Study on Exchange Rate Forecasting Using Recurrent Neural Networks

Abstract

It focuses on the problems of forecasting exchange rate that is a nonlinear time series. A dynamics systems approach and the recurrent neural networks (RNN) were employed to modeling this nonlinear time series. The delay time was calculated using mutual information method and embedding dimension was confirmed by false nearest neighbors. The dataset was reconstructed form source time series for trained and verified the neural networks model. The quadratic optimization criterion was considered which the neural networks weights update algorithm were derived using gradient descent method for hidden layer; recurrent layer and output layer. The calculation flow chart was designed for neural networks learning and emulation. The reliability and stability of neural networks was confirmed by testing dataset. The results of simulation showed that the recurrent neural networks were preferably performance for prediction the change of exchange rate.