Explainable Machine Learning for Corporate Financial Distress Classification: Evidence from Albanian Enterprises Receiving the Largest State Treasury Payments

Introduction: This study examines whether leakage-controlled and explainable machine-learning models can classify an Altman Z″-based financial-distress benchmark among Albanian enterprises receiving the largest aggregate State Treasury payments in 2024. The target is treated as an accounting-risk screen rather than an observed bankruptcy outcome.

Methods: The fixed database contains 100 enterprises, with a computable binary target for 96 firms, including 18 classified as distressed. Direct Altman components are excluded from the predictors. Penalized Logistic Regression, Random Forest, Gradient Boosting, and Support Vector Machine are evaluated through repeated stratified five-fold cross-validation with three repeats. Imputation, missing-value indicators, and scaling are fitted exclusively within the training folds, and PR-AUC is the primary evaluation metric.

Results: Random Forest and Gradient Boosting form the leading group. Gradient Boosting achieves a PR-AUC of 0.943 and balanced accuracy of 0.918, while Random Forest records a PR-AUC of 0.943, ROC-AUC of 0.986, and precision of 0.880. Support Vector Machine attains the highest recall but generates substantially more false positives. Held-out permutation importance identifies the Current Ratio as the dominant predictor, whereas Treasury-payment variables provide little stable incremental information after the accounting indicators are included.

Conclusions: Constrained tree ensembles reproduce the accounting benchmark with strong internal held-out discrimination. However, the small number of distress cases, the contemporaneous cross-sectional design, the retained-earnings proxy, and the conceptual proximity of liquidity ratios to the Altman target limit the scope of inference. The models may support preliminary professional screening, but they do not establish future bankruptcy prediction, causality, or external validity. Temporal validation against observed outcomes is required before operational use.