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

Le résumé fourni par la source

Tax administrations increasingly rely on data-driven risk models to prioritize enforcement and outreach resources, yet machine learning applications to personal income tax arrears remain scarce compared with corporate-tax and fraud-detection contexts. This study introduces TaxMind, a machine learning framework for predicting tax payment arrears, operationalized as escalation to mandatory (forced) collection, comparatively evaluating five classifiers and applying Shapley-value techniques to interpret the resulting model. A dataset of 16,010 administrative tax records was cleaned by removing 5,385 exact duplicate rows (33.6% of raw data), yielding 10,625 records across 9,264 taxpayers; a taxpayer-grouped train/test split prevented information leakage between related observations. Decision Tree, Random Forest, XGBoost, Support Vector Machine, and a Multi-Layer Perceptron were trained on eight demographic and fiscal attributes, with a signed-logarithm transformation correcting skew in the tax-amount field; interpretability was assessed using Shapley Additive Explanations (SHAP). XGBoost achieved the best discrimination (AUC = 0.779, accuracy = 70.2%, F1 = 0.699), outperforming the other models compared. SHAP identified tax amount, tax category, and taxpayer age, a socio-demographic attribute, as dominant predictors, diverging from gain-based rankings. These findings indicate that duplicate contamination and non-grouped validation can distort administrative model evaluations, and that debtor-related attributes carry under-exploited predictive value for tax segmentation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.