IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 2, June 2023, pp. 892~911

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i2.pp892-911  892

Hybrid Forex prediction model using multiple regression,

simulated annealing, reinforcement learning and technical
analysis

Hana Jamali1, Younes Chihab2, Iván García-Magariño3, Omar Bencharef1

1
Computer and System Engineering Laboratory, Faculty of Science and Technology, Cadi Ayyad University, Marrakesh, Morocco
2
Computer Sciences Laboratory, Superior School of Technology, Ibn Tofail University, Kenitra, Morocco
3
Department of Software Engineering and Artificial Intelligence, Complutense University, Madrid, Spain

Article Info ABSTRACT

Article history: Foreign exchange market refers to the market in which currencies from around
the world are traded. It allows investors to buy or sell a currency of their
Received Jul 7, 2022 choice. Forex interests several categories of stakeholders, such as companies
Revised Nov 30, 2022 that carry out international contracts, large institutional investors, via the main
Accepted Dec 12, 2022 banks, which carry out transactions on this market for speculative purposes.
One of the most important aspects in the Forex market is knowing when to
invest by buying, selling, and this through the recorded trend of a currency
Keywords: pair, but given the characteristics of the Forex market namely its chaotic, noisy
and not stationary nature, prediction becomes a big challenge for traders when
Forex it comes to predicting accuracy. This paper aims to predict the right action to
Prediction be taken at a certain moment through the development of a model that
Regression algorithm simulated combines multiple techniques such multiple regression, simulated annealing
annealing meta-heuristics, reinforcement learning and technical indicators.
Reinforcement learning
This is an open access article under the CC BY-SA license.
Technical indicators

Corresponding Author:
Hana Jamali
Computer and System Engineering Laboratory, Faculty of Science and Technology, Cadi Ayyad University
Marrakesh, Morocco
Email: [email protected]

1. INTRODUCTION
Currency is an indispensable tool in the economy of the countries. It derives its importance from
being: i) used as a reserve currency by central banks, ii) used as an invoicing currency for international
transactions, and iii) used as a vehicle currency for currency transactions. Vehicle currencies are used when
the transaction cost of trading two currencies directly is higher than the cost of trading them indirectly, via two
transactions through the vehicle currency [1].
Forex is the market in which so-called convertible currencies (currency pairs) are traded against each
other at constantly changing exchange rates, allowing it to be the most liquid market and the largest after the
debt. This global market has no borders and is physically located in no particular place. At any time, an identical
transaction can be carried out simultaneously in Paris, London, Zurich, Hong Kong or New York. Indeed,
trading volumes on Forex have been growing steadily since the 2000s.
The daily trading volume in 2019 on Forex reached 5,345 billion dollars, a 24% increase in volumes
compared to the daily transaction volume recorded, in 2013 according to the latest figures published by the
Banks of International Settlements (BIS). By comparison, France's gross domestic product (GDP) was $2,716
billion and Japan's $5,065 billion in 2019. It has traded in a day on Forex more than the annual GDP of Japan.
This makes it by far the biggest financial market in the world, far ahead of the equity markets (tiny compared
to Forex).

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  893

Forex is an over-the-counter market where the transaction is concluded directly between two
counterparties, the seller and the buyer. In other words, the transaction of purchase or sale of the currency is
made in a consensual manner between the seller and the buyer of the currency, it is an action in which only the
free will is considered. Characterized by intermediation and significant bid-ask spreads/intermediation fees [2],
[3]. Given its large transactions, Forex becomes one of the most followed markets by economic players and
institutional operators and individuals, this feature makes it a market sensitive to any economic influence, such
as budget deficits of western nations, their recovery plans, and the monetary policies of the central banks of
the world's largest powers.
Previously, when the internet was not accessible to all, only the major players in the market, such as
international banks and major financial centers, could participate in trading and currency speculation. By cons
currently, following the technological development and internet revolution, speculation and Forex trading can
be done by any major person able to trade. Trading in the Forex market requires huge capital investments. In
fact, it depends on the trading strategy you choose. So, in Forex trading, it is possible to invest sums greater
than the initial deposit thanks to the leverage, but the losses cannot exceed the balance of the account, Forex
brokers lend money to traders for that they can increase their investments in the foreign exchange market.
Leverage is there for essentially about borrowing money from a broker to increase the return on investment.
However, it also means that the trader increases the risk of losing more money. Leverage is partly responsible
for the popularity of Forex trading, as traders can easily choose the risk they want to take with a reduced capital
contribution.
The most popular currencies in the world and the most traded in Forex are the US dollar (USD), the
euro (EUR), the pound sterling (GBP), the Japanese yen (YEN) and the Swiss Franc (CHF). These currencies
make up a group of major currency pairs containing EURUSD, GBPUSD, USDJPY, USDCHF, GBPJPY,
EURGBP, EURCHF [3]; Figure 1 represents the exchange rates of the major currencies from Tuesday
February, 05 2019 at 15:13 UTC (universal time coordinated). These major pairs are characterized by lower
spreads and higher trading volume relative to other currencies.
In order to put the odds on their side, traders need to accurately predict the exchange rate at any given
time. Making a short term Forex market forecast is not an easy thing, even winning in Forex is not available
every time, but the trader can use a short term investment strategy to improve these trading decisions. In this
article, we will adopt the following plan, in the first section we will make a state of the art on the most used
techniques and investment strategies on Forex, the second section will be used to propose a short-term
investment strategy term on Forex, then the third section will be dedicated to drawing a general conclusion and
discussing the prospects envisaged.

