A Novel Approach to Forecast Crude Oil Prices Using Machine Learning and Technical Indicators

Full text

A Novel Approach to Forecast Crude Oil Prices
Using Machine Learning and Technical
Indicators
Kshitij A. Kakade 1, Kshitish S. Ghate 1, Raj K. Jaiswal 2,*, and Ritika Jaiswal 3
1 Department of Computer Science and Information Systems, Birla Institute of Technology and Science Pilani,
KK Birla Goa Campus, India; Email: {kshitijkakade2705, ghatekshitish}@gmail.com (K.A.K., K.S.G.)
2 National Forensic Sciences University, Goa Campus, India
3 Department of Economics & Finance, Birla Institute of Technology and Science Pilani, KK Birla Goa Campus, India;
Email: ritikaj@goa.bits-pilani.ac.in (R.J.)
*Correspondence: jaiswal.raaj@gmail.com (R.K.J.)
Abstract—This study proposes to use a hybrid ensemble
learning approach to improve the prediction efficiency of
crude oil prices. It combines the Long Short-Term Memory
(LSTM) with factors that influence the price of crude oil.
The information from fundamental and technical indicators
is considered along with statistical model predictions like
autoregressive integrated moving average (ARIMA)to make
one-step-ahead crude oil price predictions. A Principal
Component Analysis (PCA) approach is employed to
transform the explanatory variables. This study combines
the LSTM with PCA, jointly known as the LP model
wherein PCA transforms of the fundamental and technical
indicators are used as inputs to improve LSTM predictions.
Further, it attempts to improve these predictions by
introducing the LSTM+PCA+ARIMA (LPA) model, which
uses an ensemble learning approach to utilize the forecast
from the ARIMA model, as an additional input. Among LP
and LPA models, the LSTM model is used as a benchmark
to evaluate the performance of the hybrid models. Based on
the result, a significant improvement is seen in the LP model
over the chosen window sizes and error metrics. On the
other hand, the LPA model performs better across all
dimensions with an average improvement of 41% over the
LSTM model in terms of forecasting accuracy. Moreover,
the equivalence of forecasting accuracy is tested using the
Diebold-Mariano and Wilcoxon signed-rank tests.
Keywords—Long Short-Term Memory (LSTM), Principal
Component Analysis (PCA), ensemble learning, crude oil,
forecasting
I. INTRODUCTION
Crude oil is an indispensable non-renewable
commodity, responsible for meeting nearly a third of the
global energy demand [1]. It has far-reaching industrial
uses and is one of the most actively traded commodities
that exhibit significant volatility in its prices [2]. West
Texas Intermediate (WTI), which is traded on the New
York Stock Exchange, is widely regarded as the global
Manuscript received June 5, 2022; revised July 1, 2022; accepted July
18, 2022; published April 4, 2023.
benchmark of oil trading due to the strength of the USA
crude oil buyers, along with the global influence of the
New York Exchange [3].
Oil price changes have significant implications for
macroeconomic conditions. A substantial rise in oil prices
indicates inflation and subsequent recession for countries
that import oil, while falling prices may be detrimental to
the economic growth of oil-exporting nations. A study by
Katircioglu et al. showed that oil-price fluctuations
negatively impact the GDP, CPI, and unemployment in
Organisation for Economic Co-operation and
Development (OECD) countries in the long term [4].
Reboredo and Ugolini found a significant impact of the
oil market on the stock market for three developed and
five BRICS countries [5]. Price fluctuations have been
increasing with economic globalization and liberalization,
which has added to the overall revenue risk [6]. A
combination of supply, demand, inventory, and non-
fundamental parameters such as the exchange rate and
interest rate make the price prediction of crude oil
complex [7, 8].
This study aims to develop a deep learning framework
that incorporates fundamental and technical indicators
along with predictions from statistical models to forecast
one day ahead crude oil prices. The fundamental
variables include energy indices, stock prices of major oil
companies, interest and exchange rates, price of
substitute energy products, and other assets sharing
strong relationships with crude oil. Different technical
indicators such as the Simple Moving Average Crossover
(SMA), Relative Strength Index (RSI), Rate of change
(ROC), Moving Average Convergence Divergence
(MACD), and Bollinger Band Squeeze are also included
in the model.
A PCA transformation is used to remove the impact of
multicollinearity and reduce the input dimension of the
LSTM network. We introduce the LSTM+PCA (LP)
model, which includes the principal components of the
transformed fundamental and technical indicators
mentioned above as inputs. We also develop the
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
302
doi: 10.12720/jait.14.2.302-310

