📉 When Lower Delinquencies Don’t Lower Risk: A PD Modeling Case Study

In credit risk modeling, intuition can be misleading. Recently, I encountered a scenario where Probability of Default (PD) increased across a loan portfolio — even though delinquency counts had dropped. This counter-intuitive outcome prompted a deeper investigation into model behavior, feature design, and the subtle signals embedded in historical data.

🧩 Problem Statement: Unexpected PD Uplifts Despite Lower Delinquency Counts

A modeling issue was observed where PD uplift occurred despite improved delinquency metrics. This anomaly raised questions about the model’s sensitivity and logic.

Key Observations:

• PD values increased even as delinquency counts declined.
• The model responded more to severity than frequency of delinquency.
• Severity was captured in features like historical maximum DP and average DP, which were not normalized.
• These features retained past delinquency intensity, leading to lag effects in PD predictions.

🛠️ Solution Evaluation Framework

To address the issue, I applied a structured evaluation approach:

1. Evaluate the Solution

• Is the model functioning as intended?
• Does it solve the core problem?
• Is it aligned with business or regulatory needs?

2. Measure Performance

• Use metrics and KPIs (e.g., PD drift, Gini coefficient, stability index).
• Compare actual vs. expected outcomes across borrower segments.

3. Assess Limitations

• Identify constraints like feature lag, lack of normalization, or outdated severity signals.
• Understand how these impact model effectiveness and fairness.

4. Recommend Improvements

• Normalize severity features to reduce lag effects.
• Rebalance feature importance to reflect current borrower behavior.
• Explore alternative modeling techniques (e.g., time-decayed features, ensemble models).

5. Validate the Solution

• Confirm alignment with business logic and compliance standards.
• Ensure stakeholder buy-in through transparent evaluation and documentation.

🧠 Conclusion

The investigation revealed that PD models are highly sensitive to severity, and without proper normalization, they can misrepresent current risk levels. This case underscores the importance of:

• Feature engineering discipline
• Lag effect mitigation
• Continuous model validation

In credit risk, fewer delinquencies don’t always mean lower risk — context, history, and feature design all play a role.