What are the main credit scoring model approaches and how does logistic regression compare to machine learning methods?

Credit scoring models assign a numerical score representing the probability that a borrower will default. There are several major approaches, each with distinct trade-offs:

1. Logistic Regression (Traditional Workhorse)

This is the most widely used model in banking. It estimates the probability of default (PD) as:

PD = 1 / (1 + e^(-z))

where z = β₀ + β₁X₁ + β₂X₂ + ... + βₙXₙ

Advantages: Transparent, easy to explain to regulators, coefficients have direct interpretation (e.g., each $10,000 increase in income reduces log-odds of default by 0.15).

2. Decision Trees / Random Forests

These partition the feature space into regions. A random forest aggregates hundreds of trees to reduce overfitting.

3. Neural Networks / Gradient Boosting

These capture complex nonlinear interactions but act as "black boxes."

Practical Example: Pinnacle Bank builds two PD models for its SME portfolio. The logistic regression achieves 72% AUC (area under ROC curve), while a gradient boosting model achieves 81% AUC. However, the regulator requires model explainability under SR 11-7 guidance. Pinnacle uses the logistic model for regulatory capital and the ML model internally for screening.

Key Exam Takeaway: For FRM, understand that model validation metrics (AUC, Gini coefficient, accuracy ratio) are critical for comparing models, and that regulators generally prefer interpretable models for capital calculations.

Check out our FRM Part II Credit Risk course for more on model validation techniques.