LSTM+PCA+ARIMA (LPA) model by extending the LP
model in combination with the ARIMA model’s
forecasting abilities to enhance the accuracy of model
predictions. The ARIMA forecasts are calculated using a
252-day rolling window over the selected period and are
combined with the LP model as an additional explanatory
variable. An ensemble learning approach that involves
combining multiple learning algorithms to obtain their
collective performance generates the LPA framework.
There are only limited studies that forecast crude oil
prices based on information from fundamental, technical,
and statistical variables using a deep learning model.
The proposed model efficiency was evaluated over
four different window sizes (3, 5, 7, and 11 days). The
root Mean Squared Error (RMSE), Mean Absolute Error
(MAE), and Mean Absolute Percentage Error (MAPE)
metrics are used to evaluate the forecasting performance
of the models for one-step-ahead crude oil prices. The
results of the study strongly confirm our hypotheses. The
LP model improves performance compared to the
baseline LSTM framework over all chosen window sizes
and error metrics. The LPA is the best performing model
across all dimensions by a significant margin. We also
conduct the Diebold-Mariano (DM) and Wilcoxon
signed-rank (WS) tests to compare the predictive
accuracy of the two forecasting models.
The remaining section of the paper is organized as
follows: Section II discusses the existing literature, while
Section III outlines the data used in the study. Section IV
discusses the methodology used, and Section V presents
the experimental setup. The analysis of the results is done
in Section VI, while Section VII concludes the study.
II. RELATED WORK
Predictive frameworks for asset prices in financial
studies are generally modelled based on fundamental or
technical variables. Dees et al. consider supply and
demand factors such as OPEC production, production
capacity, oil inventories, and demand while modelling the
price of crude oil [9]. Similarly, Baumeister and Kilian
utilize fundamental economic variables to develop six
models to forecast commodity prices accurately [10].
Miao et al. tests the significance of six different
categories of variables for oil price forecasting using a
LASSO model with various supply, demand, financial,
and commodity market variables [7].
Technical analysis estimates the direction of asset
prices from trading activity, such as price movement and
volume [11, 12]. Yin and Yang use principal component
predictive regressions to systematically uncover the
components of technical indicators with oil price
forecasting power [13]. Liu and Wang et al. used moving
average rules and macroeconomic indicators to generate
density forecasts [14]. They find that the technical
indicators generate a more accurate density forecast when
compared to the macro variables.
Statistical models are well-known tools for forecasting
asset prices. Numerous classes of statistical models, such
as random walk [15], generalized autoregressive
conditional heteroskedasticity (GARCH) [16], and
ARIMA [17] have been used to forecast oil prices.
ARIMA is a linear model used for univariate time series
analysis and forecasting. Yusof and Rashid et al. applied
this model to forecast crude oil production in Malaysia
for three leading months [18]. Concurrently, Mohammadi
and Su used a hybrid ARIMA-GARCH model to model
crude oil volatility and compare the accuracy of their
framework with four other volatility models [19].
An approach to forecasting has been widely explored
in the recent times is the use of machine learning
algorithms. Kusonkhum et al. used k-Nearest Neighbours
(KNN) to predict over-budget construction projects
achieving an overall accuracy of 0.86 [20]. Several other
machine learning algorithms such as decision trees [21],
support vector machines [22], Logistic Regression and
Random Forest [23] have been used for prediction and
forecasting purposes. Artificial neural networks (ANN)
are nonlinear functions that simultaneously capture
hidden patterns between input and output variables
without any underlying assumptions. Several studies have
shown that models based on neural networks have
outperformed conventional forecasting and prediction
models. ANN is the most commonly used nonlinear AI
model. Bakshi et al. used convolutional neural network
(CNN) model, an extension of ANN, for predicting
pregnant shoppers based on their transaction history and
purchasing trends [24]. Other extensions, such as the
recurrent neural networks (RNN), use loops to iterate
over the series while maintaining an internal state that
stores information about the steps it has seen so far.
These models are efficient in modelling time series data
but are often prone to the exploding gradient problem.
The LSTM model proposed by Hochreiter and
Schmidhuber is a class of RNN models that are not
vulnerable to the vanishing gradient problem [25]. The
LSTM model has thus been widely adopted for time
series modelling as it excels at extracting patterns in an
input feature space, where the data spans long sequences.
Kubra, Sekeroglu et al. used LSTM to perform lung
cancer incidence prediction for ten European
countries [26]. However, neural networks have their own
limitations compared to machine learning algorithms.
Premsmith and Ketmaneechairat utilized the logistic
regression and Neural Network model for heart disease
detection and found that the logistic regression model
outperforms the neural network [27].
Recently, several studies have experimented with
combining the capabilities of neural networks with the
information generated through fundamental, technical,
and statistical methods to forecast asset prices. The study
indicated that RNN performed better than the two ANN
models. Dropsy employed neural networks as a nonlinear
forecasting tool for forecasting international equity risk
premia in the markets of Germany, Japan, the United
Kingdom, and the United States from 1970 to 1990 [28].
Chiroma et al. showed that evolutionary neural networks
developed using genetic algorithms can show significant
performance improvements in predicting crude oil prices
compared to known statistical models [29]. The study
compiles a comprehensive list of applications of hybrid
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
303

