Loan Risk Segmentation using LightGBM in Peer-to-Peer Lending Evaluation Systems
Annotatsiya
Peer-to-peer lending platforms intermediate credit risk between individual borrowers and individual investors without the balance-sheet buffer a traditional bank holds against its own lending losses, making accurate, well-calibrated default-risk segmentation directly load-bearing for both sides of the marketplace rather than a secondary risk-management input. LightGBM, a histogram-based gradient boosting framework using leaf-wise tree growth with depth constraints, offers strong tabular-data predictive performance at substantially lower training cost than comparable boosting or ensemble methods, properties well suited to a lending platform's combination of large applicant volume and frequent model retraining needs. This paper applies LightGBM (255 leaves maximum, learning rate 0.05, 800 boosting rounds with early stopping, L1/L2 regularisation) to 248,000 historical peer-to-peer loans (39,680 defaults; 16.0% base default rate) spanning six platform-assigned risk grades (A: 49,600; B: 62,000; C: 57,040; D: 39,680; E: 24,800; F: 14,880), using 42 borrower, loan, and macroeconomic features. On a held-out test set, LightGBM achieves$\mathbf{A U C}=\mathbf{0. 8 4 4}$and Kolmogorov-Smirnov statistic = 62.4, exceeding logistic regression (AUC 0.724; +16.6% relative), the platform's existing credit-score-only baseline (0.684; +23.4%), and a Random Forest benchmark (0.804; +5.0%), with debt-to-income ratio, credit utilisation, and payment history length jointly accounting for 49.4% of total feature importance. Risk-tier segmentation derived from the model's predicted default probability-Low Risk (74,400 loans; 4.2% realised default), Moderate Risk (99,200; 12.4%), High Risk (49,600; 24.8%), and Critical Risk (24,800; 48.2%)-reproduces the portfolio's overall 16.0% default rate under population weighting, confirming internal consistency between tier-level and aggregate outcomes. Utilized in an approval sorting context for automation purposes, the segmentation process using LightGBM algorithm helped 80.0% of applicants (198,400 of 248,000) get past the underwriters' review, which reduced average application processing time by half, from 4.8 to 2.4 hours (-50.0%), while reducing the realized portfolio-wide default rate from an 18.4% pre-deployment baseline to 12.8% post-deployment (-30.4%). The analysis conducted to check the fairness of the model revealed that the approval rate ratio between the protected sub-group of population and the reference group equals 0.809 (just within the range allowed by the four-fifths rule, but by very little), as well as 0.42 correlation between the top features of the model and an income proxy variable, requiring further monitoring rather than one-time clearance.
Hali tarjima qilinmagan