Machine Learning model for Enhancing Small Business Credit Risk Assessment and Economic Inclusion in the United State

Abstract

The lack of accessible credit is a constitutive constraint on small businesses in the US. This paper proposes, evaluates, and establishes an interpretable, bias-conscious machine learning approach for small business credit risk assessment. Based on anonymized application, repayment, and organizational operational data, we compare gradient boost, regularized generalized linear models, and tree-based learning to industry-leading scorecards, leveraging monotonic constraints, fairness-conscious weight adjustments, and a SHAP explanation layer. The research hypothesis is to validate whether machine learning systems can strengthen default AUC/KS performance while decreasing disparities in group error rates, along with increasing approval rates at equal risk. Some uplift assessment measures incremental safe approvals, as well as expected loss subject to constrained decision rules. For more comprehensive implementation, the research includes model cards, feature management, WOE/IV, feature stability, as well as champion-champion comparisons. The findings of this research confirm the hypothesis, suggesting interpretable machine learning can achieve higher levels of risk differentiation (∅AUC > X), significantly close error gaps (∅gap > Y%), and achieve inclusivity gains at equal portfolio loss. The research aims to contribute a reproducible workflow, a set of metrics, as well as evidentiary validation of the applicability of transparent machine learning in credit markets.