Revealing Pairs-trading opportunities with long short-term memory networks

Abstract

This work examines a deep learning approach to complement investors’ practices for the identification of pairs-trading opportunities among cointegrated stocks. We refer to the reversal effect, consisting in the fact that temporarily market deviations are likely to correct and finally converge again, to generate valuable pairs-trading signals based on the application of Long Short-Term Memory networks (LSTM). Specifically, we propose to use the LSTM to estimate the probability of a stock to exhibit increasing market returns in the near future compared to its peers, and we compare and combine these predictions with trading practices based on sorting stocks according to either price or returns gaps. In so doing, we investigate the ability of our proposed approach to provide valuable signals under different perspectives including variations in the investment horizons, transaction costs and weighting schemes. Our analysis shows that strategies including such predictions can contribute to improve portfolio performances providing predictive signals whose information content goes above and beyond the one embedded in both price and returns gaps.