Equity Machine Factor Models

We examine in this paper the training and test set performance of several equity factor models with a dataset of 20 years of data, 1,200 stocks and 100 factors. First, we examine several models to forecast expected returns, which can be used as baselines for more complex models: linear regression, linear regression with an L1 penalty (lasso), constrained linear regression, xgboost and artificial neural networks. Second, we present a unified framework for portfolio construction, leveraging machine learning for the whole pipeline, from the factor data to the portfolio weights, which scales to a large number of assets and predictors. The results we obtain are interesting and non trivial to interpret; non linear models models offer a more balanced outcome considering test set Sharpe ratio and turnover but linear unconstrained models show a good performance in the test set. We introduce a model-free reinforcement learning model, which uses factors to find the portfolio weights maximizing the information ratio