Machine learning portfolios with equal risk contributions: evidence from the Brazilian market

Abstract

We use machine learning methods to forecast individual stock returns in the Brazilian stock market, using a unique data set including technical and fundamental predictors. We find that portfolios formed on the highest quintile of predicted returns significantly outperform market benchmarks. However, portfolios formed on the lowest quintile of predicted returns earn positive returns and have high volatilities, making traditional long-short strategies unnatractive. To resolve this problem, we propose an equal risk contribution (ERC) ensemble approach to build a portfolio combining long-short portfolios obtained with various machine learning methods such that (i) the risk contributions of all individual long-short portfolios are equal, and (ii) the aggregate risk contribution of all long positions equals that of all short positions. The ERC ensemble portfolio outperforms, on an after cost, risk-adjusted basis, all individual machine learning long-short portfolios, as well as equally-weighted ensembles of these portfolios.