Developing a framework for forecasting financial time series using Wavelet-LSTM hybrid model

Abstract

In this paper, 4 models are fitted in EUR/USD, GBP/USD, and USD/JPY rates. First LSTM model is fitted to forecast the price at the next step and is compared with Wavelet-LSTM hybrid model. The results show that the hybrid model performs better than the LSTM model without wavelet decomposition. However, mean absolute percentage error (MAPE) shows that the hybrid model performs poor for trading purposes. Then a classification problem is considered. The objective of the problem is to predict the direction of the return on the next timestamp. Again LSTM is compared with Wavelet-LSTM model. The results show that Wavelet-LSTM performs better than LSTM. The considered simple trading algorithm achieves an annualized return of 20% with Wavelet-LSTM hybrid model.