysis. In: 4th International

Conference on Wireless Communications, Networking and Mobile Computing, 2008 (WiCOM ’08),

pp. 1–4 (2008)

63. He, K., Lai, K.K., Yen, J.: Morphological component analysis based hybrid approach for prediction

of crude oil price. In: 2010 Third International Joint Conference on Computational Science and Opti-

mization (CSO), vol. 1, pp. 423–427 (2010)

64. Azadeh, A., Moghaddam, M., Khakzad, M., Ebrahimipour, V.: A ﬂexible neural network-fuzzy math-

ematical programming algorithm for improvement of oil price estimation and forecasting. Comput.

Ind. Eng. 62(2), 421–430 (2012)

65. Malliaris, M.E., Malliaris, S.G.: Forecasting energy product prices. In: Proceedings of 2005 IEEE

International Joint Conference on Neural Networks, 2005 (IJCNN ’05), vol. 5, pp. 3284–3289 (2005)

66. Alizadeh, A., Maﬁnezhad, K.: Monthly brent oil price forecasting using artiﬁcial neural networks and

a crisis index. In: 2010 International Conference on Electronics and Information Engineering (ICEIE),

vol. 2, pp. V2-465–V2-468. IEEE (2010)

67. Xie, W., Lean, Y., Shanying, X., Wang, S.: A new method for crude oil price forecasting based on

support vector machines. In: Computational Science ICCS 2006. Lecture Notes in Computer Science,

vol. 3994, pp. 444–451. Springer, Berlin (2006)

68. Kaboudan, M.A.: Compumetric forecasting of crude oil prices. In: Proceedings of the 2001 Congress

on Evolutionary Computation, 2001, vol. 1, pp. 283–287. IEEE (2001)

69. Amin-Naseri, M.R., Gharacheh, E.A.: A hybrid artiﬁcial intelligence approach to monthly forecasting

of crude oil price time series. In: The Proceedings of the 10th International Conference on Engineering

Applications of Neural Networks, CEUR-WS284, pp. 160–167 (2007)

70. Xiao, J., He, C., Wang, S.: Crude oil price forecasting: a transfer learning based analog complexing

model. In: 2012 Fifth International Conference on Business Intelligence and Financial Engineering

(BIFE), pp. 29–33. IEEE (2012)

71. Youseﬁ, S., Weinreich, I., Reinarz, D.: Wavelet-based prediction of oil prices. Chaos Solitons Fractals

25(2), 265–275 (2005)

72. Alexandridis, A., Livanis, E.: Forecasting crude oil prices using wavelet neural networks. In: Proceed-

ings of the 5th FSDET (ET), Athens, Greece, pp. 8 (2008)

73. Fan, Y., Liang, Q., Wei, Y.-M.: A generalized pattern matching approach for multi-step prediction of

crude oil price. Energy Econ. 30(3), 889–904 (2008)

74. Mingming, T., Jinliang, Z.: A multiple adaptive wavelet recurrent neural network model to analyze

crude oil prices. J. Econ. Bus. 64(4), 275–286 (2012)

75. He, K., Xie, C., Chen, S., Lai, K.K.: Estimating VaR in crude oil market: a novel multi-scale non-

linear ensemble approach incorporating wavelet analysis and neural network. Neurocomputing 72,

3428–3438 (2009)

76. Jammazi, R., Aloui, C.: Crude oil price forecasting: experimental evidence from wavelet decomposition

and neural network modeling. Energy Econ. 34(3), 828–841 (2012)

77. Qunli, W., Ge, H., Xiaodong, C.: Crude oil price forecasting with an improved model based on wavelet

transform and RBF neural network. In: International Forum on Information Technology and Applica-

tions, 2009 (IFITA ’09), vol. 1, pp. 231–234 (2009)

78. Panella, M., Liparulo, L., Barcellona, F., D’Ecclesia, R.L.: A study on crude oil prices modeled by

neurofuzzy networks. In: 2013 IEEE International Conference on Fuzzy Systems (FUZZ), pp. 1–7

(2013)

79. Ghaffari, A., Zare, S.: A novel algorithm for prediction of crude oil price variation based on soft

computing. Energy Econ. 31(4), 531–536 (2009)

80. Yu, L., Wang, S., Lai, K.K.L Forecasting crude oil price with an EMD-based neural network ensemble

learning paradigm. Energy Econ. 30(5), 2623–2635 (2008)

81. Xiong, T., Bao, Y., Zhongyi, H.: Beyond one-step-ahead forecasting: evaluation of alternative multi-

step-ahead forecasting models for crude oil prices. Energy Econ. 40, 405–415 (2013)

82. Guo, X., Li, D.C., Zhang, A.: Improved support vector machine oil price forecast model based on

genetic algorithm optimization parameters. AASRI Procedia 1, 525–530 (2012)

123

N. Sehgal, K. K. Pandey

83. Gabralla, L.A., Jammazi, R., Abraham, A.: Oil price prediction using ensemble machine learning. In:

2013 International Conference on Computing, Electrical and Electronics Engineering (ICCEEE), pp.

674–679. IEEE (2013)

84. Sadorsky, P.: Oil price shocks and stock market activity. Energy Econ. 21(5), 449–469 (1999)

85. Lin, A.: Prediction of international crude oil futures price based on GM (1, 1). In: IEEE International

Conference on Grey Systems and Intelligent Services, 2009 (GSIS 2009), pp. 692–696. IEEE (2009)

86. Chang, K.L.: Volatility regimes, asymmetric basis effects and forecasting performance: an empirical

investigation of the WTI crude oil futures market. Energy Econ. 34(1), 294–306 (2012)

87. Vo, M.T.: Regime-switching stochastic volatility: evidence from the crude oil market. Energy Econ.

31(5), 779–788 (2009)

88. Liu, L., Wan, J.: A study of shanghai fuel oil futures price volatility based on high frequency data:

long-range dependence, modeling and forecasting. Econ. Model. 29(6), 2245–2253 (2012)

89. Zhang, X., Lai, K.K., Wang, S.-Y.: A new approach for crude oil price analysis based on empirical

mode decomposition. Energy Econ. 30(3), 905–918 (2008)

90. Huang, D., Baimin, Y., Fabozzi, F.J., Fukushima, M.: CAViar-based forecast for oil price risk. Energy

Econ. 31(4), 511–518 (2009)

91. Wei, Y.: Forecasting volatility of fuel oil futures in China: GARCH-type, SV or realized volatility

models? Phys. A Stat. Mech. Appl. 391(22), 5546–5556 (2012)

92. Chih, C.W., Chung, H., Chang, Y.H.: The economic value of co-movement between oil price and

exchange rate using copula-based GARCH models. Energy Econ. 34(1), 270–282 (2012)

93. Yang, C.W., Hwang, M.-J., Huang, B.-N.: An analysis of factors affecting price volatility of the US

oil market. Energy Econ. 24(2), 107–119 (2002)

123