Working Paper: NBER ID: w30805
Authors: Lin William Cong; Guanhao Feng; Jingyu He; Xin He
Abstract: We develop a new class of tree-based models (P-Trees) for analyzing (unbalanced) panel data using economically guided, global (instead of local) split criteria that guard against overfitting while preserving sparsity and interpretability. To generalize security sorting and better estimate the efficient frontier, we grow a P-Tree top-down to split the cross section of asset returns to construct test assets and re-cover the stochastic discount factor under the mean-variance efficient framework, visualizing (asymmetric) nonlinear interactions among firm characteristics. When applied to U.S. equities, boosted (multi-factor) P-Trees significantly advance the efficient frontier relative to those constructed with established factors and common test assets. P-Tree test assets are diversified and exhibit significant unexplained alphas against benchmark models. The unified P-Tree factors outperform most known observable and latent factor models in pricing cross-sectional returns, delivering transparent and effective trading strategies. Beyond asset pricing, our framework offers a more interpretable and computationally efficient alternative to recent machine learning and AI models for analyzing panel data through goal-oriented, high-dimensional clustering.
Keywords: Efficient Frontier; Panel Data; Asset Pricing; Stochastic Discount Factor
JEL Codes: C1; G11; G12
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| ptree (Y60) | efficient frontier (D61) |
| ptree test assets (Y10) | unexplained alphas (Y50) |
| systematic sorting of firm characteristics through ptree (C38) | performance of asset pricing models (G17) |
| ptree (Y60) | construction of diversified test assets (C52) |
| ptree (Y60) | recovery of stochastic discount factor (G19) |