The virtue of complexity in machine learning portfolios

Contrary to conventional wisdom in nance, return prediction R2 and optimal portfolio Sharpe ratio generally increase with model parameterization, even when minimal regularization is used. We theoretically characterize the behavior of return prediction models in the high complexity regime, i.e. when the number of parameters exceeds the number of observations. Empirically, we document this "virtue of complexity" in US equity market prediction. High complexity models deliver economically large and statistically significant out-of-sample portfolio gains relative to simpler models, due in large part to their remarkable ability to predict recessions