Using Tree-Based Machine Learning to Detect European Stock Mispricing

Summary

This article summarizes a study of machine-learning methods for fundamental equity valuation across 17 European countries. It estimates monthly fair values from 21 accounting variables, then defines a mispricing signal as the gap between model-implied value and market value. The comparison includes two linear regression approaches, LASSO with accounting-variable interactions, random forests, gradient-boosted trees, and an ensemble. The sample covers 1993–2019; the tree models additionally use up to 48 periods of accounting data. Feature importance is examined with SHAP values, and stocks are sorted into quintiles for portfolio and cross-sectional return tests.

The summary reports that tree-based strategies earned industry-adjusted, value-weighted risk-adjusted returns of 48–66 basis points per month. Their signals also remained significant in Fama–MacBeth regressions, whereas linear-model signals contributed little once machine-learning signals were included. LASSO improved on basic linear regression but did not match the tree strategies, and the ensemble added only a small benefit. The evidence is bounded by the study’s European universe, data filters, and portfolio design; the summary does not discuss trading costs or provide enough detail to independently reproduce the tests.

Key ideas

• The study estimates fair value from accounting data and tests whether model-to-market value gaps predict returns.
• Tree models use nonlinear relationships, interactions, and lagged accounting measures.
• The reported tree-based strategies outperform linear regression strategies in portfolio tests.
• LASSO helps relative to simple linear regression but falls short of the tree-based strategies.
• The findings apply to a screened European equity sample, and implementation costs are not addressed in the summary.

Tags

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