Momentum strategies are widely documented anomalies shown to generate excess returns, due to delayed action to firm-specific information, however, their mean returns fluctuate due to noise and changing market regimes 1. Existing ML applications use the tool to focus on the direction of price; it is unclear whether ML can accurately filter quality signals. This paper investigates whether an MLP-based filter can improve the risk-adjusted returns of a time-series momentum strategy, keeping constant conditions with the same signal logic as the traditional strategy. The ML filter is fed price-derived features, and trained on a split of 60/20/20 spanning from 2010-2026. The ML filter uses a 3-layer MLP, computed with binary cross entropy loss, where each label is deemed profitable if the 5-day returns exceed the 10 bps transaction cost of the EMA signal. When evaluated with the traditional strategy, the machine learning filter showed a decline in mean sharpe by -0.242, improvement in mean drawdown by +5.8 pp, decreased mean trade frequency of 42%, with a test accuracy of 52.13%. In conclusion, these results indicate that ML works primarily as a risk-reduction tool, reducing drawdown more consistently than improving the risk-adjusted returns, suggesting that calibration of ML is the appropriate way to evaluate the effectiveness of ML-filtered momentum-based strategies.
Advaith Prabhu (Mon,) studied this question.