Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Foreign Exchange Rates Prediction Based on Comparative
Models
Yuyao Wang
Department of AIAC, Xi’an Jiaotong-liverpool University, Suzhou, China
[email protected]Abstract. With the development of globalization, fluctuations in the foreign exchange market have
the huge potential to influence global economy. Machine Learning technology shows great potential
in the field of financial forecasting than traditional statistical methods in recent years. This research
selected statistical model Autoregressive Integrated Moving Average (ARIMA) and machine learning
models Linear Regression and Long- and Short-Term Memory (LSTM) to make training and testing
of historical data on the forex market. Root Mean Squared Error (RMSE), Mean Absolute Error (MAE),
and R-Squared were chosen as evaluation indexes to evaluate the performance of each model and
compare their performance in forex exchange rate forecasts. The results show that ARIMA and
LSTM could make better predictions than Linear Regression, between which LSTM has higher
degree of fitting data. This research provides a detailed analysis comparing the performance of
different models in forex exchange rate forecasting whose results are beneficial to guide forex market
forecast in practical application and provide reference for future research.
Keywords: Statistics, Machine Learning, ARIMA, Linear Regression, LSTM.
1. Introduction
The foreign exchange (forex) market stands as one of the most dynamic and crucial components
of the global financial system, whose many complicated factors have great impacts on foreign
exchange rates [1]. Foreign exchange rates play a crucial role in international trade, investment and
economic globalization. The fluctuations of foreign exchange rates could bring profound influence to
the economic system and market participants. Moreover, foreign exchange rate fluctuations may
affect the investment decisions of multinational corporations and domestic inflation or deflation.
Especially, sharp fluctuations in foreign exchange rates have great impacts on the stability of financial
markets, increasing investors’ worries of market risk [2]. Therefore, researchers have gradually begun
to predict foreign exchange rates by using statistical models and machine learning models in order to
help economic decision-making and policies formulation with the development of innovative
technologies, controlling market risks [1].
Statistical methods and machine learning models have been widely applied in healthcare,
education and finance to solve different problems related to data analysis [3-5], among which
predictions of foreign exchange rates has always been a popular topic. Colombo & Pelagatti, 2020
have made predictions on foreign exchange rates through statistical learning models, which
performed well in short term forecasting. What’s more, they ranked the importance of different
variables and made analysis of the correlation between the different variables and results in order to
explain how statistical models operated [6]. Apart from this, Dautel et al., 2020 used different models
of machine learning to forecast foreign exchange rates, who stated that deep learning has had huge
impacts in finance field. Therefore, four models --- long short-term memory networks, gated recurrent
units, traditional recurrent networks and feedforward networks were used to carry on the research.
They concluded that a simple neural network may have better performance than a complex deep
neural network in some respects [7]. Besides this, there is one issue that whether statistical models or
deep learning models perform better. Through the research conducted by Aggarwal & Sahani, 2020
and Bangyuan, n.d., it is obvious that deep learning models perform better on both long and short
term predictions on foreign exchange rates than statistical learning models, making great contribution
to financial researches [1, 8].
341
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
This paper aims to predict foreign exchange rates. More specifically, two foreign exchange rates
were chosen: exchanges rates of JPY to USD and exchange rates of GBP to USD in recent 5 years as
data set. Considering the data is time series data, this paper will use 3 models: Autoregressive
Integrated Moving Average (ARIMA), Linear Regression, and Long Short Term Memory (LSTM)
to predict foreign exchange rates.
2. Methodology
2.1. Data Description
This research selected two foreign exchange rates: exchange rate of JPY to USD and exchange
rate of GBP to USD in recent 5 years from February 22, 2019 to February 22, 2024 from the website:
https://cn.investing.com/currencies [9] as data set. It is obvious that these data are time series data.
The predictive objective is the change of foreign exchange rate data. The basic information of the
data set is shown in Table1, which has 1502 rows and 2 columns. According to the data set, the trend
of original data is shown in Fig. 1 and Fig. 2.
