Akur8 raised at a $700M valuation in 2024 because the pricing-actuarial workflow at most P&C carriers in 2018-2022 was a six-month black-box GBM tournament producing models with strong holdout AUC and zero defensibility under NAIC §4 governance review. The transparent-GLM-with-fine-tuned-GBM-comparator workflow Akur8 standardized - and that Earnix, Quantemplate, and the SAS Viya pricing modules have analogous versions of - is the 2026 production-line for any carrier wanting to file a personal-auto, homeowners, commercial-property, or small-business rate change through SERFF without a coverage-counsel cardiac event. The workflow takes a sample frequency dataset (call it 240,000 vehicle-year exposures across 14 territory groups with ten candidate predictors), produces a transparent Poisson GLM with interpretable coefficients, generates a model card with SHAP/PDP plots, fine-tunes a GBM as performance comparator, documents variable selection against protected-class proxy tests, pulls comparable-carrier filings via Akur8 Discover or Matrisk's filing-intelligence database, and produces the algorithm-inventory entry that Colorado Reg 10-1-1 and NY DFS Circular Letter 2024-7 both demand. AI compresses the documentation overhead - variable-selection rationale, SHAP-plot interpretation, model-card draft, SERFF-package transmittal - that historically gated the pricing-actuarial team's output. The actuary still owns the model. The chief actuary still signs the certification. This lesson is the structured workflow for building a transparent, defensible, file-ready pricing prototype.

The Dataset the Actuary Walks Into

The frequency dataset for the worked example is a 240,000-row personal-auto vehicle-year exposure dataset covering policy years 2021-2024 across 14 territory groups, with the target variable being claim count (Poisson-distributed) per exposure-year. The candidate predictors are: (1) driver age (continuous, banded at the model layer); (2) vehicle age (continuous); (3) annual mileage (continuous, banded); (4) territory group (categorical, 14 levels); (5) prior-claim count in last 5 years (continuous, capped at 3); (6) credit-based insurance score (continuous, banded - known protected-class proxy candidate requiring proxy test); (7) marital status (categorical); (8) vehicle symbol (categorical, manufacturer-driven); (9) annual mileage band (categorical, derived from continuous); (10) policy tenure with carrier (continuous, years). The actuary's job is to build a transparent GLM, document why each variable is included and how it interacts with protected-class proxies, generate the model card, fine-tune a GBM comparator for performance benchmarking, and produce the SERFF-ready filing package.

Akur8's transparent-GLM workflow (or Earnix's, or SAS Viya's pricing modules) provides the structural framework: data preparation with monotonic constraints, link function selection (Poisson with log link for frequency), penalized regression (LASSO/ridge regularization), automated variable selection with override capability, coefficient extraction with confidence intervals, and SHAP-based feature importance for model-card generation. The actuary's professional judgment intervenes at each step - the workflow surfaces the analytical choices; the actuary makes the calls that survive ASOP 41 communications scrutiny and SERFF examination.

The Transparent GLM Build

The Poisson GLM for frequency is the structural starting point. Log link, offset for exposure-years, predictors selected via penalized regression with cross-validation. The model coefficients are interpretable: each predictor's β coefficient maps to a multiplicative effect on expected frequency. A coefficient of 0.34 on the "vehicle symbol high" indicator means a 40% increase in expected frequency vs. the reference level. The interpretability is what makes the model defensible - every rate factor in the eventual filing maps to a documented coefficient that the actuary can explain in plain language during the SERFF examination conversation with the state DOI rate reviewer.

Variable selection uses LASSO regularization to suppress weakly-predictive variables. The remaining variables get the actuary's professional judgment override - keep a variable that LASSO suppressed if business rationale supports it (driver age might be retained even if the LASSO penalty pushed it toward zero, because age is a regulatory-recognized rating factor); drop a variable that LASSO kept if the variable fails protected-class proxy testing (credit-based insurance score may be dropped or constrained if proxy testing surfaces material disparate impact). Akur8's UI surfaces the LASSO results, the actuary's manual override decisions, the proxy-test results, and the final model specification. The workflow logs each decision with rationale; the documentation is the audit trail SERFF reviewers and DOI examiners need.