neural networks to forecast crude oil prices. Wu et al.
forecasted crude oil prices using a hybrid model utilizing
ensemble empirical mode decomposition, comprising
sparse Bayesian learning and ARIMA forecasts [30].
Suhermi et al. used a hybrid methodology to integrate the
ARIMA model and an ANN model and observe an
improvement in forecast accuracy compared to the non-
hybrid models [31].
The approach that differentiates our work from other
existing methods is the use of an ensemble learning
method where we train multiple LSTM layers on inputs
of crude oil prices along with explanatory variables and
statistical forecasts. The novelty is in using an additional
ANN that learns from the forecasts of the LSTM
networks to yield significantly improved predictions of
crude oil prices. The fluctuations in oil prices can be
attributed to many factors. Thus, there is a need to
incorporate relevant macroeconomic and technical
variables to predict crude oil prices accurately. The
following section details the explanatory variables used
in this study and highlights their relevance in explaining
oil prices.
III. DATA
We collect crude oil prices from the US Energy
Information Administration database, the principal
agency responsible for managing energy information.
The model is developed for WTI spot price, which is the
global standard for crude oil prices. The data is collected
for the period starting from 28-07-2000 to 13-05-2019.
The macroeconomic indicators and financial time series
are collected from the FRED Database, US Energy
Information Administration, and Nasdaq Database. Table
I presents an overview of the financial time series and
their sources. A few of the time series measuring supply
and demand levels of crude oil are available only on a
weekly scale. In such cases, a linear interpolation is used
for filling up the missing value while considering the
weekly values on the last Friday of the month. Tables I
and II describe the explanatory variables used in this
study.
TABLE I. LIST OF EXPLANATORY VARIABLES
Macroeconomic Variable
Frequency
Source
ExxonMobil Closing Price
Daily
Nasdaq Database
Royal Dutch Shell Closing Price
Daily
Chevron Closing Price
Daily
PetroChina Closing Price
Daily
Total Energies Closing prices
Daily
S&P GSCI Energy Index
Daily
U.S. Days of Supply of Crude Oil
Weekly
Energy Information
Administration
database
U.S. Ending Stocks of Crude Oil
Weekly
U.S. Exports of Crude Oil
Weekly
U.S. Field Production of Crude Oil
Weekly
U.S. Imports of Crude Oil
Weekly
US Refiner Net Input of Crude Oil
Weekly
Cushing, OK WTI Spot Price
Daily
Europe Brent Spot Price
Daily
Cushing Crude Oil Future Contract
Daily
Henry Hub Natural Gas Spot Price
Daily
3 Month Treasury
Daily
FRED Database
US Dollar Index
Daily
Effective Federal Funds Rate
Daily
Gold Price
Daily
TABLE II. LIST OF TECHNICAL VARIABLES
Technical Indicator Name
Lag length (in days)
Simple moving average (SMA) crossover
1, 3, 6, 9, 12
Moving average convergence divergence
(MACD)
(12, 26)
Price Rate of Change (ROC)
3, 6, 9, 12
Relative Strength Index (RSI)
14
Bollinger Bands (Upper, Lower, Squeeze)
20
A. Fundamental Variables
Gold is a global commodity generally used to hedge
against inflation and has shared a direct relationship with
crude oil over time. Wang and Chueh showed that both
gold and crude oil prices positively influence each other
in the short term [32]. The prices of natural gas, heating
oil, and gasoline are used in the study as they serve as
significant fuels in the energy mix. Villar and Joutz [33]
and Batten et al. [34] found a robust leading relationship
from natural gas to crude oil arising out of demand and
supply factors. Additional variables relating to the import,
export, refinery net input, field production, ending stocks,
and remaining days of supply of crude oil are also
included. These demand and supply factors related to
crude oil production are considered to account for the
short-term changes associated with production, which are
easily affected by several geopolitical and natural factors.
Consequently, there is an impact on the reserves and
demand for crude oil. The economic policy uncertainty
(EPU) index measures the uncertainty in policies related
to economic decisions. Crude oil is an asset that is
sensitive to policy changes, and the uncertainty index can
be used to capture the political environment.
B. Technical Variables
While numerous studies conclude the importance of
macroeconomic factors in impacting crude oil price
movements, a significant strand of literature also
considers the importance of technical indicators in
explaining price movements. It has been well established
that technical indicator can explicitly aid in
understanding asset price movements since investors are
known to carry out trading decisions based on the
technical analysis of historical prices [35, 36]. There are,
however, only a few studies that consider both technical
and macroeconomic indicators as significant explanatory
variables in the context of crude oil [12–14]. This study
extends the literature by including relevant trading
signals into the forecasting model in addition to the
fundamental variables already defined above. The
technical indicators chosen are the SMA, RSI, ROC,
MACD, and Bollinger Band Squeeze.
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
304