Figure 1. The exchange rates of the major currencies from tuesday February 05 th, 2019 at 15:13 UTC

2. STATE OF THE ART

As the Forex market is a dynamic environment and sensitive to any event that can disrupt its stability
in terms of sudden changes in currency exchange rates, speculation is considered a risky operation and can
contain unexpected surprises, therefore, the prediction represents an essential tool to have an overview of the
market trend during the following hours or days; because it can allow knowing when and how the trader should
react to avoid losing money and to maximizing profits of his investment in Forex. In the literature, there is a
large bibliography used different techniques and methods [4]–[6] shows success. The model proposed in this
paper is characterized by using various contributing techniques in the aim to obtain a good prediction accuracy
of currency exchange rates, the added value of this model consists in combining several types of machine
learning (supervised learning and reinforcement learning (RL)) [7], [8]. Optimization method [9] and also
technical analysis [10], [11] through the relative strength index (RSI) indicator. The proposed model will not
only help to make an investment decision but also to calculate the investment profit of a specific capital. Hence,
be gone this section provides an overview of bibliographic literature related to currency exchange rate
prediction and Forex investment.

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
894  ISSN: 2252-8938

2.1. Investing in Forex

Making decisions in business and finances is always facing the problem of risk. Reliability and risk
categories frequently are invoked together, while analyzing processes, system conditions and development
sustainability. However, often these categories are conceived as opposite to each other: when risk increases,
reliability possibility decreases, and conversely-when risk decreases, reliability possibilities increase [12].
Although trading in the Forex market is a risky business, therefore, an investment in this market must first
begin with a market study, this study must include three essentials elements: money management, fundamental
analysis, and technical analysis.
The first element, money management helps to determine investment risks in advance, develop and
improve discipline and take trading to the next level, which helps protect the investment portfolio to avoid
investment risks, and then it refers to the ability to manage earnings and trading so as not to take risks outside
the investment strategy. The second element to do a market study is fundamental analysis, this element is based
on five major factors which have a direct influence on the exchange rate of currencies, and which can help to
anticipate the Forex trend and make a fair enough prediction, these major factors are: i) economic growth, ii)
geopolitics or political stability, iii) monetary policy, iv) imports and exports, and iv) interest rates. It is very
important to consult the economic and political calendar to be able to establish a global vision on the
fundamental factors which influence the economic market and consequently the currency pair to invest. The
economic calendar is based on GDP, economic growth and the inflation rate of the investment countries,
because if there is good economic growth we can anticipate an increase in the exchange rate, and vice versa.
In addition the interest rates of a country is the most important fundamental factor in determining the exchange
rate of a currency because when a country increases its interest rate, consequently the investors are directed
toward the appreciation of the currency and thereafter the investment in this currency, suddenly they will buy
it more. Therefore, fundamental or macro-economic analysis of the Forex market is used as a tool to predict
prices in the medium and long term [13] which is not the case in our study and it will not be used in the proposed
model. It also makes it possible to assess currency prices and anticipate future market trends by relying on
economic publications from countries around the world. Indeed, several times a day, economic data is
published by financial institutions, governments, central banks or even certain private organizations, sometimes
having significant impacts on the evolution of financial market prices.
The third element: technical analysis is appeared in Japan during the 16th century, it aimed to forecast
the evolution of the price of rice, but the rise of technical analysis was reinforced in the 1970s with the
appearance of indicators. Techniques and computer development, which makes modeling more and more
sophisticated. In the case of Forex, technical analysis is used to examine changes in the market by analyzing
signals and then interpreting those signals to be able to open, close or modify a buy or sell transaction.
Technical analysis is a method that tries to predict Forex movements by looking at historical market
data including prices, volumes traded and open interest and identify trends and significant price levels with a
high probability of rebounds such as media and resistances. The technical analysis is essentially based on the
fact that certain market configurations are cyclical and that price action is repeated over time. Technical
analysts and traders use technical and mathematical indicators to make their investment decisions. These
indicators are viewed in real-time on graphs that are interpreted to identify buying or selling opportunities.
In contrast to fundamental analysis that examines economic factors and long-term trends, technical
analysis is the method most used by Forex traders to predict short-term currency price movements such as
weekly, daily or even hourly prediction [13]. Technical analysis can help you identify the trend; identify the
strength and stability of the trend over time. Technical analysis can increase discipline and decrease the
influence of emotions in your trading plan. While no system can guarantee you can identify the Forex trend
100%, technical analysis can help you create a trading plan and follow it more objectively. Forecasting in Forex
is a statistical process to help informing decisions on currency rate prediction planning and therefore short,
medium and long term investment in Forex.
Exchange rate forecasting has a direct effect on the rates themselves. If there is a well-publicized
forecast informing to rise exchange rate, people (investors) will immediately adjust the price they are willing
to pay and the forecast will come true on its own. In a sense, exchange rates become their own predictions.
Predicting whether the exchange rate will rise or fall tomorrow is about as predictable as predicting whether a
coin toss will fall as a head or tail. Either way, you'll be right about 50% of the time, regardless of your
prediction. It is said that predictions are not possible in a changing environment. Every environment changes
and a good forecasting model capture how things will change.
Two kinds of methods are used to make currency exchange rate predictions, and subsequently have
good speculation in the Forex market, the first kind is qualitative method, used in the case where there is no
data available or if existing data is not relevant for forecasting. The second type is the so-called quantitative
methods, used to exploit historical digital information (collected at regular intervals in time) while assuming

