Risk Analysis:
Lending Club

Loan default prediction under asymmetric costs

In P2P lending, false negatives (approving a loan that defaults) are far costlier than false positives (rejecting a good borrower). Standard accuracy-maximising models ignore this asymmetry. Lenders need models that can be tuned to explicitly weight the cost of default misclassification.

Three model families with cost-sensitive learning

The dataset uses Lending Club origination features: loan grade, DTI ratio, FICO score, income, employment length, and purpose. After exploratory analysis and feature engineering, three model families are trained and compared: L2-regularised logistic regression (interpretable baseline), gradient-boosted trees via XGBoost, and a shallow neural network in TensorFlow.

All models are evaluated using AUC-ROC, precision-recall curves, and expected cost at varying classification thresholds. Class weights and sample weighting are applied to embed asymmetric misclassification costs directly into training — giving lenders a cost-threshold curve they can tune to their risk appetite.

Actionable risk curves across three model types

XGBoost outperforms logistic regression and the neural network on AUC while remaining interpretable via SHAP values. The cost-sensitive framework produces threshold-specific expected cost curves that translate directly into lending policy decisions. The full analysis — methodology, code, and results — is documented in the companion report.

• →Logistic regression, XGBoost, and neural network comparison
• →Cost-sensitive training with asymmetric class weights
• →AUC-ROC, precision-recall, and expected-cost evaluation
• →SHAP feature importance for XGBoost interpretability
• →Threshold-tunable lending policy framework