Pattern Stability Analysis in Post-Crisis SME Lending: A Time-Split Classification Framework
Abstract
Concept drift is a continuing challenge in risk modelling for Small and Medium Enterprise (SME) loans, as default patterns change significantly during economic downturns. Most current studies use Random K-Fold validation, which leaks future data into the training and overestimates the performance metrics. To this end, we applied a Time-Split Classification Framework to assess the loan default risk pre- and post-2008 financial crisis in temporally reliable conditions. Our pipeline involved the collection of the U.S. SBA loan dataset (1987-2014), data cleaning, strict chronological splitting, class imbalance correction using SMOTE, model training, and feature importance tracking. We chose Light Gradient Boosting Machine (LightGBM) as the main classifier and benchmarked it against Random Forest and Logistic Regression. Results show that random validation inflates the AUC score by 1.17 percentage points (approximately $1.21 \%$), and LightGBM under time-split evaluation achieved an AUC of 0.9649. Pattern Stability Analysis showed that loan duration and borrower location were the top predictors in both periods, but after the crisis lenders put less emphasis on employee headcount and more on total loan size and projected job creation. These findings establish that developing reliable SME loan default prediction systems requires the validation of chronological models.