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  895

that the behavior of past achievements will recur in the future. In this study the method adopted is the
quantitative method, because we have a database containing the exchange rates of the EUR/USD pair.

2.2. Related work

Many papers in the literature that propose various methodologies and techniques [10], [14], [15] for
modeling the prediction of exchange rates in the Forex market, Yong et al. [13] studied the effects of different
types of inputs including: The close price as well as various technical indicators derived from the close price
are studied to determine its effects on the Forex trend predicted by an intelligent machine learning module and
it has been found that the type of input data used for Forex price prediction is a crucial element that cannot be
taken lightly. This means that the incorporation of trading rules and technical analysis as performed by the
technical analyst in the initial phase of the forecasting algorithm will help increase the accuracy of Forex price
forecasting. Samad et al. [16] uses financial information and historical price data, the problem of stock price
prediction is modeled by data mining with machine learning algorithms, namely the support vector machine
and random forest. Based on the results produced, the analysis gained more than 60% accuracy for textual
analysis (financial information) and 90% for numerical analysis (historical price data). Fattah et al. [17] used
deep neural network (DNN) to predict whether the closing price is reached at the profit which is determined
by the investor or not and improve the accuracy of the prediction. Particle swarm optimization (PSO) and
machine learning (AutoML) are used as optimizers with DNNs. Based on the experimental results, deep
learning AutoML was found to have the best accuracy rate, which ranges from 81% to 92% across all
companies, and accuracy after DNN optimization using PSO varies from 73% to 82% in all companies.
Rassetiadi and Suharjito [18] examine the external factors that can influence forecast results, looking for the
relationship between the value of indices such as S and P500 (standard and poor's) and the value of commodities
such as gold and silver in the process EUR/USD prediction. When comparing the mean squared error (MSE)
values, it turned out that the best combination was a combination of the FTSE100 (financial times stock
exchange) and natural gas values. Meng and Khushi [19] reviewed all recent stock/Forex prediction or trading
articles that used RL as their primary machine learning method and conclude that RL in stock/Forex trading is
still in its early development and further research is needed to make it a reliable method in this domain.
Carapuço et al. [20] developed a new system for short-term speculation in the foreign exchange market Forex,
based on recent RL developments and concluded that learning in the training dataset is stable and it is apparent
from a number of validations learning curves that the Q-network is indeed capable of finding relationships in
financial data that translate to out-of-sample decision making. It was also shown that the model obtained from
the training procedure can then be harnessed for profitable trading in atest dataset. Rundo [8] proposes in this
work the use of an algorithm based both on supervised deep learning and on a RL algorithm for forecasting the
short-term trend in the currency Forex market to maximize the return on investment in an high-frequency
trading (HFT) algorithm. The trading system has been validated over several financial years and on the
EUR/USD cross confirming the high performance in terms of return of investment (98.23%) in addition to a
reduced drawdown (15.97%) which confirms its financial sustainability.

3. PROPOSED SYSTEM
Following the most of conclusions reached from the bibliographic research, the combination of two
or more techniques can show better results [8], [9], [21] in terms of the exchange rate prediction accuracy, the
proposed model in this paper combines more than one method to predict exchange rates:
− Multiple regression for its ability to determine the relative influence of one or more predictor variables on
the value of the criterion.
− Simulated annealing to avoid getting locking in a local optimum.
− In order to structure efficient trading systems, RL has enormous capacities in terms of decision-making
using an intelligent agent.
− The RSI technical indicator assesses overbought or oversold conditions of the exchange rate.

3.1. The dataset

The preparation of the data used in our research went through three main processes, the first is the collection
of data from a specific time period, then the second process is data refinement, which is to eliminate unnecessary
data for our research, finally the last process is the calculation of new variables from existing data. These steps are
necessary to have good prediction results. This allows the trader to make correct investment decisions.

