Colorado SB 21-169 (the Algorithmic Fairness in Insurance Act) is the binding 2026 law that turns "unfair discrimination" from a normative aspiration into a quantitative testing requirement. Colorado Regulation 10-1-1 implements SB 21-169 for personal-auto pricing models filed after 2024, with phased expansion to homeowners (Phase 2 filed in late 2025 with the Oct 15, 2025 expansion) and life and annuity (Phase 3 in 2026), and a first compliance report due July 1, 2026. NY DFS Circular Letter No. 1 (2019) and Circular Letter 2024-7 (July 11, 2024) impose proxy-test expectations on New York filings. Connecticut Bulletin MC-25-8, California's SB 1058-driven CDI guidance, and Nevada Bulletin 24-006 form the broader 2026 fairness framework. This lesson walks the quantitative bias-detection methodology that survives a SERFF examiner's question in 2026: disparate-impact metrics with named protected-class proxies, Bayesian Improved Surname Geocoding (BISG) cautions, geographic-proxy detection, what a defensible bias test exhibit contains, and the explainability artifacts (SHAP, PDP, ALE) for high-risk pricing models. Worked DI-ratio computation with real numbers. The headline rule: a carrier filing a pricing model in Colorado, New York, Connecticut, or California in 2026 produces a quantitative bias-test exhibit per Reg 10-1-1 Phase 2 (homeowners) or Phase 1 (auto) standards, or the model does not approve - and the eight aggressive-fairness DOIs (CO, NY, CA, CT, WA, MA, NJ, IL) are all queued to demand the same exhibit by 2027.

Colorado SB 21-169 and Reg 10-1-1 Mechanics

SB 21-169 (signed 2021, effective 2023) requires insurers to test "external consumer data and information sources" and "algorithms and predictive models" for unfair discrimination against protected classes (race, color, national origin, religion, sex, sexual orientation, disability, gender identity). The statute prohibits unfair discrimination from these algorithms and authorizes the Colorado DOI to adopt regulations specifying testing requirements. Reg 10-1-1 (effective for personal-auto pricing post-2024) implements SB 21-169 with quantitative testing requirements: protected-class proxy testing via BISG; disparate-impact ratio measurement against the four-fifths threshold (0.80); explainability artifacts (SHAP, PDP, ALE); geographic-proxy testing; sensitivity analysis; documentation in a fairness exhibit attached to SERFF filings.

Phased expansion 2024-2026. Phase 1: personal-auto pricing models filed after 2024 (effective for new filings 2024 forward). Phase 2: homeowners pricing models filed after late 2025 (the Oct 15, 2025 expansion brought homeowners into scope). Phase 3: life and annuity underwriting and pricing models filed after 2026 (mid-2026 rulemaking expected). Phase 4 (proposed): commercial pricing models. The expansion sequence reflects DOI bandwidth and methodological complexity - commercial lines pose the most testing-methodology challenges because protected-class proxies are designed for personal-lines applicants, not commercial entities.

The July 1, 2026 first compliance report. Every Colorado-filing carrier must submit a first compliance report to the Division of Insurance by July 1, 2026, documenting algorithm inventory, bias-testing methodology, results, and remediation steps for any flagged disparities. The report's structure follows Reg 10-1-1's compliance-report template. The DOI staff reviews submissions and may request additional analysis, demonstrations, or remediation actions.

The Reg 10-1-1 algorithm inventory. Each algorithm used in pricing or underwriting decisions is registered: model name, version, purpose, training data, features, performance metrics, fairness-testing results, monitoring methodology, retirement criteria. The inventory is the disclosure mechanism - consumer-facing AI disclosure (per the consumer-disclosure lesson) references the algorithm registry; the DOI examiner pulls from the inventory; the AI governance committee maintains.

BISG and Its Cautions

Bayesian Improved Surname Geocoding (BISG) is the 2026 industry-standard proxy methodology where carriers do not collect protected-class data directly. BISG produces a probabilistic vector per policy across White, Black, Hispanic, Asian, Native American, Pacific Islander, Other. Source: surname distribution from the US Census Bureau Surname Distribution File plus census-block geocoding from the policy address plus a Bayesian update combining the two signals. Alternative: BIFSG (Bayesian Improved First-name Surname Geocoding) adds first-name signal where collected. The BIFSG implementation increases accuracy 4-7% on Hispanic and Asian groups; carriers using BIFSG must document the methodology and the validation in the bias-test exhibit.

