Modeling of short-term dynamics of foreign exchange rates using deep neural networks

Derbentsev Vasily
PhD in Economics, Associate Professor at the Department of Informatics and Systemology, Kyiv National Economic University named after Vadym Hetman, Kyiv, E-mail:derbv@kneu.edu.ua

Bezkorovainyi Vitalii
assistant professor at Informatics and Systemology, Kyiv National Economic University named after Vadym Hetman, Kyiv, E-mail:retal.vs@kneu.edu.ua

Ovcharenko Andrey
senior lecturer at the Department of Informatics and Systemology, Kyiv National Economic University named after Vadym Hetman, Kyiv, E-mail:ov_andrei@i.ua

Cite this article:

Derbentsev V., Bezkorovainyi V., Ovcharenko A. (2020) Modeling of short-term dynamics of foreign exchange rates using deep neural networks. Ed.: D.V. Zavadska (ed.-in-ch.) and others [Modelyuvannya korotkostrokovoy dynamiky valyutnykh kursiv z vykorystannyam hlybokykh nejronnykh merezh; za red.: D.V. Zavadska (gol. red.)], Scientific Bulletin of the Odessa National Economic University (ISSN 2313-4569), Odessa National Economics University, Odessa, No. 3-2(276-277), pp. 153-163.

Abstract

This paper is devoted to the short-term predicting of exchange rates using Deep Learning approaches (DL). The undeniable advantage of using deep networks is their ability to find hidden complex nonlinear patterns in the data, as well as to identify influential factors (carry out automatic feature extraction). For this purpose, the DL models were built on the basis of Convolutional (CNN) and Recurrent (RNN) Neural Networks. The CNN block performs the function of feature extraction, and the RNN which based on the Long-term Short-term Memory (LSTM) performs the forecast. For parameters estimating and testing the models we used daily and four hourly observations of currencies Euro/Dollar, British pound/Dollar, and cryptocurrencies (Bitcoin and Ethereum) for the period from 02/01/2015 to 12/31/2020 according to the service Yahoo Finance. As input data, we used open prices (Open), minimum (Low), maximum (High), and close prices (Close) for the corresponding timeframe. In the experimental section we compared the performance of the designed models using both daily and four-hour data sets. The accuracy of the forecasting performance was assessed by the values of the Mean Absolute Percentage Error (MAPE), which allows comparing forecast errors for different assets and models. In addition, the Mean Square Error (MSE) and the Root Mean Square Error (RMSE) were also calculated. The highest accuracy (in the sense of the MAPE metrics) was for the EUR/USD – about 0.4% for both daily and for-hour data sets. More volatile GBP/USD quotes show a larger error on both daily and four-hour quotes. But in general, an increase in the number of observations in four-hour time series reduces the model error for EUR/USD and GBP/USD. The results of the cryptocurrencies forecast turned out to be less accurate: 5.9% and 8.5%, respectively. Our study showed the prospects of using DL networks such as CNN and LSTM to predicting the short-term exchange rates. According to obtained results proposed models provide an efficient forecast for both fiat and cryptocurrencies.

Keywords

deep learning, neural networks, short-term forecasting, time series of currency quotes, cryptocurrencies.

JEL classification: G170

UD classification: 004.942:[336.743]:519.868

References

1. Международная практика прогнозирования мировых цен на финансовых рынках(сырье, акции, курсы валют) / под ред. Я.М. Миркина. Москва: Магистр, 2014. 456 с.
2. Persio, L., Honchar, O. Multitask machine learning for financial forecasting. International Journal of Circuits, Systems and Signal Processing. 2018. Vol. 12. P. 444–451.
3. В.Д. Дербенцев, Г.І. Великоіваненко, Н.В. Даценко. Застосування методів машинного навчання до прогнозування часових рядів криптовалют. Нейро-нечіткі технології моделювання в економіці. 2019. № 8. С. 65-93.
4. Hamid, S.A., Habib, A. Financial forecasting with neural networks. Academy of Accounting and Financial Studies Journal. 2014. Vol. 18(4). P. 37-55.
5. Hitam, N.A., Ismail, A.R. Comparative Performance of Machine Learning Algorithms for Cryptocurrency Forecasting. 2018. URL: https://www.researchgate.net/ publication/327415267 (дата звернення: 01.12.2020).
6. Bahrammirzaee, A. A comparative survey of artificial intelligence applications in finance:artificial neural networks, expert system and hybrid intelligent systems. Neural Comput. Appl. 2010. Vol. 19(8), P. 1165–1195.
7. Sezer O.B., Gudelek M.U., Ozbayoglu A.M. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019. Applied Soft Computing. 2020. Vol.90. Article number 106181.
8. Дербенцев В.Д., Безкоровайний В.С., Луняк І.В. Застосування методів глибокого навчання до прогнозування зміни короткострокових трендів валютних курсів. Науковий вісник Мукачівського державного університету. Секція: економіка. 2020. № 7(2). С. 75-86.
9. Kumar, D., Rath, S.K.: Predicting the Trends of Price for Ethereum Using Deep Learning Technique, in: Artif. Intell. Evol. Comput. Eng. Syst., pp. 103–114, Springer (2020)
10. Saxena, A., Sukumar, T. Predicting bitcoin price using LSTM and compare its predictability with ARIMA model. International Journal of Pure and Applied Mathematics. 2018. Vol. 119(17). P. 2591-2600.
11. Livieris, I.E., Pintelas, E., Stavroyiannis, S., & Pintelas, P. Ensemble deep learning models for forecasting cryptocurrency time-series. Algorithms. 2020. Vol. 13(5), article number 121.
12. Bao W, Yue J, Rao Y A deep learning framework for financial time series using stacked autoencoders and long-short term memory. PLoS ONE. 2017. Vol. 12(7): e0180944. doi:https://doi.org/10.1371/journal.pone.0180944