3.1.1. Data collection

The data used in this paper were collected on the site https://eatradingacademy.com [22]. These data
represent the daily exchange rates of the EUR/USD pair recorded between 10/25/2016 and 09/27/2020, so

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
896  ISSN: 2252-8938

almost 1,227 records, we chose the EUR/USD pair because it has around 25% trading volume in Forex, which
makes it ideal for short term prediction. The dataset contains the daily values of the open, close, high and low
prices. The period 2016-2020 was chosen because it was characterized by several world events having
influenced the Forex market, in particular, the election of Trump as president of the United States of America
for the presidential term 2017-2021 with all with all his decisions that havig marked the Americain economy
and consequently the world economy, thus this period was marked by the health crisis linked to the coronavirus
disease 2019 (COVID 19) pandemic with the restrictions that this entails, such as the contraction of the
economy and per capita income, which will tilt millions of people into extreme poverty. The dataset used is
separated into two parts, the training set 80% and the testing set 20%, The literature review revealed that the
testing or the validation set should be approximately from one fourth to one eight of the training set [23].

3.1.2. Data refinement

The dataset used in this research consists of eight different variables that represent closing exchange
rates over a one-month time frame. Figure 2 represents the eight variables making up the data set: the closing
exchange rate for the current day D, day D-1, day D-2, day D-3, day D- 4, of day D-5, day D-6, the average
exchange rate for the last two weeks proceeding the current day D and the average value of the last month
preceding day D. These variables are used to calculate other variables that are going to be introduced into the
dataset to help improve prediction accuracy.

Figure 2. Example of used dataset

3.1.3. Calculation of variables

Since the use of technical analysis in Forex speculation has had good results [13], we have introduced
the normalized RSI technical indicator as a variable in our dataset to design the proposed prediction model.
The choice of the RSI indicator is simply because it allows traders to quickly see if the market is overbought
or oversold, making it easy to interpret. Because of the results obtained by the literature review made on the
prediction techniques used in Forex to improve the prediction of exchange rates and to have a fairly reliable
accuracy, we propose an investing strategy in the trading market. The proposed strategy Figure 3 is composed
of several blocks which are combined to result in a better prediction of the exchange rates of the EUR/USD
pair. The first block is for collecting historical exchange rate data, then preparing the data through refining and
removing unnecessary data, then dividing the database into two parts, the training set and testing and validation
set with a percentage of 80% and 20% successively. The second block is used for the implementation of
supervised learning through the multiple regression algorithm to predict the exchange rates from day D until
day D+14. The third block was used to optimize the results obtained by the multiple regression algorithm to
achieve a better fit between the actual values of the exchange rates of the EUR/USD pair and the predicted
values obtained by the multiple regression. This optimization is made by the simulated annealing metaheuristic
and this through the use of the results obtained by the multiple regression as input variables for the simulated
annealing metaheuristic, while specifying the objective function, which is in this case the mean squared error
(MSE) calculated between the real values and the predicted values obtained by the multiple regression,
therefore the optimization is successful when the MSE arrives at a minimum point, in this case, the values
predicted by the multiple regression will be replaced by optimized values. The fourth block used RL to train
the model through the exploration of the environment which is in this case the values of the exchange rates of
the EUR/USD pair. The agent will try to maximize the rewards for achieving a good result. First, it is mandatory
to initialize the state s (t) then choose at random an action a (t) and then generate an action s (t+1). The fifth
block is used to exploit the optimized predicted values obtained by multiple regression and simulated
annealing, using RL to arrive at initial Forex investment decisions. The sixth block calculates the normalized
RSI technical indicator which will be used in combination with the results of block five to improve the accuracy
of prediction and have a final investment decision.

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  897

Figure 3. Flow diagram of the proposedmodel for rate exchange prediction

4. MODEL TECHNIQUES
4.1. Multiple regression
Multiple regressions are a family of statistics used to investigate the relationship between a set of
predictors and a criterion (dependent) variable. This procedure is applicable in a variety of research contexts
and data structures [10], [24]–[27]. This is the case in the model proposed by this study. The implementation
of the regression algorithm in this model is started by the use of the values of the exchange rates of days D,
D-1, D-2, D-3, D-4, D-5, D-6, the average of last two weeks and the average of the last month, as input values.
One of the advantages of multiple regression is the ability to determine the relative influence of one
or more predictor variables on the value of the criterion, also its simplicity of interpretation and its ease of
calculation. Figures 4 to12 illustrate a shift between the predicted values and the real values of exchange rates,
this shift widens as the prediction day moves away from day D, this shift is clearly represented by the
calculation of the coefficient of determination R² which makes it possible to measure the adequacy between
the real values and the predicted values. Figure 13 represents the coefficient of determination calculated for
the predicted days. The prediction of the exchange rate values using multiple regression method shows a high
accuracy of more than 99% for day D, but on the other hand, this precision decreases for day D+1 until day
D+14. In this case, we propose to integrate an optimization solution such as simulated annealing meta-heuristic
to adjust the predicted values with the real values.