BISG cautions in 2026.

(1) BISG is probabilistic; no policy has a 100% protected-class assignment. The methodology produces a probability vector - for example, a policyholder may carry probabilities of 0.62 White, 0.18 Hispanic, 0.12 Asian, 0.04 Black, 0.04 Other. The disparate-impact calculation weights by these probabilities rather than assigning each policy to a single group.

(2) BISG accuracy varies by group: 92-96% accurate for White and Black populations; 85-89% for Hispanic (where surname signal is strong for some sub-populations and weak for others); 78-85% for Asian (where surname signal weakens further across diverse East Asian, South Asian, and Southeast Asian sub-populations); below 70% for Native American and Pacific Islander (where sample size and surname-distribution challenges undermine the methodology). The lower accuracy on Native American and Pacific Islander populations means that disparate-impact analysis on these groups should be treated with corresponding skepticism.

(3) BISG can encode geographic-segregation patterns. The census-block geocoding component picks up residential segregation; using BISG to demonstrate fairness while the model relies on geographic features creates a circular logic that NY DFS flagged explicitly in 2025 guidance. The mitigation is geographic-proxy testing alongside BISG-based disparate-impact testing.

(4) BISG requires validation against the carrier's own data when possible. Insurance carriers do not collect race directly but can validate BISG outputs against partial samples (self-reported in voluntary surveys, claim-handling data where ethnicity is occasionally captured in third-party-claimant contexts). Validation against external benchmarks (other carriers' published validations, academic studies) is also useful. The 2026 best-practice bias-test exhibit includes BISG validation documentation.

Disparate-Impact Ratio Computation

The DI ratio is the central bias-test metric. Formula: DI = avg_predicted_outcome(protected_group) / avg_predicted_outcome(reference_group). The EEOC four-fifths rule (under Title VII employment-discrimination jurisprudence, adapted for insurance pricing): a DI ratio below 0.80 is presumptively problematic and requires either feature-set adjustment, model-form constraint, or documented business-necessity defense. The personal-auto BI severity rebuild from the pricing lesson produced DI ratios: 0.91 (Black vs. White reference), 0.94 (Hispanic vs. White), 0.97 (Asian vs. White) - all above 0.80; the model proceeds to filing.

Worked DI-ratio computation.

Step 1: BISG-imputed protected-class probabilities for each policy in the in-force book (or a representative sample if the book is large). Step 2: run production model scoring on the in-force book; capture predicted-severity per policy. Step 3: weight by imputed probability - for each protected group, sum (probability × predicted-severity) across all policies; divide by the sum of probabilities to produce the group's average predicted severity. Step 4: compute ratios. White is typically the reference group (highest predicted, lowest predicted, or majority-population depending on the carrier's methodology choice - the choice itself is documented).

Example numbers. Average predicted severity by BISG-imputed group: White $8,420; Black $7,680; Hispanic $7,950; Asian $8,180. With White as reference: Black/White = 7,680 / 8,420 = 0.91; Hispanic/White = 7,950 / 8,420 = 0.94; Asian/White = 8,180 / 8,420 = 0.97. All ratios above 0.80; model proceeds.

Step 5: sensitivity analysis. Run alternative BISG implementations (different surname-distribution vintage, different geocoding precision); run BIFSG where first-name is collected; run subset analyses by state (CO ratios, NY ratios, CA ratios separately); run alternative reference-group choices (Asian as reference, all-population average as reference). Document each sensitivity analysis result and the explanation of any divergences.

The four-fifths threshold's evolution. The 0.80 threshold comes from EEOC employment-discrimination precedent and was adopted into Colorado Reg 10-1-1 for insurance pricing. Some commentators have argued for tighter thresholds (0.85 or 0.90); others have argued the EEOC threshold is too strict for insurance where actuarially-justified rating factors may correlate with protected-class membership. The 2026 Colorado posture remains 0.80; tighter thresholds may emerge in 2027-2028.

