Foreign Exchange Rates Forecasting With Convolutional Neural Network

Using information technology for international foreign exchange rates forecasting will help investors and policy-makers get more profits and make better policies.However,for the existing models and methods always used in the international foreign exchange market,the optimal number of inputs usually lies between 3 and 10 and they cannot use all the features the international foreign exchange market has and cannot get satisfactory results when they are used for international foreign exchange rates forecasting.Therefore,for both of the academic research and engineering application,designing a model and a method to improve the long-term forecasting performance is very valuable.In this paper,to utilize the periodic characteristics of the international foreign exchange market,a novel method of transforming international foreign exchange rates data from one-dimensional architecture to two-dimensional architecture is introduced.We also propose a CNN-based model which not only could process the trading-days’ changes of international foreign exchange rates,but also could process the details of exchange rates’ different states at different times in the day.Thanks to this feature,this model can utilize the periodic characteristics of the international foreign exchange market.In our research,the most important three currency pairs,Euro against US dollar(EUR/USD),US dollar against Japanese yen(USD/JPY)and British Pound Sterling against US dollar(GBP/USD)are researched in this paper.We will compare the proposed CNN-based model with Fully Connected Artificial Neural Network(FCANN),Support Vector Regression(SVR)and Gated Recurrent Unit(GRU).Based on five performance evaluations: Mean Square Error,Mean Absolute Percentage Error,the coefficient of determination,directional symmetry and CUSUM explained variance,the final experimental results show that the proposed CNN-based model has a better performance when compared with other traditional models.