Comparative Evaluation of Machine Learning Approaches for Minority-Class Financial Distress Prediction Under Class Imbalance Constraints

This record provides the reproducibility package for the preprint "Comparative Evaluation of Machine Learning Approaches for Minority-Class Financial Distress Prediction Under Class Imbalance Constraints" by Karan Sehgal and Khawar Naveed Bhatti (arXiv:2605.14067). The package contains a single-file Python script that executes the full pipeline end-to-end on the public Taiwan Bankruptcy Prediction dataset (UCI Machine Learning Repository ID 572, CC BY 4.0): median imputation, standard scaling, stratified 80/20 split, SMOTE oversampling on the training partition only, training of six classifiers (Logistic Regression, Random Forest, AdaBoost, XGBoost, LightGBM, CatBoost) with seed 42, evaluation on the held-out test partition, and TreeSHAP attribution-stability measurement (K=50 rotated backgrounds) on the trained XGBoost model. The deposit includes all measured output artefacts: per-model classification metrics, confusion-matrix decomposition, TreeSHAP attribution stability summary (sigma_SHAP and per-feature variance), and a full environment fingerprint (library versions, dataset SHA-256, fixed random seed). Researchers can run the script on the public dataset and produce results identical to those reported here within numerical tolerance. This package also underpins a separate governance-trade-off analysis in our IEEE Cyber-AI 2026 submission "Q-MDP: Superposition-Aware State Orchestration..." which characterises the trade-off between minority-class recall and SHAP attribution stability under SMOTE and architecture choice. Released under CC BY 4.0 to support open, reproducible research into imbalance-aware classification and governance-oriented ML deployment.