The Model Card With SHAP and PDP Plots

The model card is the artifact that documents the model for regulatory and governance review. Following the NAIC Model Bulletin §4 framework and the model-card guidance from the AI Risk Management Framework (NIST AI RMF), the card has eight sections: (1) model purpose and use case (pure-premium frequency component for personal auto rate filing in 14 territory groups); (2) data sources and time period (carrier's policy and claims data 2021-2024); (3) model architecture (Poisson GLM with log link, LASSO regularization); (4) variable list with definitions and data lineage; (5) performance metrics (holdout deviance, Gini coefficient, lift charts by decile); (6) SHAP feature-importance plot and partial-dependence (PDP) plots for each predictor; (7) protected-class proxy test results and disparate-impact analysis; (8) limitations, operating constraints, and human-review checkpoints.

The SHAP plot shows the magnitude and direction of each predictor's contribution to predicted frequency across the dataset. Driver age and prior-claim count typically dominate; vehicle symbol and territory group follow; credit-based insurance score (if retained) sits in the middle with the proxy-test caveat. The PDP plots show the predicted frequency as a function of each predictor holding others at their mean - a U-shape on driver age (high at young and old, low at middle), an increasing relationship on prior claims, a step function on vehicle symbol bands. AI generates the plots from the model output; the actuary interprets and writes the model-card narrative. The card is the foundational artifact every downstream filing reference points to.

The Fine-Tuned GBM Comparator

The GBM (Gradient Boosting Machine, typically XGBoost or LightGBM) comparator is the benchmark - it captures interactions and non-linearities the GLM may miss, and its performance metrics tell the actuary how much predictive power the transparent GLM gives up for interpretability. The GBM is fine-tuned with the same predictors, the same train/test split, and the same evaluation framework as the GLM. Typical results on a clean frequency dataset: GBM gains 2-7% on holdout deviance and 1-4 points on Gini coefficient over the GLM. The actuary documents the gap and decides whether the gap is material enough to warrant the GBM in production (rarely; the regulatory and interpretability costs typically exceed the predictive gain) or whether the GLM is the production model with the GBM serving as performance benchmark and feature-engineering input (typical 2026 outcome).

The GBM comparator also surfaces interaction patterns the GLM does not capture. SHAP interaction plots from the GBM show that, for example, prior-claim count interacts with driver age (younger drivers with prior claims show steeper frequency lift than older drivers with prior claims) or that vehicle symbol interacts with territory (urban high-symbol vehicles show non-additive frequency vs. rural high-symbol vehicles). The actuary uses these signals to add specific interaction terms to the GLM (e.g., a driver_age × prior_claims interaction term) or to inform business judgment about rating-plan structure (territory-specific symbol curves). The GBM doesn't replace the GLM; it improves the GLM through structured feature engineering surfaced by the comparator.

Protected-Class Proxy Testing and the Credit-Score Decision

The most consequential analytical decision in the workflow is what to do with credit-based insurance score. Credit-based insurance score is a regulatorily-recognized rating factor in many states but a known proxy for race and ethnicity in many other states' interpretations. Colorado Reg 10-1-1 explicitly requires disparate-impact testing on every variable in a rating model; New York DFS Circular Letter 2024-7's proxy test applies similar discipline. The proxy test compares the predicted frequency for protected classes (using inferred demographic data via BISG - Bayesian Improved Surname Geocoding - or other approved methods) against the predicted frequency for non-protected classes; statistically significant disparate impact triggers either variable removal, variable constraint (e.g., capping the coefficient), or a documented justification of business necessity.

For the worked example, the proxy test on credit-based insurance score may show a 12-18% disparate impact ratio between predicted-frequency-for-Black-policyholders and predicted-frequency-for-White-policyholders (using BISG inference). The actuary's options: (a) drop credit-based insurance score from the model (loss of predictive power, increased rate uncertainty); (b) constrain the coefficient (cap the multiplicative effect at a threshold, e.g., max 1.4x base frequency for the highest credit band, vs. unconstrained 2.1x); (c) document business necessity per NY DFS guidance with parallel testing showing the variable's predictive power is not duplicable through proxy-free variables. Each path has a different SERFF filing posture and a different examination conversation. The actuary's decision is documented in the model card, the proxy-test exhibit, and the algorithm-inventory entry.

