The paper introduces the Momentum Transformer, a deep-learning architecture for time-series momentum trading that combines LSTM layers for local processing with multi-head self-attention for long-term dependencies. Unlike standard LSTM-based Deep Momentum Networks (DMNs), which struggle with regime changes and long-term patterns, the attention mechanism provides direct connections to all previous time-steps, enabling the model to learn longer-term dependencies and adapt to new market regimes. The architecture also includes a Variable Selection Network (VSN) that filters out noisy inputs and provides interpretable variable importance over time.

Backtesting on a portfolio of 50 liquid futures from 1990–2020 shows that the Momentum Transformer significantly outperforms classical TSMOM strategies and LSTM-based DMNs across multiple risk-adjusted metrics, including Sharpe, Sortino, and Calmar ratios. The improvement is especially pronounced during periods of high nonstationarity, such as 2015–2020 and the SARS-CoV-2 crisis. The addition of a changepoint detection (CPD) module further enhances performance, demonstrating that attention and CPD are complementary. The model's interpretability is highlighted through attention patterns that reveal which past time-steps and factors are most important, providing insights into the trading strategy's decision-making process.