CreditRisk-Lens: Regulatory-Grade Admission Scoring

📋 Executive Summary

This project delivers a state-of-the-art Admission Model (Application Scorecard) developed to optimize credit underwriting. By leveraging Python-based analytics, the model predicts the Probability of Default (PD) to ensure balanced risk-taking and profitable lending. The pipeline adheres to typical banking governance standards, focusing on transparency, stability, and predictive discrimination.

---

🛠️ Technical Methodology

1. Data Engineering & Feature Selection

The model utilizes 9 core credit features, including income, employment length, and loan intent.

• Categorical Encoding: Implementation of One-Hot Encoding to translate qualitative labels into machine-readable formats.
• Feature Pruning: Dropping variables with low Information Value (IV) to prevent model noise and ensuring no "leakage" variables are used that wouldn't be available at the time of application.

2. The Modeling Framework

The primary "Champion" model is a Logistic Regression, chosen for its transparency and alignment with regulatory requirements for explainability.

The Sigmoid Link Function: $$PD = \frac{1}{1 + e^{-(\beta_0 + \beta_1x_1 + ... + \beta_nx_n)}}$$

3. Separation & Predictive Power

The model was validated using industry-standard metrics to ensure robust separation between "Good" and "Bad" loan populations:

• ROC-AUC: Achieved a stable score of 0.86, indicating excellent discrimination.
• KS Statistic (Kolmogorov-Smirnov): Calculated at 0.56, significantly exceeding the 0.40 benchmark for high-performance retail models.
• Stability Audit: Comparison of Training vs. Testing AUC showed a variance of $<1%$, proving the model generalizes well to unseen data.

---

📈 Credit Scorecard Scaling

To translate raw probabilities into actionable business decisions, the log-odds were scaled into a traditional 300-850 point system using the PDO (Points to Double the Odds) methodology.

Scaling Formula: $$Score = Offset + Factor \times \ln\left(\frac{P(Good)}{P(Bad)}\right)$$

• Target Score: 600 points at 50:1 odds.
• PDO: 20 points (A 20-point increase represents a halving of default risk).

---

🚀 Key Responsibilities Demonstrated

• Model Re-development: Transitioned from raw data to a production-ready scorecard.
• Monitoring & Back-testing: Automated the comparison of predicted PDs against actual default rates.
• Innovation: Utilized advanced visualization (Diverging Histograms) to assist the Risk Committee in setting optimal cut-off thresholds.

---

📂 Project Structure

• data/: Raw and cleaned credit datasets.
• notebooks/: Detailed step-by-step development from EDA to Model Validation.
• src/: Python scripts for data pipeline and scaling logic.
• requirements.txt: Environment dependencies for reproducibility.

💻 Tech Stack

• Language: Python
• Libraries: Scikit-Learn, Pandas, NumPy, Seaborn, SciPy
• Environment: VS Code / Jupyter

---

Developed with a focus on Responsible Lending & Advanced Risk Analytics.