The SMA is a momentum-based indicator that
generates a buy (sell) signal when a shorter-term (
periods) SMA crosses above (below) the longer-term (
periods) SMA. Ten crossover signals are obtained for
crude which are generated using ,  = 1, 3, 6, 9, 12. The
RSI is a momentum oscillator that measures the speed
and change of asset price movements and is used to
identify overbought and oversold conditions. The MACD
is a trend following indicator calculated by subtracting
the exponential moving average (EMA) over nine periods
by the 26 periods EMA. Price ROC is another
momentum-based technical indicator that compares the
current asset price with the price from a certain number
of previous periods. We compute the ROC for 3, 6, 9, and
12 days. Unlike MA and ROC, Bollinger Bands are chart
indicators of technical analysis. We use the upper and
lower band and the band squeeze in our analysis. These
indicators, which depend only on the closing price of the
financial asset, are chosen in line with previous studies
such as [11, 14, 36, 37]. Crude oil is one of the most
actively traded commodities on the exchange, and these
indicators could provide some predictive power in
determining the price trends for the model.
TABLE III. DESCRIPTIVE STATISTICS OF THE CRUDE OIL PRICES
Crude Oil Price (in $)
Min
17.5
Max
145.31
Mean
63.1076
Std
26.4272
Skew
0.3322
Kurtosis
-0.7658
Count
4575
Table III describes the summary statistics of the WTI
crude oil spot prices. There is a total of 4575 data points.
The series has a mean of 63.107 $. Furthermore, a skew
of 0.332264 and a kurtosis of −0.765801 exist within the
series. Fig. 1 depicts a plot of the crude oil prices, with a
vertical line drawn to mark the split between train and
test data points, as used by the LSTM networks.
Figure 1. Actual WTI crude oil price from 28-07-2000 to 13-05-2019.
IV. METHODOLOGY
A. Principal Component Analysis (PCA)
We transform the fundamental and technical variables
, ( = 1, 2, ...) using PCA [38] to reduce the
dimensionality of the data and remove multicollinearity.
A defined number of principal components is obtained ,
( = 1, ...,  ≤ ), which are independent of one another.
The PCA technique works by extracting diffusion
indexes as a linear combination of the predictors.
Working directly with the raw data without standardizing
would lead to improper transformations since more
weight would be assigned to those variables with
relatively higher variances, mainly when the variables are
measured in different units. Thus, all variables are
standardized before applying the PCA transform for
dimensionality reduction.
+ + ... +
+ + ... +
(1)
+ + ... +
where , denotes the ℎ Eigen value of the th principal
component. The first , (<) principal components that
represent a majority of the total information in terms of
the variance ratio [8] are selected as inputs to the models.
B. ARIMA Model
The ARIMA model introduced by Box and Jenkins is a
statistical technique used for analysing time series
forecasting [39]. The ARIMA model uses a linear
relationship between the predicted value as a function of
a certain number of lagged observations and the lagged
values of the residual errors. In general, the ARIMA
model is expressed as
󰇛󰇜󰇛󰇜
󰇛󰇜 (2)
where,
󰇛󰇜󰇛󰇜󰇛󰇜 󰇛󰇜 (3)
󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜 (4)
  (5)
 denotes the observations,  denotes the backshift
operator, and  denotes the white noise sequence, ~
 (0, 2).  ( = 1, 2, ...),  ( = 1, 2, ...,), and  are
model parameters.  denotes the order of differencing.
C. LSTM Model
Traditional feed-forward networks have been extended
to develop RNN, which can forecast long data sequences
by utilizing internal loops derived from input sequences.
RNNs maintain an internal cell state that updates every
step of the series, thus remembering and producing
results based on past observations. However, RNNs face
certain drawbacks due to their inability to factor in errors
...
...
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
305

from older observations while training, making them
inefficient to model long-run dependencies. The
vanishing (or exploding gradients) problem where
weights allocated while training are too small (or large) is
one of the major concerns regarding the usability of
RNNs. A class of RNN that avoids the issues mentioned
above is the LSTM model [25]. It operates with the help
of memory cells or states that remember information in
the long run and forget past data that is unnecessary.
󰇛󰇜 (6)
󰇛󰇜 (7)
󰇛󰇜 (8)
󰇛󰇜 (9)
  (10)
󰇛󰇜 (11)
Figure 2. Diagram of LSTM model.
Fig. 2 illustrates the components of the LSTM
comprising of a memory cell () along with three gates:
an input gate (), a forget gate (), and an output gate
(). The computations associated with the cell state (),
the hidden state (ℎ), and the three gates are described in
Eqs. (6)–(11). 0 = 0 and ℎ0 = 0 are initialized prior to
calculations. The input is represented by ℎ and the
hidden state by 
, at a given time t. The value  (input
modulate gate) determines the amount of new
information received by the cell for every time step. 0
and 0 are weight matrices in these equations,  are a
bias term,  is a sigmoid function,  is the hyperbolic
tangent function, and the symbol  denotes element-wise
multiplication.
D. Test for Comparing Equivalence of Forecast
Accuracy
Diebold-Mariano (DM) test and Wilcoxon signed-rank
(WS) test are utilized to compare the models’ predictive
forecasting accuracy. It provides a framework to
determine whether the difference in the predictive
accuracy of the models is significant for forecasting
purposes or is just a result of the choice of data.
• Diebold-Mariano (DM) test
The DM test evaluates each forecast’s quality by a
predefined loss function g of the forecast error [40]. The
null hypothesis of equal predictive accuracy is defined as
() =0, where 󰇛󰇜. The DM statistics
are obtained as shown in Eq. (12),
 

󰇛󰇜
where, 
󰇡󰇢

 and 󰇛󰇜 is
aconsistent estimate of󰇛󰇜.
• Wilcoxon-Signed (WS) Rank Test
The WS Rank test examines whether the difference of
forecasting accuracy based on zero-median loss
differential is statistically significant. The null hypothesis
is given by median () = 0. If the out-of-sample loss
distribution is symmetric, both the tests should give
consistent results. WS statistic is given in Eq. (13),
 󰇛󰇜󰇛󰇜

