Credit Risk Assessment Using Alternative Financial and Behavioral Data

Abstract

Credit scoring draws on transaction records, bill payments, and digitally observed behavior to evaluate applicants whose conventional credit files are sparse. Whether these signals improve lending decisions without creating new forms of exclusion remains unresolved. This study synthesizes pre-2023 research on alternative financial and behavioral data and develops a transparent, simulation-based comparison of conventional and augmented credit-risk models. Thirty-five peer-reviewed journal articles inform the synthesis; a separately identified conference proceeding provides context on small-business finance in Bangladesh. The demonstration generates 24,000 synthetic borrower records with six conventional predictors, eight consent-dependent alternative indicators, and an explicitly specified default mechanism. Four classifiers are evaluated on held-out cases using discrimination, precision-recall, calibration-sensitive loss, and operational screening measures. In the simulation, augmenting logistic regression increases receiver operating characteristic area from 0.740 to 0.810 and average precision from 0.387 to 0.521; random forests and gradient boosting achieve areas of 0.796 and 0.803, respectively. Shapley-based explanations identify bill-payment regularity, inflow volatility, and debt burden as prominent simulated predictors. These figures describe a controlled demonstration, not observed improvements in real lending portfolios, because the outcome mechanism deliberately contains behavioral information. The reviewed field evidence nonetheless indicates that properly governed alternative data can complement formal credit histories, particularly for thin-file applicants. A viable implementation therefore requires temporally valid measurement, consent and minimization, external validation, probability calibration, group-level fairness audits, and accessible adverse-action explanations. The contribution is an evaluation design that separates predictive feasibility from credible claims of financial inclusion, while making the assumptions behind the numerical illustration explicit.