Behavioral Financial Signals and Explainable AI in Credit Risk Assessment: Evidence from Utility Arrears and Gambling Activity

Abstract

Traditional consumer credit scoring models frequently fail to identify rising default risks in transition economies because they rely almost exclusively on formal macroeconomic indicators and unchanging demographic data. This paper offers a preliminary empirical assessment of digital gambling participation and utility payment arrears as stand-in risk indicators in Kazakhstani consumer credit scoring. Using a unique survey-derived dataset from Kazakhstan (N = 41 survey respondents, yielding a finalized modeling cohort of N = 233 active borrowers), we develop, tune, and evaluate five predictive models—Logistic Regression (LR), Decision Trees, Artificial Neural Networks (ANNs), Random Forests, and Extreme Gradient Boosting (XGBoost). Models were optimized using stratified 5-fold cross-validation on the training partition (N ≈ 163) and benchmarked across both a SMOTE-balanced evaluation pool (Deval, N = 473) and an empirical 30% holdout test set (Demp, N ≈ 70). The results provide localized baseline evidence on incorporating non-traditional behavioral data into credit evaluation frameworks in data-constrained emerging economies. While the raw empirical cohort (𝒟emp) exhibits a dominant majority distribution with an 88.84% default/late-payment rate, a balanced resampled evaluation pool (𝒟eval, N = 473) was constructed to ensure unbiased model optimization. The empirical results show that tree-based ensemble structures considerably outperform conventional methods, establishing a clear performance hierarchy. With an accuracy of 83.1% and an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.854, the classical LR baseline has the lowest predictive power and is unable to map complex, interactive behavioral hazards. On the other hand, the best architecture is the improved XGBoost ensemble engine, which achieves an AUROC of 0.976 on the balanced evaluation pool (𝒟eval). The XGBoost framework successfully reduces institutional asymmetric loss constraints, according to a micro-structural audit of the underlying confusion matrix. Evaluated on 𝒟eval, the model achieves a specificity of 97.4% (TN = 228) and a sensitivity of 96.2% (TP = 230), reducing catastrophic FNR underwriting errors while preserving high loan onboarding efficiency. Furthermore, comparative testing directly on the unadjusted empirical cohort (𝒟emp) confirms strong diagnostic robustness, demonstrating that high class-separation performance generalizes effectively to skewed real-world credit distributions. These findings demonstrate that incorporating real-time behavioral abnormalities into regularized ensemble pipelines offers a very reliable method of reducing risk. In conclusion, this research provides emerging market financial institutions and regulators with an empirical guide to enhance credit underwriting accuracy, strike a compromise between inclusive lending and capital preservation, and shield the retail banking portfolio from localized default contagion. Although the size of the available dataset does not allow us to attain a definitive conclusion, these results point to a promising direction to further investigate with the same methods proposed in this work and larger datasets. Indeed, they suggest that behavioral data has the potential to enhance traditional credit-scoring techniques when properly vetted and integrated with machine-learning techniques.