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
898  ISSN: 2252-8938

Figure 4. Comparison of the test set predictions of day D got by multiple regression (red line) with real
output values (blue line)

Figure 5. Comparison of the test set predictions of day D+1 got by multiple regression (red line) with real
output values (blue line)

Figure 6. Comparison of the test set predictions of day D+2 got by multiple regression (red line) with real
output values (blue line)

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  899

Figure 7. Comparison of the test set predictions of day D+3 got by multiple regression (red line) with real
output values (blue line)

Figure 8. Comparison of the test set predictions of day D+4 got by multiple regression (red line) with real
output values (blue line)

Figure 9. Comparison of the test set predictions of day D+5 got by multiple regression (red line) with real
output values (blue line)

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
900  ISSN: 2252-8938

Figure 10. Comparison of the test set predictions of day D+6 got by multiple regression (red line) with real
output values (blue line)

Figure 11. Comparison of the test set predictions of day D+7 got by multiple regression (red line) with real
output values (blue line)

Figure 12. Comparison of the test set predictions of day D+14 got by multiple regression (red line) with real
output values (blue line)

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  901

Figure 13. Multiple R-squared calculated between real values and predicted values

4.2. Simulated annealing

Simulated annealing is a metaheuristic method used more in industry and which is based on heating
two metals and then cooling them slowly or rapidly, which represent completely different properties [28]. So
it is intended to solve optimization problems, therefore, it is used in this model to optimize the currency
prediction results of the EUR/USD pair obtained by using the multiple regression algorithm, the principle of
use was to take the outputs of the EUR/USD exchange rate prediction results obtained by applying the multiple
regression algorithm as input variables for the simulated annealing, certainly with other parameters such as
temperature and limit threshold , to establish an optimization algorithm Figure14 that will allow the predicted
values to be adjusted with the actual values. Figures 15 to 23 illustrate the results obtained by integrating the
optimization meta-heuristic.
The results of the integration of the simulated annealing meta-heuristic Table 1, showed a prediction
improvement compared to the results obtained by the use of the multiple regression algorithm, the smaller
errors of prediction MSE in Table 1 is better. In other words, the use of several techniques shows more
efficiency. Precision compared to the use of a single technique [9]. Hence, we thought of introducing another
type of learning, which is RL to predict an action (decision) to be taken (buy, sell, hold), using results obtained
by multiple regression and simulated annealing.

Figure 14. Simulated annealing algorithm

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
902  ISSN: 2252-8938

Figure 15. Comparison of the test set optimized predictions of day D got by Simulated annealing (red line)
with real output values (blue line)

Figure 16. Comparison of the test set optimized predictions of day D+1 got by Simulated annealing (red line)
with real output values (blue line)

Figure 17. Comparison of the test set optimized predictions of day D+2 got by simulated annealing (red line)
with real output values (blue line)

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  903

Figure 18. Comparison of the test set optimized predictions of day D+3 got by Simulated annealing (red line)
with real output values (blue line)

Figure 19. Comparison of the test set optimized predictions of day D+4 got by Simulated annealing (red line)
with real output values (blue line)

Figure 20. Comparison of the test set optimized predictions of day D+5 got by Simulated annealing (red line)
with real output values (blue line)

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
904  ISSN: 2252-8938

Figure 21. Comparison of the test set optimized predictions of day D+6 got by Simulated annealing (red line)
with real output values (blue line)

Figure 22. Comparison of the test set optimized predictions of day D+7 got by Simulated annealing (red line)
with real output values (blue line)

Figure 23. Comparison of the test set optimized predictions of day D+14 got by Simulated annealing (red
line) with real output values (blue line)

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  905

Table 1. Comparaison of MSE obtained by using multiple regression and simulated annealing
Day MSE
With multiple regression With multiple regression and simulated annealing
D 0.0000204076 0.00002029581
D+1 0.00004046837 0.00004036597
D+2 0.00006178088 0.0000616916
D+3 0.00008198048 0.00008193095
D+4 0.00008193095 0.0001014745
D+5 0.0001205481 0.0001205404
D+6 0.0001393285 0.0001390773
D+7 0.0001567872 0.0001561272
D+14 0.000254101 0.0002530860

4.3. Reinforcement learning

Scientific research in the field of artificial intelligence is mainly devoted to the development of
programs to solve complex problems without the intervention of human beings, these programs are designed
using learning algorithms that extract lessons from repeated experiments by proceeding through the notions of
test and error, as in the case of RL, which allows an autonomous agent to learn the actions to take, from
experiments, in order to optimize a quantitative reward over time. This makes it possible to develop intelligent
solutions to solve complex management problems [8], [19], [20]. RL is, after supervised and unsupervised
learning, the third possibility for teaching algorithms to make decisions for themselves. In other words, develop
intelligent solutions to solve complex management problems. The major advantage of RL is that it does not
need any data for conditioning, unlike supervised learning which must be data fed beforehand. Either RL allows
the machine to learn through interactions with the environment and to apply what it has learned to solve
complex problems in whether the human being has to enter data or intervene in one way or another. Noting
that the RL models are inspired by the behaviorist approach to psychology and learning. At the turn of the 20th
century, American psychologists like Watson, Thorndike, Skinner and Pavlov rejected the study of mental
processes from an introspective approach [29]. They state that these cannot be studied objectively and that only
measurable and observable behaviors can and should be. Their work on learning sheds light on the links
between a situation (or stimulus), a behavioral response and its consequence.

