Definition
Measures the degree to which AI-driven pricing produces systematically different premium outcomes across demographic groups after controlling for actuarially justified risk factors, revealing proxy discrimination that alternative data features may introduce.
The Measurement Gap
AI pricing models incorporate alternative data features (credit-based scores, consumer behavior, social media signals, IoT data) that may correlate with protected class membership through geographic, economic, or behavioral proxies. A model can produce actuarially sound aggregate results while systematically overcharging or undercharging specific demographic groups. Legacy metrics cannot detect this. Loss ratio measures aggregate profitability, not whether that profitability is evenly distributed across demographic segments. Combined ratio captures expense and loss efficiency without any fairness dimension. Complaint ratio tracks explicit dissatisfaction but misses policyholders who never complain because they do not know they are being overcharged relative to their risk. Retention rate measures who stays and who leaves, but cannot distinguish actuarially justified attrition from price-driven disparate impact. Rate filing approval confirms compliance with filed factors, not whether the model's output pattern produces disparate outcomes through unfiled proxy effects. None of these metrics operate in the dimension of cross-group outcome comparison after controlling for legitimate risk factors.
Decision Supported:
Are the alternative data features in the pricing model producing outcomes that would fail a disparate impact analysis under state insurance regulatory standards, and if so, which features are the primary drivers of the disparity?
Formula
Plain language: For each protected class proxy group (defined by census tract demographics, since individual protected class data is not collected), compare the average model-predicted premium to the average actuarially expected premium based on traditional rating factors only. The gap, if it varies systematically by demographic proxy, is the disparate impact signal.
Notation:
PDII_g = (P_model_g - P_actuarial_g) / P_actuarial_g
PDII = max|PDII_g| across all demographic proxy groups g
Variables:
| Variable | Definition | Unit |
|---|---|---|
| P_model_g | Average AI-model-predicted premium for policies in demographic proxy group g | Dollars |
| P_actuarial_g | Average premium using only traditional actuarially justified rating factors for the same policies | Dollars |
| g | Demographic proxy group (defined by census tract income quintile, racial/ethnic composition quartile, or similar geographic proxy) | Classification |
Report both the maximum absolute PDII across groups and the group-specific values. Positive PDII_g indicates group g pays more under the AI model than traditional factors would predict.
Proposed formula, pending validation pilots.