Using Boosted Trees to Extract Nonlinear Alpha from Analyst Expectations

Summary

This report studies whether boosted tree models can extract stock return signals from analyst expectation factors that linear models may miss. The authors note that these factors can be complex, have substantial missing data, and correlate with earnings and market capitalization factors. Their approach uses rolling models with both linear and boosted tree methods, alongside statistical adjustments to reduce correlations with those established factors.

The analysis covers Chinese stocks from early 2007 through October 2020. The report says tree model predictions had an average correlation of 10% with a baseline model built from earnings and size factors, while predictions from the tree and linear models correlated by 50%. The tree approach was more suitable for smaller and mid-sized stocks, while the linear model fit larger stocks better. Combining the two models equally outperformed using the linear model alone, and added returns in CSI 300 and CSI 500 enhancement tests. These are historical results; the summary does not specify transaction costs or establish that the patterns persist out of sample.

Key ideas

• Boosted trees can model nonlinear relationships in analyst expectation factors and can accommodate missing inputs.
• The study uses rolling forecasts and adjusts predictions to reduce their correlation with earnings and size factors.
• Tree model signals were less correlated with the baseline than with linear model signals.
• The report finds different size profiles for tree and linear predictions and improved historical portfolio results from combining them.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.