Table 1. Data Information
Column Non-Null Count Dtype
Exchange Rate of JPY to USD 1305 non-null Float64
Exchange Rate of GBP to USD 1305 non-null Float64
Figure 1. Trend of JPY to USD (Photo/Picture credit: Original)
Figure 2. Trend of GBP to USD (Picture credit: Original)
2.2. Data Pre-Processing
Before making the fitting and predicting of the model, the research preprocessed the time series
data from the data set. Firstly, the ‘Date’ column is set as the index column. Secondly, this research
drawn the box plot of the two rates respectively (shown in Fig. 3), from which the outliers in the
column ‘Exchange Rate of GBP to USD’ are found. In order to deal with the outliers, the research
342
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
calculated the value of quartile: Q1 (25th percentile) = 0.7518, Q3 (75th percentile) = 0.8094, IQR
(Interquartile Range) = 0.0576, the lower bound is 0.6654, and the upper bound is 0.8958. Then this
research replaced the outliers with the median of column ‘Exchange Rate of GBP to USD’ and got
the processed data (shown in Fig. 4).
Figure 3. Box Plot (Picture credit: Original)
Figure 4. Processed Data (Picture credit: Original)
Furthermore, this research Use 'MinMaxScaler' from the ‘sklearn' to standardize the time series
data and scale the data into a specific range [-1, 1]. Next, this research set the time window size
‘lookback’ to 100 and split the data into training and test sets. By iterating the time series data, the
previous' lookback 'time step is taken as the input feature, and the value of the next time step is taken
as the target value. After the above processing, the data is transformed into a format suitable for
LSTM model training, and important time information is retained, which helps to improve the model's
interpretability and generalization ability.
2.3. ARIMA
ARIMA represents ‘Autoregressive Integrated Moving Average’ which is made up of
autoregression (AR), difference (I), and moving average (MA). This model could capture both linear
and nonlinear relationships of data, which is a common method to analyze time series data, which is
a common statistical model used to predict changes in time series. The workflow of ARIMA is shown
in Fig. 5 [10].
343
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Figure 5. ARIMA Workflow (Picture credit: Original)
Taking the ‘Exchange Rate of JPY to USD’ as example, the research first used ADF test to test
the stationarity of time series data and found that the p-value was greater than the significance level
(usually 0.05), the null hypothesis could not be rejected, that is, the data was not stationary. Therefore,
the research made difference on the data, and ADF test is conducted on the column data after two
differences, and the p-value is less than 0.05, so the data passes the stationarity test. Then, LB test
was used to conduct white noise test on the data, and it was judged that it was not a white noise
sequence, so it could continue to be analyzed. ACF and PACF charts (shown in Fig. 6 and Fig. 7)
were drawn to estimate the values of p and q. The last 100 data were selected from the whole column
of data, and 90% were divided into training sets and 10% into test sets. Furthermore, the research
tried to find the optimal ARIMA model parameters through iteration, which were finally determined
as p=10, d=1, q=3. The parameters of ‘Exchange Rate of GBP to USD’ were determined by the same
operation as p=12, d=3, q=2.
Figure 6. ACF for JPY to USD (Picture credit: Original)
344
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Figure 7. PACF for GBP to USD (Picture credit: Original)
2.4. Linear Regression
Linear Regression is a simple statistical method used to model the relationship between a
dependent variable and one or more independent variables by fitting a linear equation to observed
again. The workflow of linear regression is shown in Fig. 8. This research put ‘Exchange Rate of JPY
to USD’ and ‘Exchange Rate of GBP to USD’ respectively into the linear regression model for fitting
training [11].
Figure 8. Linear Regression Workflow (Picture credit: Original)
2.5. LSTM
LSTM, also called ‘Long-short Term Memory’, is a type of Recurrent Neural Network (RNN)
architecture, specifically designed to address vanishing gradient problem in traditional RNNs, and to
capture long-term dependencies in sequential data. The workflow of LSTM is shown in Fig. 9 [12].
