Portfolio 3: End-to-End Credit Risk Modelling & Portfolio Risk Analytics (UK)

📎For more details, click here

🎯 Objective

Developed a credit risk model to support lending decisions and reduce default-related losses for a UK private bank to predict Probability of Default (PD) and quantify portfolio risk via LGD, EAD, and Expected Loss (EL) under adverse macroeconomic conditions. The project aims to support risk-based pricing, capital allocation, and regulatory-aligned decision-making under Basel II/III principles.

⚙️ Tasks

• Data Preprocessing & Cleaning: Handled missing values, treated outliers, and standardised numerical features to ensure data quality and model stability.
• Feature Engineering (WoE/IV): Transformed variables using Weight of Evidence (WoE) and selected features based on Information Value (IV) to enhance predictive power and maintain interpretability.
• Data Splitting & Validation: Implemented a train/test/holdout framework to ensure robust model evaluation and prevent data leakage.
• Model Development (Scorecard & Benchmarking): Built an interpretable Logistic Regression scorecard model and benchmarked performance against ensemble methods (e.g., Random Forest, XGBoost).
• Class Imbalance Handling: Applied resampling techniques (ADASYN, SMOTE variants) to improve detection of rare default events.
• Threshold Optimisation: Tuned classification thresholds using ROC analysis and Youden’s Index to balance sensitivity and specificity, prioritising recall.
• Model Evaluation & Validation: Assessed performance using ROC-AUC, Gini coefficient, KS statistic, CAP curve, and confusion matrix metrics to evaluate discrimination and classification effectiveness.
• Risk Quantification (LGD, EAD, EL): Implemented LGD and EAD estimation and calculated Expected Loss (EL = PD × LGD × EAD) to translate model outputs into monetary risk measures.

💡 Results

• Model performance: Logistic Regression achieved ROC-AUC = 0.9524, Gini = 0.9048, and KS = 0.7703, with 92.73% recall, effectively capturing the majority of defaulters.
• Risk drivers: Loan-to-value, interest rate and regularity of income inflows (IV > 0.5) were identified as the primary predictors, outperforming traditional variables such as income level and employment length.
• Business impact: Reduced missed defaults to 554 out of 7,617, significantly limiting loss exposure. CAP analysis showed that ~90% of defaulters were captured within the top 20% of the highest-risk segment, enabling more targeted and efficient underwriting.
• Financial integration: The Expected Loss (EL) framework translated model outputs into monetary terms (e.g., rejecting a small business loan with EL = £39.7k versus approving a home improvement loan with EL = £42), supporting automated credit decision-making and portfolio risk optimisation.