Application of recurrent neural networks to momentum trading

In recent years, deep learning techniques have been developed to handle complexity in the data which can not be achieved with traditional econometric techniques. However, the application of these techniques to real world problems have not been studied extensively. In this paper, we aim to investigate whether recurrent neural networks using a LSTM-autoencoder can contribute to financial time series forecasting and especially, trading strategies. We will combine different regression methods with a modified momentum strategy suggested by Kim (2019) and evaluate its performance on S&P 500 data. Our finding indicates that regression methods using LSTM-autoencoders lead to an improved profitability performance and predictive accuracy performance compared to regression methods based on shallow learning or non-recurrent deep learning algorithms. This implies that preserving sequential information is crucial in time series forecasting, for which recurrent neural networks are designed.