A Machine Learning Framework for Predicting Tax Payment Arrears: Comparative Model Evaluation and Shapley-Value Interpretability

Tax administrations increasingly use data-driven risk models to prioritize collection resources, yet machine learning applications to personal income tax (PIT) arrears remain limited. This study introduces TaxMind (v1.0), an interpretable risk management framework for predicting whether PIT obligations will progress to mandatory collection and for integrating debtor- and debt-related information into risk-based segmentation. Administrative records from the Lebanese Tax Administration were analyzed; after removing 5385 exact duplicates from 16,010 records, the final dataset contained 10,625 obligations across 9264 taxpayers. Five tuned classifiers were evaluated using a taxpayer-grouped train/test design, with SHAP used for model interpretation. XGBoost achieved the highest observed discrimination (ROC-AUC = 0.784; accuracy = 0.708; F1-score = 0.704), closely followed by Random Forest (ROC-AUC = 0.780); the leading tree-based models substantially outperformed logistic regression benchmarks. SHAP identified Total Tax Amount, Tax Category 2, and Age as the three leading individual encoded features, while the Age contribution was nonlinear and varied across Tax Categories. The principal SHAP ranking was highly consistent under Random Forest. Top-decile ranking by predicted probability captured only 0.4% of the monetary exposure of realized mandatory collection cases, versus 92.7% under probability-weighted exposure ranking. These findings show that debt-related information remains central, but selected debtor characteristics add predictive value, supporting TaxMind as a model-agnostic early-warning framework for preventive tax debt management.