RESEARCH · GLOBAL
Classification loss outperforms regression in ML-driven portfolio construction
Researchers compared classification vs. regression across gradient boosting, random forests, and neural networks for stock selection. Classification ensemble achieved Sharpe ratio 1.83 vs. 1.11 for regression—a 65% improvement on out-of-sample testing.
WHY IT MATTERS
Portfolio managers optimizing ML stock-picking pipelines often default to regression; this work validates classification setups and guides hyperparameter tuning for better risk-adjusted returns.
Source: arXiv · 2026-06-24