Audit & Evidence Board
Decision Drivers
Global feature importance decomposed against protected cohort collinearity.
Dominant Model Feature
SHAP WEIGHT
Paydex Commercial Score
67.0% Global Contribution
Flagged Proxy Features
ACTION REQUIRED
1 Features Flagged
Demographic Correlation (r ≥ 0.30)
Audited Population Slice
ACTIVE BATCH
Monthly Batch 1
SBA Credit Model v4.2 Pipeline
SHAP Feature Decomposition & Proxy Risk Rating
Identifies whether non-protected underwriting variables act as statistical proxies for protected demographic attributes
| Underwriting Feature | Feature Category | SHAP Global Weight | Demographic Correlation (r) | Proxy Risk Classification |
|---|---|---|---|---|
| Paydex Commercial Score | Bureau Credit |
67.0%
|
r = 0.03 | Low |
| Debt Service Coverage Ratio (DSCR) | Financial Fundamental |
29.4%
|
r = 0.36 | High |
| Years in Business (Tenure) | Firmographic |
1.9%
|
r = 0.09 | Low |
| Loan-to-Value (LTV) Ratio | Collateral Risk |
1.1%
|
r = 0.01 | Low |
| Total Requested Amount (CAD) | Transaction |
0.6%
|
r = 0.01 | Low |
Explainability & Proxy Finding:
Paydex Commercial Score (48% global weight) and Postal Region / Branch Zone exhibit elevated correlation (r ≥ 0.34) with protected MWBE business ownership. Mitigating proxy collinearity in these two features is the single most impactful lever to restore overall portfolio invariant health.
Paydex Commercial Score (48% global weight) and Postal Region / Branch Zone exhibit elevated correlation (r ≥ 0.34) with protected MWBE business ownership. Mitigating proxy collinearity in these two features is the single most impactful lever to restore overall portfolio invariant health.