Deep Learning Meets Statistical Arbitrage: An Application of Long Short-Term Memory Networks to Algorithmic Trading

Abstract

In this article, the authors study the utility of deep-learning approaches in statistical arbitrage under the generalized pairs-trading paradigm. Stock returns are regressed on a set of risk factors derived using principal component analysis, and the long short-term memory (LSTM) structure is employed to forecast directions of idiosyncratic residuals. Daily market-neutral trades are constructed based on the predicted signals. The authors compare their results with the influential relative value (RV) model by Avellaneda and Lee (2010) on the universe of S&P 500 Index (S&P 500) stocks. Model evaluations are performed on two distinct periods (2001–2007 and 2015–2021) to alleviate the survivorship bias resulting from the S&P 500 composition changes over time and to study the robustness of these two models in two distinct eras. Their findings suggest that the LSTM model consistently and significantly outperforms the RV model across the two periods when transaction costs are accounted for. However, in the transaction cost–free world, the outperformance is modest even though it is still consistent.