What question did this study set out to answer?

The aim is to predict and explain default on SBA 7(a) loans by utilizing detailed loan data and advanced statistical methods.

July 12, 2026International Journal of Computer Science and Artificial Intelligence0 citationsOpen Access

Predicting Small Business Loan Default Risk Using U.S. SBA 7(a) FOIA Data

SKShivanand R Koppalkar

Key Points

The aim is to predict and explain default on SBA 7(a) loans by utilizing detailed loan data and advanced statistical methods.
Utilized data from SBA FOIA public portal, consisting of 132,459 loans with an 8.84% default rate.
Applied five inferential statistical tests and five supervised classifiers to analyze the data.
Pipeline executed in Python using libraries like pandas and scikit-learn for data processing and model development.
LightGBM model achieves AUC-ROC of 0.9703 and PR-AUC of 0.8420, exceeding a baseline of 8.84%.
All models performed above 0.93 AUC-ROC, with up to 92% recall on the rare default class.
Cross-state loans show a default rate nearly double that of same-state loans.

Abstract

Small businesses sustain nearly half of U.S. private-sector employment, yet a meaningful share of the loans that fund them end in charge-off, and conventional underwriting rarely captures the nonlinear ties among industry, loan structure, and geography that drive that risk. This study predicts and explains default on U.S. Small Business Administration (SBA) 7(a) guaranteed loans using the structural fields the agency itself records, and translates the findings into pricing, oversight, and program-design guidance for lenders, SBA program managers, and policymakers. The work pairs five inferential statistical tests with five tailored supervised classifiers, one pair per research question, so that each hypothesis earns both inferential and predictive evidence. Statistical methods span chi-square with Cramer V, logistic regression with likelihood-ratio and ANOVA tests, a two-proportion z-test, and a Cox proportional-hazards survival model; the machine learning line-up adds XGBoost, LightGBM, Random Forest, and a multilayer perceptron neural network. Data come from the SBA FOIA public portal: two loan-level 7(a) extracts totalling 903,617 rows and 44 columns covering FY2010 through FY2025. After harmonization, duplicate removal, outlier winsorization, and a final-status plus FY2018 filter, the labeled working sample holds 132,459 loans at an 8.84 percent default rate, with six engineered features (NAICS sector, term and loan-size buckets, guarantee percentage, cross-state flag, and disbursement lag). The pipeline runs in Python on CPU using pandas, NumPy, scikit-learn, XGBoost, LightGBM, statsmodels, SciPy, lifelines, and seaborn. On the held-out test partition the LightGBM rate-structure model leads with AUC-ROC of 0.9703 and PR-AUC of 0.8420, far above the 8.84 percent prevalence baseline, while every model exceeds 0.93 AUC-ROC and reaches up to 92 percent recall on the rare default class. Four of five nulls are rejected; cross-state loans default at roughly twice the same-state rate. The calibrated risk insights can be embedded in lender loan-pricing and approval workflows and in SBA portfolio-monitoring dashboards as a complement to traditional credit underwriting. Keywords: machine learning; LightGBM; XGBoost; SBA 7(a) Loans; Credit Default Risk; Gradient Boosting; Class Imbalance; PR-AUC; FOIA Loan-level data

AI에게 질문

Bookmark

View Full Paper