Geographic-Proxy Detection

Independent of BISG, geographic features (zip codes, census tracts, territory codes) can act as proxies for protected-class membership. Geographic-proxy testing measures whether the carrier's geographic features correlate materially with majority-minority census-tract concentration.

Methodology. For each geographic feature in the model, compute correlation between the feature value and the census-tract minority-share variable from American Community Survey data (the rolling 5-year ACS estimates). Colorado Reg 10-1-1 threshold: 0.30 (correlations above this require justification or removal). NY DFS Circular Letter 2024-7 threshold: 0.25 (tighter, reflecting DFS staff's emphasis on geographic proxies in New York's residentially-segregated urban geography). California signaled under SB 1058 a threshold likely to follow Colorado's 0.30 standard.

Three-territory drop example from the BI rebuild. The personal-auto BI rebuild's initial feature set included 22 territory codes. Geographic-proxy testing on territory codes against census-tract minority share produced: 19 territories with correlation 0.04-0.27 (below the 0.30 threshold); 3 territories with correlation 0.34-0.41 (above threshold). The 3 territories were dropped at feature engineering; new territory codes created with broader geographic aggregation (combining adjacent territories to dilute the segregation signal); correlation re-tested on the revised features (all below 0.30). The Reg 10-1-1 filing exhibit documents the drop with reason codes ("territory 14 dropped at feature engineering due to correlation 0.41 with census-tract minority share; reaggregated into broader territory 14a") and the alternative territory feature engineering rationale.

The 0.25 vs. 0.30 jurisdictional difference. A model that passes Colorado's 0.30 threshold may fail NY DFS's 0.25 threshold. The carrier filing in both jurisdictions either runs separate model variants (operationally complex) or holds to the tighter 0.25 standard across the book (cleaner operationally). The 2026 best-practice carrier holds to the tightest standard across the book.

NY DFS Proxy Test Specifics

NY DFS Circular Letter No. 1 (2019) addresses the use of external consumer data in underwriting and pricing. Circular Letter 2024-7 (July 11, 2024) extends to AI/ML decision systems specifically and to the broader category of automated decision systems used by NY-domiciled and NY-writing insurers. Key requirements: (a) external-data sourcing documentation (where does the carrier obtain non-traditional data, including third-party scores, telematics signals, social-media signals if used); (b) proxy-test results using BISG or equivalent methodology; (c) explainability artifacts (SHAP for feature attribution, PDP and ALE for feature-effect visualization); (d) consumer-disclosure language consistent with the disclosure framework; (e) DFS examination authority including targeted exams.

The NY DFS approach to numeric thresholds. NY DFS does not specify a numeric DI threshold like Colorado's 0.80. DFS asks for testing and documented analysis; DFS staff evaluate the carrier's methodology and the explanatory quality of the bias-test exhibit. A carrier presenting a DI ratio of 0.91 with documented sensitivity analysis and clear explanation typically passes DFS scrutiny; a carrier presenting a single DI ratio without sensitivity analysis or geographic-proxy testing typically receives follow-up questions.

The tighter geographic-proxy threshold (0.25). Reflects DFS staff's emphasis on geographic proxies in New York's segregated urban geography. The bias-test exhibit should treat the 0.25 threshold as the binding constraint for NY filings.

The external-data documentation requirement. Every external data source used in the model must be documented: the source vendor (e.g., LexisNexis Risk Solutions, Verisk, Experian), the data type (credit-based insurance score, MVR, CLUE, telematics), the carrier's contracted access and usage rights, the carrier's validation of the data quality. The documentation accompanies the bias-test exhibit in the SERFF filing.

What a Defensible Bias-Test Exhibit Contains

The bias-test exhibit attached to a SERFF filing is the artifact regulators read first when evaluating the model. Length: 12-20 pages typically.

Section 1 - Scope and methodology. Protected classes tested (race, ethnicity, sex, and the broader SB 21-169 list); proxy methodology (BISG or BIFSG with rationale); reference group selection and rationale; sensitivity analyses planned and executed.

