An Explainable Ensemble Statistical Learning Framework for Early Prediction of Corporate Financial Distress

Al-Juboor, Fatema S. i; Sabri, Husam

doi:10.22068/ijiepr.37.4.2789

IN PRESS Back to the articles list | Back to browse issues page

‎ 10.22068/ijiepr.37.4.2789

An Explainable Ensemble Statistical Learning Framework for Early Prediction of Corporate Financial Distress

Fatema S. i Al-Juboor¹

, Husam Sabri²

1- University of Information Technology and Communication, Iraq
2- University of Baghdad, Iraq , hossam.mostafa.coart.uobaghdad.edu.iq@gmail.com

Abstract: (20 Views)

Corporate financial distress prediction is an important decision-support problem for investors, regulators, creditors, and corporate managers. Existing financial distress models often provide strong predictive performance but may suffer from limited interpretability, inadequate treatment of longitudinal firm-level observations, information leakage during preprocessing, and insufficient uncertainty assessment. This study proposes an Explainable Ensemble Statistical Learning Framework (EESLF) that integrates statistical feature engineering, survival analysis, nonlinear ensemble learning, Bayesian probability estimation, and explainable artificial intelligence within a leakage-controlled predictive architecture. The framework uses liquidity, leverage, profitability, cash-flow, and market-related financial indicators. PCA is employed to assess dimensional redundancy, while LASSO, Elastic Net, and mutual-information analysis identify interpretable original financial variables. Cox proportional hazards modelling provides time-to-distress information, while Random Forest, XGBoost, and Bayesian logistic regression capture complementary nonlinear and probabilistic risk patterns. Their outputs are transformed to a common probability scale and combined using weights estimated exclusively from training data. To address the longitudinal nature of firm-year observations, the revised evaluation uses chronological out-of-time validation, with preprocessing, feature selection, imbalance correction, hyperparameter tuning, and ensemble-weight optimization performed independently within the training periods. SHAP and partial-dependence analyses provide model-based explanations and assess explanation stability across evaluation periods. The dataset contains 10,936 firm-year observations from 1,284 publicly listed companies covering 2012–2024. The results are evaluated using discrimination, calibration, uncertainty, and survival-specific performance measures. The framework provides a structured approach to trustworthy early-warning analysis while recognizing that SHAP-based explanations represent predictive attribution rather than causal inference and that independent external validation remains necessary.

Keywords: Bayesian Logistic Regression, Corporate Financial Distress, Ensemble Learning, Explainable Artificial Intelligence, Survival Analysis.

Full-Text [PDF 836 kb] (7 Downloads)

Type of Study: Research | Subject: Material Managment
Received: 2026/08/9 | Accepted: 2026/08/24