Enhancing the Momentum Strategy Using Recurrent Neural Networks

The momentum strategy of buying assets which performed well in the past and shorting those which performed poorly has been shown to generate persistent abnormal returns. Existing literature suggests combining this concept with predictive models in order to determine trading positions for momentum portfolios. In this paper, we propose the use recurrent neural networks in combination with such momentum strategies. Using data from 1953 to 2019, our empirical results show that utilizing recurrent neural networks results in similar profitability compared to classical methods. However, as main advantage, we find that up to a 15% improvement can be made with respect to Sharpe ratio. Additionally, during times of relative economic turmoil, recurrent neural networks are found to perform significantly better than the benchmark strategies.