Section 2 - BISG methodology. Sources (Census Bureau Surname Distribution File version, geocoding methodology and precision); implementation details (Bayesian-update procedure, threshold for protected-class assignment if any); validation if available (self-reported sample, external-benchmark comparison).

Section 3 - In-force book DI ratios. Protected-group average predicted outcome, DI ratios with four-fifths threshold reference, confidence intervals where computed.

Section 4 - Sensitivity analyses. Alternative BISG implementations (different vintages, different geocoding precision); BIFSG comparison; subset analyses by state, by new-business-vs-renewal cohort, by tenure band; alternative reference-group choices.

Section 5 - Geographic-proxy test. Per-feature correlation table with census-tract minority share; threshold reference (0.30 Colorado, 0.25 NY DFS); decisions on failed features (dropped, re-aggregated, justified with documented business necessity).

Section 6 - Feature-level decomposition. SHAP contribution to the DI gap by feature; no single feature should contribute disproportionately to the DI gap; if a feature does, document the business-necessity rationale.

Section 7 - Explainability artifacts. SHAP values for material features; PDP plots for the top features; ALE plots for features with significant correlation (ALE corrects for feature-correlation bias that PDP exhibits).

Section 8 - Model-form constraints. Monotonicity constraints (e.g., predicted severity must increase monotonically with age above 50 on personal-auto BI); group constraints (e.g., no segment receives a rate change above +20% or below -25%); regulatory-mandated constraints (specific to jurisdiction).

Section 9 - Conclusion. The model passes bias testing under Colorado Reg 10-1-1, NY DFS Circular Letter 2024-7, Connecticut Bulletin MC-25-8, California SB 1058 (where applicable). Sign-off chain: the actuary (FCAS or equivalent credential) plus the chief actuary plus the chief data officer or AI governance committee chair.

The eight aggressive-fairness DOIs (CO, NY, CA, CT, WA, MA, NJ, IL) all expect this exhibit by 2026 in some form; the carrier with a uniform exhibit template across the book minimizes operational complexity.

SHAP, PDP, ALE for High-Risk Pricing Models

SHAP (Shapley values). Decomposes per-policy predictions into per-feature contributions. The SERFF examiner can ask: "For this representative policyholder with these specific feature values, what is each feature contributing to the predicted severity?" The actuary returns a quantitative decomposition - feature A contributes +$1,200, feature B contributes -$340, feature C contributes +$2,800, etc., summing to the predicted severity. SHAP is the most transparent feature-attribution method available. On a GLM the values compute exactly via the closed-form additive structure; on a GBM or ensemble the values approximate via Shapley sampling. The actuary documents which method (exact vs. sampled) and the sampling parameters.

PDP (Partial Dependence Plot). Shows the marginal effect of a feature averaged across the book. The plot visualizes how the predicted outcome moves as the feature value changes, holding other features at their distributional average. PDP is contaminated by feature-correlation bias when features are correlated - for example, driver-age, vehicle-age, and prior-accidents are correlated on a personal-auto model; the PDP for driver-age averages across all combinations of vehicle-age and prior-accidents, including unrealistic combinations.

ALE (Accumulated Local Effects). Corrects PDP by computing local differentials within slices of the feature; each slice uses observations actually in that slice. ALE is feature-correlation robust. Colorado examiner asked for ALE on the last filing cycle (informal reports from carrier actuarial teams in 2025); the actuary must include ALE in addition to PDP for material features. The bias-test exhibit's Section 7 includes ALE for the top 5-10 features.

The explainability triad. SHAP for per-policy decomposition; PDP for marginal effect intuition; ALE for feature-correlation-robust marginal effect. All three together give the examiner the explainability artifacts needed to evaluate the model's fairness and the actuary's methodology.

The 2026 Cross-Jurisdictional Bias-Test Pattern

The carrier filing in multiple aggressive-fairness states builds a uniform bias-test exhibit that satisfies the tightest jurisdiction's requirements and submits the same exhibit to all jurisdictions. The uniform pattern minimizes operational complexity, signals to regulators that the carrier takes fairness testing seriously, and allows the chief actuary to defend a single set of results rather than reconcile divergent jurisdiction-specific exhibits.

