Deep momentum networks with market trend dynamics

Abstract

Time-series momentum (TSMOM) trading strategies manage positions based on the persistence of return trends. Although long short-term memory (LSTM) deep neural architectures can enhance TSMOM, their performance often deteriorates during abrupt market trend changes. This study aims to improve TSMOM performance, particularly at critical moments marked by significant shifts in long- and short-term trends. To achieve this, we combine short- and long-term signals into a comprehensive market-state representation, employing supervised learning to incorporate these market dynamics into the proposed model. In our experiments, we generate market-state features, referred to as MTDP scores, by numerically capturing changes in market trends via an extreme gradient boosting (XGBoost) process. These MTDP scores are then applied within an LSTM-based trading strategy. A backtest on 99 continuous futures (1995–2021) demonstrates that incorporating MTDP scores enhances the Sharpe ratio, indicating the effectiveness of embedding market-state information in TSMOM. Notably, an 8-week fast momentum look-back window performed best over stable periods (1995–2019). However, during extreme market downturns, such as the COVID-19 crisis, a 20-week fast momentum window not only outperformed shorter windows (4- and 8-week signals) but also recovered more rapidly. These findings suggest that TSMOM strategies can benefit from dynamically adjusting fast momentum windows, consistently generating profitable opportunities even amid turbulent conditions.