 (13)
where󰇛󰇜󰇥  
.
E. Measures of Prediction Errors
The RMSE, MAE), and MAPE are the three-error
metrics that have been utilized to evaluate the out-sample
forecast of the various prediction’s models. Several past
works have used these metrics to measure the out-sample
efficiency of the training models [41, 42]. In terms of the
real and forecasted prices of crude oil, the error metrics
are shown in Eqs. (14)–(16),
 
󰇛󰇜

 (14)
 


 (15)
 



 (16)
where ,  is the crude oil price, ,  is the
predicted value for the particular time , and  is the
number of observations.
V. EXPERIMENTAL SETUP
The fundamental and technical time-series panel is
transformed using PCA, and the first three principal
components represents that 80% of the original data are
used in this study. The study establishes an LSTM model
with a single layer of 64 nodes trained on a single crude
oil price input to provide a forecasting capability
benchmark. This study investigates how combining the
information from fundamental and technical indicators
along with ARIMA forecasts into hybrid LSTM models
can improve forecasting accuracy. Therefore, the simple
LSTM model trained on the historical crude oil prices is
used as the baseline for evaluating all proposed hybrid
models for the remainder of the analysis. The
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
306

performance of the models is compared by calculating the
RMSE, MAE, and MAPE error measures.
This study proposes two hybrid frameworks to
improve the LSTM predictive accuracy. The first model
is a multi-input LSTM network (LP) model to evaluate
whether information from the chosen explanatory
variables can improve prediction accuracy. It involves
training the LSTM on the time series of crude oil prices
and the PCA transformed series. Three principal
components are chosen to cumulatively represent
approximately 80% of the variance in the original dataset.
The second model builds on the existing model by using
information from the predictions of the ARIMA models
as an additional input to investigate the advantage of
adding statistical forecasts. The ARIMA forecasts are
generated using a moving rolling window of 252 days.
This model is fit every 252 days to obtain one-step-ahead
estimates for the oil prices. Fig. 3 contains a flowchart
illustrating the construction of the hybrid model. The
ensemble learning-based LPA model has two LSTM
models which are trained independently as shown in the
figure. The upper LSTM model is identical to the LP
model, trained on the PCA and crude oil price inputs
while lower LSTM model is trained purely on ARIMA
forecasts. Further, ANN with two layers with 256 and
128 nodes, respectively, is used to learn how to best
combine the input predictions to make the best output
prediction.
Figure 3. Flowchart of proposed final model.
The proposed models are run with each LSTM and
hybrid model repeatedly, trained using inputs taken in
four separate rolling windows of sizes 3, 5, 7, and 11
days. Our choice of window sizes is motivated by the
periods they represent, where three days represent an
intra-weekly period, and 5 and 7 days describe weekly
periods. The 11 days’ window represents a biweekly time
horizon. Each hybrid model was trained twenty-five
times independently, and the average of the forecast over
all the iterations was used to compute the accuracy. We
aim to produce a more reliable and consistent prediction
accuracy by averaging all the iterations.
VI. RESULTS AND DISCUSSIONS
This study evaluates the model predictions’ errors for
the one-day ahead crude oil price forecasts over the out-
of-sample period. The following section details the
comparisons between the performances of the models.
The LSTM model is the baseline for all other proposed
hybrid models.
Table IV shows the results for the LSTM, LP, and
LPA models. The LP and LPA hybrids outperform the
plain LSTM in all the loss functions across all four
window sizes. Table V shows the percentage
improvement in loss function values for LP and LPA
over plain LSTM. An improvement of around 30% is
observed when information from the explanatory
variables, captured by the PCA transformation, is used as
an input into the LSTM to improve the forecast accuracy
of crude oil. Notably, the model using a window size of
11-days outperforms other variations, with improvements
of 35.19%, 37.72%, and 42.10%, for the RMSE, MAE,
and MAPE metrics. Hence, from the results, it is evident
that including explanatory variables that account for
crude oil’s fundamental and technical nature substantially
improves the performance of the predictive neural
network. It is observed that the improvements for the
final hybrid, LPA model are more significant than the LP
counterparts across all window sizes and error metrics.
For LPA, a window size of 11 days is optimal,
showcasing performance improvements of 48.85%,
53.20%, and 50.65% for the RMSE, MAE, and MAPE
metrics. The results support the hypothesis that the
ARIMA predictions contain additional explanatory
information about crude oil price movement, allowing the
ensemble learning-based LPA hybrid models to produce
better forecasts than the plain LSTM and LP models.
TABLE IV. ERROR METRICS FOR EACH PROPOSED MODEL
Model
WIN_SZ
RMSE
MAE
MAPE
L.5TM
3
1.1943
0.94149
0.01673
5
1.19075
0.94127
0.01676
7
1.28716
1.02294
0.01807
11
1.31064
1.04087
0.01836
LP
3
0.90373
0.74632
0.01223
5
0.90982
0.7515
0.01226
7
0.78477
0.61272
0.0101
11
0.77403
0.58636
0.00969
LPA
3
0.70213
0.51716
0.00945
5
0.74154
0.55388
0.01002
7
0.61423
0.44175
0.00822
11
0.61088
0.44062
0.00826
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
307