The 2026 carrier benchmark. A pricing actuary preparing a Colorado, NY, Connecticut, and California filing in 2026 produces one bias-test exhibit using BISG plus geographic-proxy testing at the 0.25 (tighter NY) threshold plus sensitivity analyses plus SHAP/PDP/ALE plus the standard nine-section structure. The exhibit attaches to the SERFF filing in each jurisdiction. Approval timelines vary (Colorado 30-90 days, NY 45-120 days, California Prop 103 75-180 days, Connecticut 45-90 days), but the underlying exhibit is the same.

Approval-conditional remediation. Where a jurisdiction's examiner finds a specific concern (e.g., the Colorado examiner asks for additional sensitivity on the BIFSG implementation; the NY DFS asks for additional documentation on the external-data sources), the actuary responds to that jurisdiction with supplemental analysis without modifying the core exhibit. The supplemental responses are filed with the SERFF submission and become part of the carrier's filing record.

Key Takeaways

• Colorado SB 21-169 plus Reg 10-1-1 is the binding quantitative fairness requirement. Phased expansion 2024-2026 from personal auto to homeowners (Oct 15, 2025 expansion) to life and annuity (Phase 3 2026) to commercial (proposed Phase 4). First compliance report due July 1, 2026; algorithm inventory mandated; consumer-facing AI disclosure references the inventory.
• BISG is the 2026 industry-standard proxy methodology. Accuracy varies by group: 92-96% White/Black; 85-89% Hispanic; 78-85% Asian; below 70% Native American/Pacific Islander. Cautions: probabilistic not deterministic; can encode geographic-segregation patterns producing circular logic; requires validation where possible against self-reported samples or external benchmarks.
• DI ratio formula: avg_predicted_outcome(protected) / avg_predicted_outcome(reference). Four-fifths threshold 0.80. Personal-auto BI worked example: 0.91 (Black vs. White), 0.94 (Hispanic vs. White), 0.97 (Asian vs. White) - all above threshold; model proceeds. Sensitivity analysis with alternative BISG implementations, BIFSG, subset analyses, alternative reference groups is non-negotiable.
• Geographic-proxy threshold 0.30 Colorado / 0.25 NY DFS. Three territory codes dropped at feature engineering on the BI rebuild for correlation 0.34-0.41 with census-tract minority share. Re-engineered territories all below threshold. Reg 10-1-1 filing exhibit documents the drop with reason codes.
• NY DFS Circular Letter No. 1 (2019) plus Circular Letter 2024-7 (July 11, 2024) impose external-data, proxy-test, explainability, and disclosure requirements. Tighter geographic-proxy threshold (0.25) reflects NY's segregated geography. No specific DI numeric threshold; DFS evaluates methodology and explanation quality.
• Bias-test exhibit 12-20 pages, nine sections. Methodology, BISG details, DI ratios, sensitivity analyses, geographic proxy, feature decomposition (SHAP contribution to DI gap), SHAP/PDP/ALE explainability, model-form constraints (monotonicity, group constraints), conclusion with sign-off chain. Eight aggressive-fairness DOIs (CO, NY, CA, CT, WA, MA, NJ, IL) all expect by 2026.
• SHAP per-policy decomposition; PDP marginal feature effect; ALE feature-correlation-robust corrected PDP. Colorado examiner asked for ALE on last filing cycle; ALE necessary on models with correlated features (driver age, vehicle age, prior accidents on auto).
• BIFSG (first-name added to BISG) increases accuracy 4-7% on Hispanic and Asian groups. Where carrier collects first-name, BIFSG worth implementing; methodology documented in exhibit with validation.
• 2026 cross-jurisdictional pattern: uniform bias-test exhibit satisfying tightest jurisdiction's requirements (NY DFS 0.25 geographic threshold) filed across all aggressive-fairness states. Approval timelines vary (CO 30-90 days, NY 45-120 days, CA 75-180 days, CT 45-90 days); core exhibit unchanged; jurisdiction-specific supplemental responses filed where examiners request.