From probability of default to usable risk bands

Problem

Borrower-level risk models are useful only when their outputs can be evaluated and translated into interpretable segments. This project estimates probability of default and then converts the FICO distribution into statistically structured rating bands.

Workflow

Synthetic data → feature inspection → train/test split → logistic PD model → model evaluation → dynamic FICO bucketing → rating validation.

Method

The PD model uses FICO score, debt-to-income, credit lines, and employment tenure in a logistic-regression specification.

FICO boundaries are selected with dynamic programming rather than naive equal-width or equal-frequency bins. Candidate segments are evaluated with log-likelihood subject to a minimum bucket size.

Validation

• Model performance evaluation.
• Monotonicity verification.
• Seeded synthetic-data generation.
• Unit tests for the bucketing algorithm.
• Explicit model and deployment limitations.