Stimulus⟶Response⟶Consequence

Thorndike postulates the existence of learning laws. On one hand,the law of exercise states that the
link between stimulus and response is strengthened by exercise and that the probability of response increases
with the number of trials performed. On the other hand, the law of effect states that the link between the
stimulus and the response is strengthened or weakened by the immediate effect of its consequences.
RL is different from supervised learning because it does not need inputs and outputs to learn, it is the
agent who decides what to do in a specific situation. so, in the proposed model, the environment is the Forex
market, the stocks are buy, sell and hold, the state is represented by the exchange rates predicted by the
combination of the multiple regression algorithm and the metaheuristic of simulated annealing (outputs from
block 3). Figures 24 to 32 represent the initial decisions obtained based on the predictions of the exchange rates
for days D, D+1, D+2, D+3, D+4, D+5, D+6, D+ 7, D+14. Also the model proposes to calculate the total gains
and the percentage of investment according to the predictions obtained, Table 2 and Table 3 present the results
for an investment of €10,000 and €50,000.

Figure 24. Day D buying signal and selling signal using reinforcement learning

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
906  ISSN: 2252-8938

Figure 25. Day D+1 buying signal and selling signal using reinforcement learning

Figure 26. Day D+2 buying signal and selling signalusing reinforcement learning

Figure 27. Day D+3 buying signal and selling signal using reinforcement learning

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  907

Figure 28. Day D+4 buying signal and selling signal using reinforcement learning

Figure 29. Day D+5 buying signal and selling signal using reinforcement learning

Figure 30. Day D+6 buying signal and selling signal using reinforcement learning

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
908  ISSN: 2252-8938

Figure 31. Day D+7 buying signal and selling signal using reinforcement learning

Figure 32. Day D+14 buying signal and selling signal using reinforcement learning

Table 2. Total gains and the percentage of investment of 10,000€

Day Investment money Total gains Investment percentage
D 10,000 2.320307 0.023203%
D+1 10,000 2.265305 0.022653%
D+2 10,000 1.285376 0.012854%
D+3 10,000 1.440015 0.014400%
D+4 10,000 1.833289 0.018333%
D+5 10,000 1.461165 0.014612%
D+6 10,000 1.093787 0.010938%
D+7 10,000 1.528127 0.015281%
D+14 10,000 1.672714 0.016727%

Table3. Total gains and the percentage of investment of 50,000€

Day Investment money Total gains Investment percentage
D 50,000 1.913611 0.003827%
D+1 50,000 0.589392 0.001179%
D+2 50,000 1.131534 0.002263%
D+3 50,000 1.379849 0.002760%
D+4 50,000 1.949362 0.003899%
D+5 50,000 1.647245 0.003294%
D+6 50,000 1.297847 0.002596%
D+7 50,000 2.578427 0.005157%
D+14 50,000 1.598468 0.003197%

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  909

4.4. RSI indicator

The RSI indicator (for relative strength index) is an indicator used by traders to measure the magnitude
of price changes and highlight areas of overbought or oversold exchange rates. In addition, the graphical
representation of the RSI makes it possible to distinguish three situations Figure 33:
− An extreme band from 0 to 30 symbolizing an oversold zone.
− The zone of neutrality with a low median of 30 to 50 and a high median of 50 to 70.
− An extreme band from 70 to 100 symbolizing an overbought zone.
When the RSI indicator indicates an overbought situation, we can therefore go short. When the
indicator indicates an oversold market, one can go long. The idea underlying this strategy is as follows. An
overbought market is ready to enter a selling phase, an oversold market in a buying phase.
The reason behind choosing the RSI indicator is simply because it allows you to quickly see whether
the market is overbought or oversold, which makes it easy to be interpreted. By the combination of RL outputs
and RSI technical indicator results, the model proposed can take investment decisions using rules mentioned
in Table 4. The scenario proposed for the trader to make the investment decision is that the result of the RL
algorithm is identical to the result obtained by the RSI indicator; otherwise, the trader must put his investment
on hold.

Figure 33. RSI indicator of exchange rate EUR/USD

Table 4. Decision rules of investment using RL and RSI technical indicator

RL decision RSI technical indicator decision Final decision
BUY BUY BUY
BUY SELL HOLD
SELL BUY HOLD
SELL SELL SELL

5. CONCLUSION
This paper represents the continuity of a previously published work of a model proposed to combine
two techniques which are: multiple regression and simulated annealing, the results showed that the combination
of two techniques gives good precision of prediction. From this conclusion, we proposed an improvement of
the model through the introduction of new techniques which are learning by reinforcement and RSI indicator.
All about harnessing the strengths of these techniques to refine investment decision-making, which must be
made based on the combination of the four techniques. It must be proven that the purpose of this model is not
only to maximize investment profits for traders, but also to limit the risks associated with the prediction and
this through the multitude and diversity of techniques used in terms of type to reassure better prediction
compared to using a single technique, we used supervised learning, meta heuristics, RL and technical analysis.
As a future work and following the current political and economic situation which is characterized by its
disturbed and non-stationary nature, we aim to extend the application domain of this model to contain the oil
market to predict the daily movements of this marketand analyze its influence on the global economy.
Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)
910  ISSN: 2252-8938

