Exploring Possible Improvements to Momentum Strategies with Deep Learning

Momentum is a well-known phenomenon that pervades in all major stock markets. Strategies that exploit this phenomenon have been documented to generate abnormal returns persistently over time. In this paper, we study these portfolio designs and explore possible improvements, employing machine-learning and deep-learning models in the return prediction. To forecast the returns of the best- and worst-performing stocks, we use support vector regression and the latest deep-learning techniques such as stacked (denoising) autoencoders and state-of-the-art generative adversarial networks. Our analysis focuses on the comparison of financial performance measures, such as average return, Sharpe ratio and maximum drawdown, between the simple momentum strategies and the advanced deep-learning models. The results confirm our expectations, namely that the complex models can improve momentum portfolios, mainly by avoiding large losses in times of crises. For generative models, the reduction of 50% in the maximum drawdown is achieved, compared to a basic momentum strategy. Furthermore, our results indicate a slight fading of the negative momentum, such that the stocks losing value in the past cannot be easily exploited by complex models to generate abnormal returns.