Predicting Corporate Financial Distress in Pakistan: An Explainable Machine-Learning Early Warning System for KSE-100 Firms

Abstract

The purpose of this study is to explore the use of machine learning (ML) for corporate distress prediction on the KSE-100 listed non-financial firms in Pakistan. We present a comparative study of traditional statistics (logistic regression and discriminant analysis) to advanced machine learning models which are based on random forests, gradient boosting, Sharpley neuaral network using archival financial as well as macroeconomic data during 2005–2023. Distress is measured by Altman Z-scores and firm-specific filings of default, bankruptcy, or restructuring. We find that ML is much better than traditional methods, especially at a longer forecasting horizon. One year ahead predictions achieved AUC values after 0.90, with the highest short-term prediction accuracy obtained by the gradient boosting. The ensemble models and neural networks retained reasonable accuracy (AUC ≈ 0.80) for 3–5-year horizons, whereas the performance of conventional methods plunged. Importance scores analysis identifies the leverage, profitability, liquidity and efficiency ratios as core predictors in models while incorporating macroeconomic variables– credit Gross domestic product (GDP). Growth; Inflation; Exchange rate movement- enhances significance and robustness of predictive model. These results have implications for regulators, firms and investors in emerging markets including the usefulness of ML-based early warning systems. The combination of firm and macro level indicators with interpretable ML frameworks lays the groundwork for active risk management and policy intervention

Keywords: financial distress; early warning system; machine learning; KSE-100; financial ratios; non-financial firms