CAN MACHINE LEARNING BETTER PREDICT ACCOUNTING-BASED FIRM PERFORMANCE IN THE POST-ESG ERA?

Abstract

This study aims to examine whether the rise of the ESG era improves the predictability of firm 
performance. We focus on a specific point in time when the number of firms with ESG scores reported 
by ASSET4ESG increases exponentially, which we define as the ESG era. This study focuses on the 
Malaysian context, which exhibits a clear distinction of increase number of firms with ESG scores since 
2020 compared to the other countries. Accordingly, we employ machine learning (ML) models 
(Random Forest, Gradient Boosting Machines, and XGBoost), and compare them with linear models 
(OLS and LASSO) to evaluate their predictive performance for firm performance in ESG era (2020-2024) 
relative to the pre-ESG era period (2001–2019). Our findings show that ML models achieve 10-14% 
higher predictability for ROA compared to linear models during the ESG era. When ESG-related 
variables are incorporated into the independent variable matrix, the predictive performance of ML 
models further improves to 19-25%. Using SHAP analysis, we find that the Altman Z-score, sales 
efficiency, and industry growth exert the strongest influence on firm performance, while the ESG 
combined score ranks below these financial indicators. This suggests that ESG contributes less directly 
to firm performance, but its interaction with financial variables enhances the overall predictability of 
firm performance.