High-dimensional portfolio optimization based on tree-structured factor model

Abstract

Certain characteristics are known to predict cross-sectional expected stock returns and risk exposures. We propose a novel portfolio optimization procedure to incorporate multiple characteristic information, which requires minimal parameters and no stringent assumptions. Instead of investing in individual stocks, we perform portfolio optimization on a large panel of characteristic portfolios generated by a tree-structured portfolio sorting method, which can capture the non-linearities and interactions among stock characteristics to help predict expected returns and covariances. Simulations show that our tree-structured Lasso-based mean–variance (MV) strategy has better out-of-sample Sharpe ratios than the three benchmark strategies: the stock-based MV strategy, the stock-based global minimum variance (GMV) strategy, and the double-sorted portfolio-based MV strategy. We use daily stock data from the Chinese A-share market from 2002-04-01 to 2022-12-31 to compare the out-of-sample performance of the tree-structured Lasso-MV with other classical strategies, such as GMV and equally weighted (EW) strategies, etc. The empirical results suggest that the tree-structured Lasso-MV strategy can achieve a higher Sharpe ratio, a smaller standard deviation, and a lower turnover. These results are robust to different levels of granularity.