Using Tree-Based Machine Learning to Detect European Stock Mispricing
Summary
This research summary compares linear valuation models with machine-learning methods for estimating the monthly fundamental value of stocks in 17 European countries. It constructs a mispricing signal from the difference between estimated fair value and observed market value, then tests whether the signal predicts subsequent returns. The models include pooled and benchmark linear regressions, LASSO with accounting variables and their interactions, random forests, gradient-boosted trees, and an ensemble. The sample spans 1993–2019 and uses 21 accounting measures; tree models also incorporate lagged accounting information.
The reported evidence favors tree-based methods: value-weighted, industry-adjusted long-short portfolios produce risk-adjusted returns of 48–66 basis points per month, and machine-learning signals remain significant in Fama–MacBeth regressions where linear signals have little marginal predictive power. LASSO improves on basic linear regression but does not match the tree strategies; combining tree models offers a modest advantage. These findings are specific to the article’s European sample, screening rules, and portfolio construction. The summary gives no implementation costs, and its claims depend on the underlying study’s data and methodology.
Key ideas
- The study estimates equity fair value from accounting data and uses deviations from market value as a return-prediction signal.
- Tree-based models capture nonlinearities and interactions that linear valuation models may miss.
- The reported tree-based strategies outperform linear approaches in risk-adjusted portfolio tests.
- LASSO improves predictive ability over simple linear regression but does not match the tree-based portfolios.
- The evidence covers a screened sample of European stocks and does not establish performance after implementation costs.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.