345
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Figure 9. LSTM Workflow [13]
The neural network model used in this research adopts the structure of 2-layer LSTM, whose input
dimension is set to 1 which means the number of features per time step is 1, the hidden dimension is
set to 32, and the output dimension is 1. During the training process, the model was trained for 200
rounds. This research used the LSTM layer and the fully connected layer, where the LSTM is
responsible for processing the long-term dependency of the time series data, and the fully connected
layer maps the output of the LSTM layer to the final output dimension. Furthermore, the mean square
error Loss function (MSE Loss) was used as the loss function of the model, Adam was used as the
optimizer of the model, and the learning rate was set at 0.05. Over 200 training cycles, the research
trained the training set repeatedly and recorded the loss value for each training cycle.
3. Results and Discussion
3.1. Results
3.1.1 ARIMA
The results of ARIMA were shown in Fig. 10 to Fig. 13.
Figure 10. Results of JPY to USD (Photo/Picture credit: Original)
346
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Figure 11. Results of GBP to USD (Photo/Picture credit: Original)
Figure 12. Predictions of train set (GBP to USD) (Photo/Picture credit: Original)
Figure 13. Predictions of test set (GBP to USD) (Photo/Picture credit: Original)
347
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
ARIMA model has the advantages of simplicity and interpretability. The architecture of the model
is straightforward, making it easy to grasp and execute. However, the processes involved are complex,
and it demands strict conditions for time series data. ARIMA might not yield optimal results for
datasets characterized by nonlinearity, lack of stationarity, or non-Gaussian distributions.
3.1.2 Linear Regression
The two comparison graphs between the actual value and the predicted value were drawn and
shown in Fig. 14 and Fig. 15.
Figure 14. Results of JPY to USD (Photo/Picture credit: Original)
Figure 15. Results of GBP to USD (Photo/Picture credit: Original)
Linear Regression can freely add a variety of independent variables, including time trends,
seasonal factors, and other influencing factors, which is easy to understand and implement, so as to
better capture the dynamic characteristics of time series data. The parameters of the model are
intuitively understandable, yet it falls short in accurately fitting nonlinear data and shows increased
sensitivity to outliers.
3.1.3 LSTM
Through the visual analysis of the loss value, the change trend of the loss value in the training
process of the model was observed (shown in Fig. 16 and Fig. 17), and the convergence of the model
and the training effect were judged.
348
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Figure 16. MSE Loss of JPY to USD (Photo/Picture credit: Original)
Figure 17. MSE Loss of GBP to USD (Photo/Picture credit: Original)
The comparison result graphs between the actual value and the predicted value by LSTM are
shown in Fig. 18 and Fig. 19.
Figure 18. Results of JPY to USD (Photo/Picture credit: Original)
349
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
Figure 19. Results of GBP to USD (Picture credit: Original)
LSTM can effectively capture long-term dependencies in time series data through memory unit
and gating mechanism, which is suitable for complex time series patterns. It can also adaptively learn
the features and patterns of time series data. However, when LSTM deals with large-scale data, the
training time is long, the data requirements are also very high, and the predicted data has a certain
lag.
3.2. Discussion
This paper selected 3 indexes to evaluate the three models: Root Mean Square Error (RMSE),
Mean Absolute Error (MAE), R-squared. The comparative results of the indexes of the three models
are shown in Table 2 and Table 3.
Table 2. Evaluation of Three Models (JPY to USD)
Model / Index RMSE MAE R-squared
ARIMA 0.2823 0.1457 0.3469
Linear Regression 7.3439 6.3274 0.7669
LSTM 1.0786 0.8805 0.8390
Table 3. Evaluation of Three Models (GBP to USD)
Model / Index RMSE MAE R-squared
ARIMA 0.0017 0.0012 0.5494
Linear Regression 0.0383 0.0312 0.5777
LSTM 0.0041 0.0031 0.9163
In order to compare the prediction accuracy and error variance of the models, the RMSE and MAE
values of the three models were observed. Among them, the RMSE and MAE values of ARIMA are
the smallest, from which it can be seen that the prediction error of ARIMA is the smallest and the
prediction effect is more accurate. In this respect, the predictive ability of ARIMA is higher than that
of Linear Regression and LSTM.