REFERENCES
[1] R. K. Lyons, “New perspective on FX markets: order-flow analysis,” International Finance, vol. 4, no. 2, pp. 303–320, Jul. 2001,
doi: 10.1111/1468-2362.00075.
[2] A. Geromichalos and K. M. Jung, “An over-the-counter approach to the Forex market,” International Economic Review, vol. 59,
no. 2, pp. 859–905, 2018, doi: 10.1111/iere.12290.
[3] R. K. others Lyons, The microstructure approach to exchange rates. Citeseer, 2001.
[4] S. Achchab, O. Bencharef, and A. Ouaarab, “A combination of regression techniques and cuckoo search algorithm for FOREX
speculation,” in World Conference on Information Systems and Technologies, 2017, vol. 569, pp. 226–235, doi: 10.1007/978-3-
319-56535-4_23.
[5] C. Evans, K. Pappas, and F. Xhafa, “Utilizing artificial neural networks and genetic algorithms to build an algo-trading model for
intra-day foreign exchange speculation,” Mathematical and Computer Modelling, vol. 58, no. 5–6, pp. 1249–1266, 2013,
doi: 10.1016/j.mcm.2013.02.002.
[6] L. Mendes, P. Godinho, and J. Dias, “A Forex trading system based on a genetic algorithm,” Journal of Heuristics, vol. 18, no. 4,
pp. 627–656, 2012, doi: 10.1007/s10732-012-9201-y.
[7] M. A. H. Dempster and Y. S. Romahi, “Intraday FX trading: an evolutionary reinforcement learning approach,” in International
Conference on Intelligent Data Engineering and Automated Learning, Springer Berlin Heidelberg, 2002, pp. 347–358.
[8] F. Rundo, “Deep LSTM with reinforcement learning layer for financial trend prediction in FX high frequency trading systems,”
Applied Sciences (Switzerland), vol. 9, no. 20, p. 4460, 2019, doi: 10.3390/app9204460.
[9] H. Jamali, O. Bencharef, A. Nabaji, K. El Housni, and Z. Asebriy, “Improving regression models using simulated annealing for
stock market speculation,” in World Conference on Information Systems and Technologies, 2018, vol. 745, pp. 295–304,
doi: 10.1007/978-3-319-77703-0_30.
[10] O. Chantarakasemchit, S. Nuchitprasitchai, and Y. Nilsiam, “Forex rates prediction on EUR/USD with simple moving average
technique and financial factors,” in 17th International Conference on Electrical Engineering/Electronics, Computer,
Telecommunications and Information Technology, ECTI-CON 2020, 2020, pp. 771–774, doi: 10.1109/ECTI-
CON49241.2020.9157907.
[11] M. A. H. Dempster, T. W. Payne, Y. Romahi, and G. W. P. Thompson, “Computational learning techniques for intraday FX trading
using popular technical indicators,” IEEE Transactions on Neural Networks, vol. 12, no. 4, pp. 744–754, Jul. 2001,
doi: 10.1109/72.935088.
[12] A. V. Rutkauskas, A. Miečinskiene, and V. Stasytyte, “Investment decisions modelling along sustainable development concept on
financial markets,” Technological and Economic Development of Economy, vol. 14, no. 3, pp. 417–427, Sep. 2008,
doi: 10.3846/1392-8619.2008.14.417-427.
[13] Y. L. Yong, D. C. L. Ngo, and Y. Lee, “Technical indicators for Forex forecasting: a preliminary study,” in International Conference
in Swarm Intelligence, 2015, vol. 9142, pp. 87–97, doi: 10.1007/978-3-319-20469-7_11.
[14] Y. Chihab, Z. Bousbaa, M. Chihab, O. Bencharef, and S. Ziti, “Algo-trading strategy for intraweek foreign exchange speculation
based on random forest and probit regression,” Applied Computational Intelligence and Soft Computing, vol. 2019, 2019,
doi: 10.1155/2019/8342461.
[15] B. Omar, G. C. Daniel, B. Zineb, and C. J. Aida, “A comparative study of machine learning algorithms for financial data prediction,”
in International Symposium on Advanced Electrical and Communication Technologies, ISAECT 2018 - Proceedings, Nov. 2018,
pp. 1–5, doi: 10.1109/ISAECT.2018.8618774.
[16] P. H. Damia Abd Samad, S. Mutalib, and S. Abdul-Rahman, “Analytics of stock market prices based on machine learning
algorithms,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 16, no. 2, pp. 1050–1058, 2019,
doi: 10.11591/ijeecs.v16.i2.pp1050-1058.
[17] D. A. Fattah, A. A. Naim, A. S. Desuky, and M. S. Zaki, “AutoKeras and particle swarm optimization to predict the price trend of
stock exchange,” Bulletin of Electrical Engineering and Informatics, vol. 11, no. 2, pp. 1100–1109, 2022, doi: 10.11591/eei.v11i2.3373.
[18] R. Rassetiadi and Suharjito, “Foreign exchange prediction based on indices and commodities price using convolutional neural
network,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 18, no. 1, pp. 494–501, Apr. 2019,
doi: 10.11591/ijeecs.v18.i1.pp494-501.
[19] M. Khushi and T. L. Meng, “Reinforcement learning in financial markets,” Data, vol. 4, no. 3, p. 110, 2019,
doi: 10.3390/data4030110.
[20] J. Carapuço, R. Neves, and N. Horta, “Reinforcement learning applied to Forex trading,” Applied Soft Computing Journal, vol. 73,
pp. 783–794, 2018, doi: 10.1016/j.asoc.2018.09.017.
[21] B. Jubert de Almeida, R. Ferreira Neves, and N. Horta, “Combining support vector machine with genetic algorithms to optimize
investments in Forex markets with high leverage,” Applied Soft Computing Journal, vol. 64, pp. 596–613, Mar. 2018,
doi: 10.1016/j.asoc.2017.12.047.
[22] “EA Trading Academy - online courses | strategy builders | expert advisors,” Trading Academy. https://eatradingacademy.com/
(accessed Oct. 25, 2016).
[23] R. Pardo, Design, testing, and optimization of trading systems, vol. 2. John wiley & Sons, 1992.
[24] J. F. Hair, Multivariate data analysis, 7th edision Prentice Hall International, 2010.
[25] L. Plonsky, Advancing quantitative methods in second language research. Routledge New York: Routledge New York, 2015.
[26] L. Plonsky and H. Ghanbar, “Multiple regression in L2 research: a methodological synthesis and guide to interpreting R2 values,”
Modern Language Journal, vol. 102, no. 4, pp. 713–731, Dec. 2018, doi: 10.1111/modl.12509.
[27] B. G. Tabachnick, L. S. Fidell, and J. B. Ullman, Using multivariate statistics, 6th edision vol. 5. pearson Boston, MA, 2007.
[28] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated annealing,” Science, vol. 220, no. 4598, pp. 671–680,
May 1983, doi: 10.1126/science.220.4598.671.
[29] L. Thorndike and D. Bruce, Animal intelligence: experimental studies. Routledge, 2017.

