Machine learning (ML) is increasingly prominent in modern finance and is being adopted across a wide range of applications. However, using ML to extract alpha remains nontrivial. Reference [1] proposes a novel approach to improving the risk-adjusted returns of momentum strategies using machine learning.
It employs three ML methods—linear regression, random forest, and neural networks—to predict next month’s information coefficients (ICs) using information available at the end of the previous month. These predicted ICs are then combined with momentum factors to construct momentum and reversal strategies.
The authors pointed out,
Machine learning helps improve momentum strategies by capturing strong correlations between stocks and their returns using information coefficients (IC). Shallow machine learning enables us to trace the sources of advantage and adapt to non-linear correlations that traditional methods might overlook. We confirm that a unidirectional momentum strategy can yield significant returns. Furthermore, we observe that implementing a bidirectional approach leads to even more pronounced gains. We find that shorting weak stocks outperforms longing strong stocks, indicating the persistence of the "winners keep winning, losers keep losing" phenomenon. Additionally, we observe that price reversals do not occur within a month. Optimizing the number of stocks in the portfolio leads to the highest excess returns within a certain threshold range. However, implementing inverse trades using machine learning models does not yield positive returns, indicating that the predictive ability of the models may be inadequate to identify profitable opportunities in these situations. Finally, we discover that momentum factors with longer holding periods are more powerful predictors, accurately capturing the correlation between stocks and their returns. We assume that momentum factors with shorter observation periods are less effective, possibly due to the effect of mean reversion. Our research underscores the importance of accurately predicting IC values and the effectiveness of momentum strategies in generating positive returns.
In short, the paper applies machine learning models to momentum investing by predicting future information coefficients and using them to construct momentum and reversal portfolios. It finds that bidirectional momentum strategies outperform unidirectional ones, longer-horizon momentum signals are more informative, and portfolio performance is maximized when the number of selected stocks is optimized within a threshold.
This represents an interesting twist on applying machine learning to investing, using predicted information coefficients to guide portfolio construction rather than directly forecasting returns.
Let us know what you think in the comments below or in the discussion forum.
References
[1] Chang, H.-L., Cheng, H.-W., Lan, Y.-M., & Yu, J.-P. (2025). Machine Learning in Momentum Strategies. EFMA
Post Source Here: Improving Momentum Strategies with Machine Learning
source https://harbourfronts.com/improving-momentum-strategies-machine-learning/