In order to compare the degree of fit of the data, the research observed the R-squared values of the
three models. The R-squared value of LSTM is closest to 1. It can be seen that LSTM has the highest
degree of fitting to the data. In this respect, the data fitting capability of LSTM is better than that of
ARIMA and Linear Regression.
4. Conclusion
This research used comparative models to make predictions on foreign exchange rates. The models
are ARIMA, Linear Regression, and LSTM. The evaluation indexes chose are RMSE, MAE, R-
squared. At last, this research makes a conclusion that ARIMA predicts more accurately than other
350
Highlights in Business, Economics and Management FTMM 2024
Volume 40 (2024)
models while LSTM has the highest degree of fitting to data. The limitation is that this research did
not take seasonal and other factors into consideration during the predictions. Although this research
mainly focuses on the role of ARIMA, Linear Regression, LSTM in predicting foreign exchange rates,
their application potential in other fields is still broad. In the future, the future study will seek to
expand the research to explore the potential applications of statistical models and machine learning
in other fields such as healthcare, education, and more. This will help deepen the understanding of
how statistics and machine learning work in different contexts and provide a theoretical and practical
basis for their further application.
References
[1] Bangyuan Z. Critical Comparisons on Deep Learning Approaches for Foreign Exchange Rate Prediction.
Retrieved February 29, 2024, from https://arxiv.org/ftp/arxiv/papers/2307/2307.06600.pdf.
[2] Lal M, Kumar S, Pandey DK, Rai VK, Lim WM. Exchange rate volatility and international trade. Journal
of Business Research, 2023, 167: 114156.
[3] Qiu Y, Wang J, Jin Z, et al. Pose-guided matching based on deep learning for assessing quality of action
on rehabilitation training. Biomedical Signal Processing and Control, 2022, 72: 103323.
[4] Biju A K V N, Thomas A S, Thasneem J. Examining the research taxonomy of artificial intelligence, deep
learning & machine learning in the financial sphere—a bibliometric analysis. Quality & Quantity, 2024,
58 (1): 849-878.
[5] Qiu Y, Chang C S, Yan J L, et al. Semantic segmentation of intracranial hemorrhages in head CT scans.
2019 IEEE 10th International Conference on Software Engineering and Service Science (ICSESS). IEEE,
2019: 112-115.
[6] Colombo E, Pelagatti M. Statistical learning and exchange rate forecasting. International Journal of
Forecasting, 2020.
[7] Dautel AJ, Härdle WK, Lessmann S, Seow H-V. Forex exchange rate forecasting using deep recurrent
neural networks. Digital Finance, 2020.
[8] Aggarwal P, Sahani AK. Comparison of Neural Networks for Foreign Exchange Rate Prediction. IEEE
Xplore, November 1, 2020.
[9] Investing. Today's exchange rate of foreign exchange of foreign currency to real-time query _ the currency
market and movements of affection Investing.com Retrieved March 16, 2024, from
https://cn.investing.com/currencies.
[10] Hayes A. Autoregressive Integrated Moving Average (ARIMA), December 18, 2022. Investopedia.
https://www.investopedia.com/terms/a/autoregressive-integrated-moving-average-arima.asp.
[11] Zach. Introduction to Simple Linear Regression, December 11, 2018. Statology.
[12] Brownlee J. A Gentle Introduction to Long Short-Term Memory Networks by the Experts, July 19, 2017.
Machine Learning Mastery.
[13] Chang M, Bae S, Cha G, Yoo J. Aggregated Electric Vehicle Fast-Charging Power Demand Analysis and
Forecast Based on LSTM Neural Network. Sustainability, 2021, 13 (24): 13783.
351
