An Explainable Machine Learning Framework for Early Warning of Profitability Downside Risk in Listed Companies

Abstract

Profitability deterioration often appears before bankruptcy, default, or formal financial distress. This paper develops an explainable machine learning framework to identify listed companies with high next-year profitability downside risk. Using a public firm-year panel constructed from SEC Financial Statement Data Sets and Stooq historical stock price data, this study examines U.S. non-financial listed companies from 2015 to 2024. High-risk observations are defined as firms whose next-year ROA change falls in the bottom 30 percent within the same industry-year group. Logistic Regression, Random Forest, XGBoost, and LightGBM are evaluated under a chronological validation design. XGBoost performs best in the out-of-sample test set, with an AUC of 0.836 and a PR-AUC of 0.653. SHAP results indicate that revenue growth, operating margin, leverage, ROA, operating cash flow growth, and stock volatility are the main risk drivers.