Machine learning in empirical asset pricing: a critical review on the UK Stock Market

This study asks a number of questions: (1)why are the existing risk factor models ineffective at explaining returns within the UK stock market, and can elements of ML be used to improve them? (2) can ML models be leveraged to provide economic value to investors within the UK equity market by modelling the high-order interaction and non-linear effects typical linear models struggle to incorporate? and (3)can model agnostic interpretability measures be used to gain insight into the non-linear relationships and the complex interaction structure of equity returns anomaly variables while providing insight as to how ML models make their decisions?

Using a data set spanning over 20 years, using monthly returns, and several stock characteristic and fundamental variables of UK firms, we find that: (1) Linear risk factor models, even with decision tree-based factors, fail to consistently explain the cross section of returns; (2) a random forest trained on a small set of historically-lagged stock characteristics, including market capitalisation and momentum, was able to inform investment decisions that yielded an average risk-adjusted monthly return of 1.33%.