Int J Artif Intell, Vol. 12, No. 2, June 2023: 892-911

Int J Artif Intell ISSN: 2252-8938  911

BIOGRAPHIES OF AUTHORS

Hana Jamali a Ph.D. student in the Cadi Ayyad University, Marrakesh,

Morocco, on the topic of development of an investment strategy on the stock exchange data
mining approach. She has several international publications on the topic of her on going
thesis. She can be contacted at email: [email protected].

Dr. Younes Chihab He received the doctorate thesis in Network Security from
the Faculty of Sciences in the Cadi Ayyad University, Marrakech, Morocco, in December
2013. His research was in artificial intelligence, signal processing and machine learning. In
2019, he received the PHD degree in Computer Sciences from the Faculty of Sciences, Ibn
Tofail University, Kenitra, Morocco. His research interests include the signal and image
processing and coding, networking and artificial intelligence. Actually, he is a professor and
member of the Computer Sciences Research Laboratory (CSRL), Superior School of
Technology, Ibn Tofail University, Kenitra, Morocco. He can be contacted at email:
[email protected].

Iván García-Magariño is a lecturer and a contributor to the GRASIA research

group at the Complutense University of Madrid. Prior to commencing this position in 2018,
he was a PhD assistant professor in University of Zaragoza (2014-2018), and before he was
a lecturer at the Madrid Open University (2010-2014). Iván was awarded his Ph.D.in
Computer Science Engineering from the Complutense University of Madrid in 2009, and
was a recipient of an FPI scholarship from 2006 to 2010. His main research interests include
agent-based simulators, multi-agent systems, and agent-oriented software engineering.
Among journals, book chapters, conferences and workshops, he has over 100 publications
(over 47 in journals with ISI Thomson JCR). He actively collaborates with the EduQTech
research group from the University of Zaragoza onm-health projects, the HCI Lab from the
University of Udine (Italy) onhuman-computer interaction projects, Institute of Technology
Blanchardstown (Ireland) on datamining projects, and the Multicultural Alzheimer
Prevention Program (MAPP) of Massachusetts General Hospital and Harvard University
(US) on m-health projects. He is editor in several journals and guest editor in several special
issues. He can be contacted at email: [email protected].

Dr. Omar Bencharef received the D.Sc. degree (Doctor Habilitatus D.Sc.) in
computer science from the Faculty of Sciences and Technology, Cadi Ayyad University,
Marrakech, Morocco, in April 2018. He is a Professor of Computer Science in the Faculty
of Sciences and Technology, Cadi Ayyad University, since 2013. His research interests
include the signal and image processing and coding, networking and artificial intelligence.
He can be contacted at email: [email protected].

Hybrid Forex prediction model using multiple regression, simulated annealing … (Hana Jamali)