TABLE V. PERCENTAGE IMPROVEMENT IN LOSS FUNCTION VALUES
FOR LP AND LPA OVER PLAIN LSTM
Model
WIN_SZ
RMSE
MAE
MAPE
LP
3
24.333
20.733
26.923
5
23.823
20.183
26.743
7
34.293
34.923
39.613
11
35.193
37.723
42. 103
LPA
3
41.213
45.073
43.543
5
37.913
41.173
40.093
7
48.573
53.083
50.843
11
48.853
53.203
50.653
Figure 4. Comparison of prediction errors of LPA to the baseline plain
LSTM.
Figure 5. Comparison of prediction errors of LP for baseline plain
LSTM.
Figs. 4 and 5 show the improvement in one-step-ahead
forecast performance considering the 3-day window size
as the benchmark model. Numerous studies have
concluded that crude oil prices are unpredictable using
traditional econometric methods. The best estimate of
future oil prices is the current price itself [43, 44].
However, the relationships between various
macroeconomic, geopolitical, supply, demand, and
technical factors play a crucial role in explaining oil price
behavior [9, 45]. In line with the findings of
Miao et al.’s [7] and Baumeister and Kilian’s [10]
research, this study shows a significant confirmation of
our hypothesis that the information from traditional
forecasting methods further improves price predictions.
Table VI shows the results of the DM and WS tests for
equal forecast accuracy. The null hypothesis of these tests
suggests that the forecasting models have comparable
accuracy, and hence comparisons made between them are
not significant. The p values obtained from the tests,
which involve comparing all model forecasts, are
reported in the table. Results associated with the DM tests
are reported above the diagonal of each table, while those
of the WS tests are reported below the diagonal. At the
95% significance level, all the p values suggest that the
null hypothesis is rejected, indicating that the out-of-
sample forecast accuracy obtained from each model is
significantly different from the other.
TABLE VI. DIEBOLD-MARIANO & WILCOXON SIGNED-RANK TEST
RESULTS
Model
WIN_SZ
L.5TM
LP
LPA
3
L.5TM
0.00
0.00
LP
0.00
0.00
LPA
0.00
0.00
5
L.5TM
0.00
0.00
LP
0.00
0.00
LPA
0.00
0.00
7
L.5TM
0.00
0.00
LP
0.00
0.00
LPA
0.00
0.00
11
L.5TM
0.00
0.00
LP
0.00
0.00
LPA
0.00
0.00
VII. CONCLUSION
Based on experimental results, both the hybrid models
are found to outperform the simple LSTM model. The
experimental result showed 20% improvement in the
forecasting accuracy for all loss functions for the LP
model. The proposed model shown 35.19%, 37.72%, and
42.10% improvement for the RMSE, MAE, and MAPE,
respectively by using a window size of 11. Thus, it is
evident that adding explanatory fundamental and
technical variables helps to improve the forecasting
ability of the neural network. The LPA model performs
well, and provides 40% improvement for all the error
metrics. Models trained on 11 days’ window size are
again found to be the better performing model. Thus,
ARIMA model forecasts add explanatory power to the
framework, improving the forecasting accuracy. Hence,
this study verified that adding information from
fundamental, technical, and statistical methods associated
with distinct economic characteristics of crude oil prices
as inputs to the proposed LSTM base model significantly
improves the one-step-ahead forecasting accuracy of
crude oil prices.
Multiple days ahead forecasts of crude oil price
movements can be a further extension to this study. This
would be pertinent to those with longer investment time
horizons. Fluctuations in crude oil prices affect oil
producers, governments, oil-dependent industries, and
traders. Crude oil also plays a crucial role in the hedging
strategies of manufacturers and investors. Crude prices
being a commodity of international significance are one
of the most important fuel sources and are subject to
periods of extreme volatility. The findings of this study
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
308

