Foreign Exchange Rate Forecasting using Autoencoder and LSTM Networks

Abstract

The problem of predicting foreign currency exchange rate is the problem in which many researchers in forecasting community have been interested. The exchange rate changes occur hourly, even seconds, thus producing correlated time series. To enable accurate forecasting on such correlated time series data, this work proposes a deep learning model which combines two approaches: autoencoder and Long Short-Term Memory (LSTM) network. The model employs an LSTM-based autoencoder in order to extract features from input dataset well. The model also uses LSTM-based network as a forecaster, which provides accurate and robust forecasting. Experimental results on four exchange rate datasets suggest that the proposed model is effective and outperforms five other comparative methods: shallow neural network (ANN), deep belief network (DBN), convolutional neural network (CNN), LSTM network and another form of combining autoencoder and LSTM network.