Comparable-Carrier Filing Intelligence via Akur8 Discover and Matrisk

Before submitting the SERFF filing, the actuary wants to know how other carriers handle the same variables in the same states. Akur8 Discover and Matrisk's filing-intelligence database aggregate SERFF-public filings by carrier, line, state, and rating factor; the actuary queries "personal auto credit-based insurance score in California 2024-2026" and gets the carrier-by-carrier coefficient ranges, the constraint approaches, the documented justifications, and the DOI examiner questions per filing. The intelligence informs the actuary's filing posture: if seven peer carriers cap the credit-score coefficient at 1.4x and the carrier under review proposes 2.1x, the DOI examiner will ask why; the actuary either matches the peer cap or documents the business necessity for the wider band.

The filing-intelligence pass also surfaces emerging trends: which states recently published bulletins constraining specific variables, which DOIs are actively examining bias-testing exhibits, which approval patterns indicate softer or harder review. The actuary's filing strategy incorporates the intelligence; the SERFF transmittal letter cites peer-carrier patterns where favorable; the rate-filing memorandum addresses anticipated DOI examiner questions. AI surfaces the intelligence; the actuary integrates it into the filing strategy.

The SERFF Filing Package the Actuary Assembles

The SERFF (System for Electronic Rate and Form Filing, NAIC-operated) submission package contains four core documents: (1) cover transmittal letter explaining the filing's purpose, the rate change requested, the effective date, and the methodology highlights; (2) actuarial memorandum documenting the model, the variable selection, the bias testing, and the rate-change calculation; (3) rate-change exhibit showing the proposed rates by territory, class, and rating tier vs. current rates; (4) supporting actuarial appendices (model card, SHAP/PDP plots, proxy-test results, comparable-carrier intelligence summary). Each document references the others; together they comprise the filing.

The actuarial memorandum is the centerpiece. It walks the DOI rate reviewer through the data, the methodology, the variable selection rationale (why each variable is in the model and how it was tested for proxy effects), the model performance, the bias-testing results, and the rate-change calculation. The memorandum is signed by the appointed actuary under the Code of Professional Conduct of the American Academy of Actuaries and the applicable ASOPs (23 data quality, 38 model evaluation, 41 communications, 56 modeling). AI drafts the structural sections and integrates the model output; the actuary writes the methodology and variable-selection sections in the actuary's voice and reviews the entire memorandum for accuracy and defensibility before signing.

The Colorado Reg 10-1-1 Algorithm Inventory Entry

Colorado Reg 10-1-1 (Governance and Risk Management Framework Requirements for Insurers' Use of External Consumer Data and Information Sources, Algorithms, and Predictive Models) requires every insurer writing in Colorado to maintain an inventory of every algorithm or predictive model used in underwriting, rating, claims, and marketing. The inventory entry has specific required fields: model name and version, intended use, data sources, training/validation methodology, performance metrics, bias-testing results, human-review controls, model-update cadence, retirement protocol, and accountable officer. The Colorado DOI publishes inventory examination expectations; insurers must respond on demand within defined timeframes.

For the worked example's frequency GLM, the inventory entry reads: "Personal Auto Frequency GLM v3.1, intended use: pure-premium frequency component for personal auto rate filings in Colorado and other states where the model is approved. Data sources: carrier's policy and claims data 2021-2024, BISG-inferred demographic data for proxy testing. Methodology: Poisson GLM with log link, LASSO regularization for variable selection, manual override applied per actuary's professional judgment. Performance: holdout deviance [X], Gini [Y]. Bias testing: proxy test on credit-based insurance score showed 14% disparate impact ratio (Black vs. White inferred via BISG); coefficient constrained at 1.4x cap per state-specific requirements. Human-review controls: appointed actuary signs every model release; quarterly performance monitoring against holdout; annual recalibration cycle. Model-update cadence: annual unless material drift detected. Retirement protocol: documented in the model lifecycle policy. Accountable officer: [Chief Actuary Name]." Each AI-touched model and predictive component gets its own inventory entry; multistate carriers maintain a unified inventory; algorithmic-inventory examinations look for completeness and accuracy across the registry.

