RESEARCH · GLOBAL
Walk-forward benchmark: XGBoost outperforms LightGBM and logistic regression on SPY same-day directional forecasts
arXiv study backtests XGBoost, Random Forest, LightGBM, and logistic regression on predicting daily SPY direction using only opening-bell data. Walk-forward validation from 1993–2024 shows tree ensembles beat linear models but transaction costs limit profits.
WHY IT MATTERS
Validates ensemble tree methods for short-horizon equity alpha; realistic cost assumptions strip illusion of profits. Useful baseline for trading desks evaluating new ML models.
Source: arXiv q-fin · 2026-08-28