Credit Risk Scorecard Best Practices
This guide covers regulatory and industry best practices for developing, validating, and monitoring scorecard models using ScoreCardModel.
Feature Selection
IV Thresholds (Regulatory Standard):
| IV Range | Label | Recommendation |
|---|---|---|
| < 0.02 | Useless | Reject — no predictive power |
| 0.02 – 0.1 | Weak | Accept only with business justification |
| 0.1 – 0.3 | Medium | Good candidate |
| 0.3 – 0.5 | Strong | Excellent candidate |
| > 0.5 | Suspicious | Investigate for data leakage or overfitting |
- Minimum 3 bins per feature for regulatory acceptance
- At least 5% of population in each bin (merge rare bins)
- No pairwise feature correlation above |r| > 0.7
- 8–15 features recommended for a robust scorecard
WOE Requirements
Monotonicity:
- Monotonic or directionally interpretable WOE patterns are preferred by regulators
- Non-monotonic WOE may indicate overfitting or poor binning
- Use check_monotonicity() from the diagnostics module to verify
WOE Ranges: - Cap WOE values at ±3.0 to prevent extreme scores from single bins - WOE of 0 means the bin's risk matches the population average - Positive WOE = higher proportion of goods (lower risk)
Bin Quality:
- No bin should have zero counts of either good or bad events
- Laplace smoothing (adjusted method) handles zero-count bins
- For small datasets, use empirical_logit method (Agresti correction)
Scorecard Construction
PDO and Base Odds:
- Common settings: PDO = 20 (doubling odds adds 20 points), Base Odds = 50:1 at 600 points
- Verify factor calculation: factor = PDO / ln(2)
- Verify offset: offset = Base Points - factor * ln(Base Odds)
Points Distribution: - Feature point ranges should be proportional to their contribution - A feature contributing 50+ points of range is significant - Features with < 10 points range may not justify their inclusion
Model Validation
Holdout Testing: - Minimum 30% of data held out for validation - Time-based out-of-time (OOT) validation required for regulatory submissions - KS on validation set should be within 0.05 of training KS
Stability Monitoring (PSI):
| PSI Range | Interpretation |
|---|---|
| < 0.1 | No significant shift |
| 0.1 – 0.25 | Minor shift — investigate |
| > 0.25 | Major shift — consider retraining |
- Monitor PSI monthly or quarterly
- Compare current score distribution to development distribution
Backtesting: - Compare predicted vs actual default rates quarterly - Hosmer-Lemeshow test for calibration assessment - Track feature-level WOE pattern stability
Regulatory Context (Basel / CCAR)
Basel II/III: - Scorecards must be interpretable (no black-box models) - Model documentation must include: development data, variable definitions, binning rationale, WOE calculations, and validation results - Annual model review required
CCAR (Comprehensive Capital Analysis and Review): - Model use must be clearly defined - Performance monitoring must be ongoing - Model limitations must be documented
FCRA / Reg B (Adverse Action):
- Reason codes must be provided to declined applicants
- The top 2–4 contributing features (in points) should be the reason codes
- Use plot_scorecard_waterfall() to visualize per-feature point contributions
Scorecard Templates
| Template | When to Use |
|---|---|
BaseScorecard |
Standard development with all features |
ConservativeScorecard |
Regulatory submissions, limited feature sets |
| Custom Pipeline | When you need non-LogisticRegression models (use Pipeline) |