What the Actuary and the Chief Actuary Each Own

The pricing actuary owns the model build, the variable-selection decisions, the bias-test results documentation, the model card, the SERFF filing memorandum, and the algorithm-inventory entry. The chief actuary owns the actuarial certification under ASOP 41 - signing the certification that attests to the model's compliance with applicable ASOPs, the appropriateness of the methodology, and the reasonableness of the rate change. The two roles do not overlap; both are non-delegable under the actuarial qualification standards. AI accelerates the documentation and analytical work - variable-selection summary, model-card draft, bias-test report formatting, comparable-carrier intelligence integration - that historically consumed actuarial time without producing professional-judgment value. The actuary's time concentrates on the analytical and professional-judgment work; the chief actuary's time concentrates on the certification review.

The rate-filing manager owns the SERFF submission mechanics - uploading the package, responding to DOI examiner questions, tracking the approval timeline. The chief underwriting officer owns the implementation when the rate is approved - building the rate factors into PolicyCenter or Duck Creek, updating the underwriting guidelines, training the underwriters on the new rates. AI accelerates each role's work but does not substitute for any. The 2026 pricing-actuarial team using Akur8 (or Earnix, or SAS Viya pricing) at a mid-market regional carrier produces 25-40% more filings per actuary-year vs. the pre-Akur8 baseline; the increase is in documentation throughput, not in cuts to professional judgment time.

Key Takeaways

• The transparent-GLM-with-fine-tuned-GBM-comparator workflow Akur8 standardized (and Earnix, SAS Viya pricing modules support) is the 2026 production-line for any P&C rate filing through SERFF. 240,000 vehicle-year sample dataset, 10 candidate predictors, Poisson GLM with log link, LASSO regularization, GBM comparator for benchmarking, model card for governance.
• The Poisson GLM build uses log link, exposure-year offset, LASSO regularization for variable selection, and actuary's professional-judgment override on retention or exclusion decisions. Coefficient interpretability is the defensibility foundation - every rate factor in the filing maps to a documented coefficient.
• The model card has eight sections per NAIC Model Bulletin §4 and NIST AI RMF guidance: purpose, data sources, architecture, variable list with lineage, performance metrics, SHAP/PDP plots, protected-class proxy test results, and limitations/operating constraints/human-review checkpoints. AI generates the structure; the actuary writes the narrative.
• The GBM comparator typically gains 2-7% on holdout deviance and 1-4 points on Gini vs. the GLM; the actuary documents the gap and decides whether to keep GLM in production (typical) or migrate to GBM (rare given interpretability costs). GBM SHAP interaction plots surface interaction terms to add to the GLM, improving the transparent model through structured feature engineering.
• Protected-class proxy testing on credit-based insurance score is the most consequential analytical decision: drop, constrain, or document business necessity per NY DFS Circular Letter 2024-7 and Colorado Reg 10-1-1 disparate-impact testing. BISG (Bayesian Improved Surname Geocoding) is the typical inference method; 12-18% disparate impact triggers constraint or removal.
• Akur8 Discover and Matrisk's filing-intelligence database aggregate SERFF-public filings by carrier, line, state, and rating factor; the actuary queries comparable carriers to inform filing strategy and anticipate DOI examiner questions. Where peer carriers cap a coefficient at 1.4x and the carrier proposes 2.1x, the DOI examiner will ask why.
• The SERFF filing package has four core documents: cover transmittal, actuarial memorandum, rate-change exhibit, and supporting actuarial appendices (model card, SHAP/PDP, proxy-test results, comparable-carrier intelligence). The actuarial memorandum is the centerpiece signed by the appointed actuary under ASOPs 23, 38, 41, 56.
• Colorado Reg 10-1-1 requires algorithm-inventory entries with specific fields: model name and version, intended use, data sources, methodology, performance metrics, bias testing, human-review controls, update cadence, retirement protocol, and accountable officer. Multistate carriers maintain unified inventory; the entry serves examination response and discovery.
• 2026 mid-market regional carriers using Akur8 produce 25-40% more filings per actuary-year vs. the pre-Akur8 baseline. The increase concentrates in documentation throughput; the pricing actuary's and chief actuary's professional-judgment time is preserved and amplified, not reduced.