Hybridizing Structural Credit Risk and Machine Learning for Corporate Distress Prediction: Evidence from Indonesian Non-Financial Public Firms

Abstract

This study develops an explainable early-warning framework for predicting corporate distress among non-financial firms listed on the Indonesia Stock Exchange. Using a firm-month panel of 107,448 observations from 2014 to 2024, the study constructs a 12-month forward distress label based on PKPU and bankruptcy events. The analysis compares Logistic Regression, Random Forest, XGBoost, and a hybrid XGBoost model incorporating Merton-based structural indicators, evaluated using ROC-AUC, PR-AUC, precision, recall, and F1-score under a time-based split. The results show that tree-ensemble models outperform Logistic Regression, with XGBoost achieving the strongest standalone rare-event performance, including PR-AUC of 0.151 and F1-score of 0.217. Adding Merton structural indicators does not improve aggregate ROC-AUC or PR-AUC, but improves recall and F1-score, indicating incremental detection value at the operational threshold. SHAP analysis shows that distress predictions are mainly driven by solvency, leverage, retained earnings, debt-servicing capacity, profitability, asset structure, and market signals. The model captures 66.7% of distress events with an average lead time of 8.5 months. The study contributes an interpretable hybrid framework for corporate distress early warning in an emerging-market setting.