Boosting portfolio choice in the Big Data Era

Markowitz’s mean-variance portfolio optimization is either inefficient or impossible when the number of assets becomes relatively large. To overcome this difficulty, we propose several component-wise boosting learning methods that, in a linear regression specification, can iteratively select the assets (variables) with the largest contribution to the fit from a huge number of assets. Based on dataset consisting of 897 assets with 624 observations obtained from Ken French data library, we assess the performance of tangency portfolios estimated using our methods. We find that our methods substantially outperform the 1/N portfolio in terms of various popular metrics. For example, our component-wise LogitBoost can reach an out-of-sample Sharpe ratio of 1.03, while the 1/N portfolio only achieves a Sharpe ratio of 0.27