THE U-SHAPED ESG PERFORMANCE PARADIGM: GRADIENT-BOOSTED MACHINE LEARNING EVIDENCE CHALLENGING SUSTAINABILITY-FINANCIAL ORTHODOXY

Abstract

ESG policy is often implemented as if lower ESG risk should automatically produce better equity outcomes. I test that premise in U.S. large-cap stocks with a five-year panel of 500 S&P 500 constituents and Sustainalytics ESG Risk Ratings frozen at January 2019. Firms are sorted into sector-normalized Low, Medium, and High ESG-risk cohorts (167, 166, and 167 firms), which makes it possible to ask whether ESG is priced along a straight line or across distinct portfolio roles.

Prediction is handled with a Light Gradient Boosting Machine (Light GBM) and then translated into trades under institutional constraints. Out-of-sample validation uses five-fold expanding-window cross-validation. Precision reaches 56.15%, recall 78.84%, and F1-beta 74.19% with beta = 0.7. The backtest assumes $10 million of starting capital, $1 million per position, a 0.5 trade probability threshold, 5% take-profit, a 20-day maximum holding period, and execution frictions.

What emerges is not a monotonic ranking but a U-shaped structure. High-risk firms generate the strongest absolute and risk-adjusted outcomes (240.35% cumulative return; 28.33% annualized return; Sharpe 1.19; Calmar 0.84). Low-risk firms provide the strongest downside protection, with maximum drawdown of 29.08%. Medium-risk firms do not lead on either return or protection. Pairwise tests, non-parametric tests, bootstrap confidence intervals, multiple-comparison correction, and sector and regime checks all support the same ordering.

The implication is narrower and more useful than a blanket ESG verdict. In this sample, ESG works better as a segmented risk signal than as a single quality ladder. The proposed Governance-Driven Risk Premium remains only an interpretation of the High-risk result; this design does not identify it causally.