may be of relevance to risk management professionals
who want to understand the behaviour associated with the
crude oil market to avoid potential losses. The results
would also aid policymakers concerned with maintaining
commodity market stability.
CONFLICT OF INTEREST
The authors declare no conflict of interest.
AUTHOR CONTRIBUTIONS
In this work, Kshitij Kakade, Kshitish Ghate, and
Ritika Jaiswal have conceptualized and implemented the
idea, while Raj K. Jaiswal has done review, correction
and checked for result correctness and validation.
REFERENCES
[1] R. Rapier. (2021). Fossil fuels still supply 84 percent of world
energy — And other eye openers from BP’s. [Online]. Available:
https://www.forbes.com/sites/rrapier/2020/06/20/bp-review-new-
highs-in-global-energy-consumption-and-carbon-emissions-in-
2019/?sh=1af3546266a1
[2] R. Sari, S. Hammoudeh, and U. Soytas, “Dynamics of oil price,
precious metal prices, and exchange rate,” Energy Economics, vol.
32, no. 2, pp. 351–362, 2010.
[3] Z. Cen and J. Wang, “Crude oil price prediction model with long
short-term memory deep learning based on prior knowledge data
transfer,” Energy, vol. 169, pp. 160–171, 2019.
[4] S. T. Katircioglu, K. Sertoglu, M. Candemir, and M. Mercan, “Oil
price movements and macroeconomic performance: Evidence
from twenty-six OECD countries,” Renewable and Sustainable
Energy Reviews, vol. 44, pp. 257–270, 2015.
[5] J. C. Reboredo and A. Ugolini, “Quantile dependence of oil price
movements and stock returns,” Energy Economics, vol. 54, pp.
33–49, 2016.
[6] Y. J. Zhang and Y. M. Wei, “The dynamic influence of advanced
stock market risk on international crude oil returns: An empirical
analysis,” Quantitative Finance, vol. 11, no. 7, pp. 967–978, 2011.
[7] H. Miao, S. Ramchander, T. Wang, and D. Yang, “Influential
factors in crude oil price forecasting,” Energy Economics, vol. 68,
pp. 77–88, 2017.
[8] T. Yao and Y. Zhang, “Forecasting crude oil prices with the
Google index,” Energy Procedia, vol. 105, pp. 3772–3776, 2017.
[9] S. Dées, A. Gasteuil, R. Kaufmann, and M. Mann, “Assessing the
factors behind oil price changes,” European Central Bank,
Working Paper Series, no. 855, January 2008.
[10] C. Baumeister and L. Kilian, “Forecasting the real price of oil in a
changing world: A forecast combinationapproach,” Journal of
Business & Economic Statistics, vol. 33, no. 3, pp. 338–351, 2015.
[11] L. C. Hsu, “Investment decision making using a combined factor
analysis and entropy-based topsis model,” Journal of Business
Economics and Management, vol. 14, no. 3, pp. 448–466, 2013.
[12] S. P. Das, N. S. Achary, and S. Padhy, “Novel hybrid SVMTLBO
forecasting model incorporating dimensionality reduction
techniques,” Applied Intelligence, vol. 45, no. 4, pp. 1148–1165,
2016.
[13] L. Yin and Q. Yang, “Predicting the oil prices: Do technical
indicators help?” Energy Economics, vol. 56, pp. 338–350, 2016.
[14] L. Liu, Y. Wang, and L. Yang, “Predictability of crude oil prices:
An investor perspective,” Energy Economics, vol. 75, pp. 193–205,
2018.
[15] A. Murat and E. Tokat, “Forecasting oil price movements with
crack spread futures,” Energy Economics, vol. 31 no. 1, pp. 85–90,
2009.
[16] Y. Wei, Y. Wang, and D. Huang, “Forecasting crude oil market
volatility: Further evidence using GARCH-class models,” Energy
Economics, vol. 32, no. 6, pp. 1477–1484, 2010.
[17] Y. Xiang and X. H. Zhuang, “Application of ARIMA model in
short-term prediction of international crude oil price,” in
Advanced Materials Research, Trans Tech Publications Ltd., 2013,
vol. 798, pp. 979–982.
[18] N. M. Yusof, R. S. A. Rashid, and Z. Mohamed, “Malaysia crude
oil production estimation: An application of ARIMA model,” in
Proc. 2010 International Conference on Science and Social
Research, 2010, pp. 1255–1259.
[19] H. Mohammadi and L. Su, “International evidence on crude oil
price dynamics: Applications of ARIMA-GARCHmodels,”
Energy Economics, vol. 32, no. 5, pp. 1001–1008, 2010.
[20] W. Kusonkhum, K. Srinavin, N. Leungbootnak, P. Aksorn, and T.
Chaitongrat, “Government construction project budget prediction
using machine learning,” Journal of Advances in Information
Technology, vol. 13, no. 1, pp. 29–35, Feb. 2022.
[21] A. Alshammari, R. Almalki, and R. Alshammari, “Developing a
predictive model of predicting appointment no-show by using
machine learning algorithms,” Journal of Advances in Information
Technology, vol. 12, no. 3, pp. 234–239, 2021.
[22] W. Jinxiao, S. Jiaxin, H. Dexin, and Z. Xiaoyu, “Internet financial
news and prediction for stock market: An empirical analysis of
tourism plate based on LDA and SVM,” Journal of Advances in
Information Technology, vol. 10, no. 3, pp. 95–99, 2019.
[23] T. Daghistani and R. Alshammari, “Comparison of statistical
logistic regression and random forest machine learning techniques
in predicting diabetes,” Journal of Advances in Information
Technology, vol. 11, no.2, pp. 78–83, 2020.
[24] N. A. Bakshi, P. R. Kolan, and B. Behera, N. Kaushik, and A. M.
Ismail, “Predicting pregnant shoppers based on purchase history
using deep convolutional neural networks,” Journal of Advances
in Information Technology, vol. 9, no. 4, pp. 110–116, 2018.
[25] S. Hochreiter and J. Schmidhuber, “Long short-term memory,”
Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[26] T. Kubra, B. Sekeroglu, and C. Ozkan, “Lung cancer incidence
prediction using machine learning algorithms,” Journal of
Advances in Information Technology, vol. 11, no. 2, pp. 91–96,
2020.
[27] J. Premsmith and H. Ketmaneechairat, “A Predictive model for
heart disease detection using datamining techniques,” Journal of
Advances in Information Technology, vol. 12, no. 1, pp. 14–20,
2021.
[28] V. Dropsy, “Do macroeconomic factors help in predicting
international equity risk premia?: Testing the out-of-sample
accuracy of linear and nonlinear forecasts,” Journal of Applied
Business Research, vol. 12, no. 3, p. 120, 2011.
[29] H. Chiroma, S. Abdulkareem, and T. Herawan, “Evolutionary
neural network model for west Texas intermediate crude oil price
prediction,” Applied Energy, vol. 142, pp. 266–273, 2015.
[30] Y. X. Wu, Q. B. Wu, and J. Q. Zhu, “Improved EEMD-based
crude oil price forecasting using LSTM networks,” Physica A:
Statistical Mechanics and its Applications, vol. 516, pp. 114–124,
2019.
[31] N. Suhermi, D. Suhartono, D. Prastyo, and B. Ali, “Roll motion
prediction using a hybrid deep learning and ARIMA model,”
Procedia Computer Science, vol. 144, pp. 251–258, 2018.
[32] Y. S. Wang and Y. L. Chueh, “Dynamic transmission effects
between the interest rate, the US dollar, and gold and crude oil
prices,” Economic Modelling, vol. 30, pp. 792–798, 2013.
[33] A. Villar and F. L. Joutz, “The relationship between crude oil and
natural gas prices,” Energy Information Administration, Office of
Oil and Gas, vol. 1, pp. 1–43, 2006.
[34] J. A. Batten, C. Ciner, and B. M. Lucey, “The dynamic linkages
between crude oil and natural gas markets,” Energy Economics,
vol. 62, pp. 155–170, 2017.
[35] R. Novy-Marx, “Is momentum really momentum?” Journal of
Financial Economics, vol. 103, no. 3, pp. 429–453, 2012.
[36] C. J. Neely, D. E. Rapach, J. Tu, and G. Zhou, “Forecasting the
equity risk premium: The role of technical indicators,”
Management Science, vol. 60, no. 7, pp. 1772–1791, 2014.
[37] S. Demir, K. Mincev, K. Kok, and N. G. Paterakis, “Introducing
technical indicators to electricity price forecasting: A feature
engineering study for linear, ensemble, and deep machine learning
models,” Applied Sciences, vol. 10, no. 1, p. 255, 2019.
[38] S. Wold, K. Esbensen, and P. Geladi, “Principal component
analysis,” Chemometrics Intell. Lab. Syst., vol. 2, no. 1–3, pp. 37–
52, 1987.
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
309

