Credit Score Enhancement Using Gradient Boosting Machine in Retail Lending and Credit Evaluation

Credit scoring, a statistical measure of an applicant's likelihood of default based on financial and behavioural information that is available, is the industry staple credit risk management tool of the retail lending sector and one of the key drivers of credit access, pricing and portfolio performance. Traditional logistic regression scorecards, built on credit bureau data based on a credit scoring paradigm formalized in the 1950s, have known accuracy limitations in thin-file populations (those who do not have much credit bureau history) and also have a range of Gini coefficients, which have stagnated around 0.62–0.68 for mainstream retail lending, after decades of scorecard feature engineering improvements. Ensemble methods like Gradient Boosting Machine (GBM) models, which fit sequentially decision trees on the residuals of previous boosting steps with learning rate shrinkage to avoid overfitting, have shown to have better discriminative power in credit risk applications due to their capability of modeling non-linear interactions of features and irregular distribution patterns of the repayment behavior, which are not possible for logistic regression. In this paper, we test the performance of LightGBM, a high-efficiency GBM implementation, for the task of credit default prediction in retail lending using a data including 124,800 loan applications from four retail lenders in two countries (United Kingdom: three lenders, United States: one lender) for the period January 2022 – December 2024. The default rate for the dataset is defined as 90 or more days past due or charge-off within 24 months of origination and is 14.8% (18,470) for the dataset. Seven categories of 58 features are built: traditional credit bureau data (12 features); income and employment (8); banking behaviour (10); loan application characteristics (7); behavioural and digital engagement (9); alternative credit data (8); and macroeconomic context (4). On 20% held-out test set (n = 24,960), LightGBM delivers AUC-ROC of 0.928, Gini coefficient of 0.856 and KS statistic of 0.641, which is 36.3% higher than the AUC-ROC of logistic regression (0.628). The five most important default predictors by SHAP variance, which explain 84.8% of the variance in the model's output, are credit utilisation ratio (22.8% of SHAP variance) 12-month payment history score (19.4%), debt-to-income ratio (16.2%), banking behaviour index (14.6%), and alternative credit data composite (11.8%). Deployed across the four lenders' approval decision workflows, the GBM scoring model reduces portfolio non-performing loan (NPL) ratio from 14.8% to 9.2%-a 37.8% relative reduction-while maintaining a credit approval rate of 72.4%, compared with 68.2% under the logistic regression policy, demonstrating that GBM enhancement expands credit access for well-qualified applicants currently declined by less discriminating models.

Hali tarjima qilinmagan