Deep Momentum
Log in to collectAcademic paper
Strategy in a nutshell
The paper develops a deep momentum (DM) strategy using US equity market data from CRSP (1955–2017) and machine learning techniques. Stocks are classified into return deciles using deep neural networks (DNNs), with multiclass classification predicting future return classes and regression forecasting absolute returns. The strategy addresses bimodal return distributions in momentum and other firm characteristics, applying five reclassification methods to improve prediction accuracy. Portfolios are formed based on predicted classes, and performance is evaluated in terms of returns, Sharpe ratios, and robustness compared to conventional strategies. The DM approach captures nonlinear patterns and multiple features, enhancing the predictive power of momentum-based investment strategies.
Economic rationale
The DM framework leverages machine learning to reconcile predicted returns with actual financial performance, mitigating bimodality issues in momentum stocks. By estimating return distributions via DNNs and applying reclassification techniques, it uncovers hidden nonlinear information, improving portfolio profitability and robustness. Empirical evidence demonstrates superior performance over traditional strategies, even accounting for transaction costs, highlighting the value of integrating advanced ML methods into asset pricing and portfolio construction. This approach advances the understanding of stock return predictability and the role of feature-driven models in investment decision-making.