[39] G. E. P. Box and G. M. Jenkins, Time Series Analysis Forecasting
and Control, 2nd ed. San Francisco: Holden-Day, 1976.
[40] F. X. Diebold and R. S. Mariano, “Comparing predictive
accuracy,” Journal of Business Economic Statistics, vol. 20, no. 1,
pp. 134–144, 2002.
[41] A. Fuertes, M. Izzeldin, and E. Kalotychou, “On forecasting daily
stock volatility: The role of intraday information and market
conditions,” International Journal of Forecasting, vol. 25, no. 2,
pp. 259–281, 2009.
[42] H. Y. Kim and C. H. Won, “Forecasting the volatility of stock
price index: A hybrid model integrating LSTM with multiple
GARCH-type models,” Expert Systems with Applications, vol. 103,
pp. 25–37, 2018.
[43] J. D. Hamilton, “Understanding crude oil prices,” The Energy
Journal, vol. 30, no. 2, 2009.
[44] R. Alquist, L. Kilian, and R. J. Vigfusson, “Forecasting the price
of oil,” Handbook of Economic Forecasting, vol. 2, pp. 427–507,
2013.
[45] R. K. Kaufmann, S. Dees, A. Gasteuil, and M. Mann, “Oil prices:
The role of refinery utilization, futures markets and non-
linearities,” Energy Economics, vol. 30 no. 5, pp. 2609–2622,
2008.
Copyright © 2023 by the authors. This is an open access article
distributed under the Creative Commons Attribution License (CC BY-
NC-ND 4.0), which permits use, distribution and reproduction in any
medium, provided that the article is properly cited, the use is non-
commercial and no modifications or adaptations are made.
Kshitij Kakade completed his undergraduate with
a master in economics and a bachelor in computer
science from BITS Pilani, K.K. Birla Goa Campus.
His interests are in the areas of applications of ML
in finance.
Kshitish Ghate is a final-year undergraduate
student pursuing a master in economics and a
bachelor in computer science from BITS Pilani,
K.K. Birla Goa Campus. His interests are in the
areas of machine learning, natural language
processing, financial markets, and behavioral
economics with a focus on applications of AI for
social good
.
Raj K. Jaiswal is working as an assistant professor
at SCSDF, National Forensic Sciences University
Goa. He has received a Ph.D. degree from the
Department of Information Technology, National
Institute of Technology Karnataka, Surathkal, India.
His research areas of interest include vehicular ad
hoc network, SDN/NFV, and cyber security.
Ritika Jaiswal is working as an assistant professor
at the Department of Economics & Finance, BITS
Pilani, KK Birla Goa Campus. She has done her
Ph.D. in the area of commodity derivatives
(finance) form the School of Management, National
Institute of Technology Karnataka, Surathkal, India.
Her research interests include financial economics,
market microstructure, risk management and energy financ
e.
Journal of Advances in Information Technology, Vol. 14, No. 2, 2023
310