AQR: How to use Machine Learning to improve Your Investment Strategy

The folks over at AQR recently wrote a piece on how to use machine learning in stock portfolio construction. Specifically stock portfolio selection and timing. The article is quite technical but do provide some food for thought for less sophisticated investors on how AI can be implemented in an investment strategy.

Here are the key takeaways:

• The Virtue of Complexity: Machine learning techniques can significantly improve stock selection strategies by capturing ‘nonlinear relationships’ between predictor variables and stock returns. Complex models that incorporate numerous nonlinear factors can outperform simple linear approaches by 50-100% in terms of portfolio performance.

• Nonlinear Relationships Matter: The relationship between investment signals (like value and momentum) and stock returns is ‘complex and nonlinear’. Traditional linear models miss critical nuanced interactions between different investment signals. By using machine learning techniques to generate nonlinear predictors, investors can develop more sophisticated portfolio allocation strategies.

• Performance Across Different Signal Sets: The research demonstrated the “virtue of complexity” across three different signal sets: Value and Momentum, Fama-French 5-Factor Model plus Momentum, Defensive-Oriented Signals

In each case, complex models generated substantially higher Sharpe ratios compared to simple linear models, with improvements ranging from 50-100%.

• Caution Against Indiscriminate Data Mining: While complexity can improve portfolio performance, researchers warn against randomly adding predictor variables. Signal relevance is crucial – Including unrelated or noisy variables can rapidly degrade portfolio performance. The nonlinear predictors must have a meaningful relationship to expected returns.

• Regularization is Key: To manage the challenges of complex models with many parameters, techniques like ridge regression help identify precise portfolio weights while preventing overfitting. This allows investors to leverage model complexity without sacrificing statistical reliability.