Chapter 32 — Further Reading

Sources are grouped by the book's three citation tiers: Tier 1 (verified canonical — institutions, frameworks, and public regulatory actions we can stand behind), Tier 2 (real practices and bodies of knowledge whose exact citation you should confirm before quoting), and Tier 3 (illustrative and constructed — the teaching material in this chapter). Numbers in the chapter are illustrative unless tied to a Tier-1 source.

Tier 1 — Verified canonical

  • The National Association of Insurance Commissioners (NAIC) — the standard-setting body for U.S. insurance regulation. Its work on artificial intelligence and big data in insurance (including principles for the use of AI and ongoing model-governance and accelerated-underwriting reviews) is the central regulatory reference for this chapter. Start at the NAIC's published materials on AI/Big Data.
  • Colorado SB21-169 — the Colorado statute restricting insurers' use of external consumer data and algorithms/predictive models that could result in unfair discrimination, with implementing regulation by the Colorado Division of Insurance. A landmark, real state action directly on the §32.1 / §32.5 / Ch.35 proxy-discrimination problem.
  • The Fair Credit Reporting Act (FCRA) — governs the use of consumer-report data (including credit-based insurance scores) that frequently enters pricing models as features; defines adverse-action obligations. (Owned in Chapter 8; load-bearing here for feature legality.)
  • The McCarran-Ferguson Act and state rate-regulation law — the framework under which a model's output is still a filed rate subject to approval (Chapter 4). Predictive-model rate filings live inside this regime.
  • The Genetic Information Nondiscrimination Act (GINA) — the federal limit on the use of genetic information; relevant as the clearest statutory line on a category of data a model may not use (the fuller life-and-genetics treatment is in Chapters 17 and 35).
  • The Casualty Actuarial Society (CAS) and the Society of Actuaries (SOA) — the professional actuarial bodies whose published syllabi, monographs, and research underpin GLM and predictive-modeling practice in insurance. Their materials are the authoritative technical reference for the methods in §32.2–§32.6.

Tier 2 — Attributed, specifics to confirm

  • The GLM as the personal-lines pricing standard — it is well established in the industry that multivariate GLMs became the dominant ratemaking method in personal auto and homeowners, spreading from the U.K. motor market in the 1990s–2000s outward. Attribute the pattern; confirm any specific carrier, date, or figure before quoting it (see Case Study 1).
  • Standard actuarial texts on predictive modeling and GLMs for insurance ratemaking — there is a recognized body of practitioner literature (CAS monographs and widely-used textbooks on GLMs for ratemaking and on predictive modeling applications in actuarial science). Consult the current editions for the Poisson-frequency / gamma-severity construction, model validation, and lift/Gini methodology rather than relying on memory for specifics.
  • Gradient boosting and machine learning in insurance — the use of GBMs and other ML for risk selection, triage, and feature discovery (alongside GLMs for filed pricing) is widely reported industry practice; the accuracy-vs-interpretability trade and the use of interpretability methods (e.g., SHAP, partial dependence) are well attested. Confirm specifics in current technical literature.
  • Image- and satellite-based property underwriting — the use of aerial and satellite imagery analyzed by machine-learning models to assess roof condition, wildfire fuel, and property characteristics is an established and growing practice (and connects to Chapter 31's data-enrichment treatment). Attribute the practice; avoid quoting vendor-specific accuracy claims as fact.
  • "Price optimization" restrictions — numerous U.S. states have issued bulletins or taken action restricting or prohibiting price optimization (pricing to a customer's tolerance rather than risk). Attribute the regulatory trend; confirm the specific states and dates before citing.

Tier 3 — Illustrative / constructed (this chapter)

  • Harbor Steel & Fabrication — the constructed progressive-project account; the model's 7/10 score and the underwriter's documented override to a 6 are illustrative teaching constructions (frozen in _continuity.md §5).
  • All worked figures — the GLM relativity example (1.85, 1.30, 1.20, 0.92), the lift chart deciles (45%…190%), the Gini values, and every loss ratio and combined-ratio number in the chapter are constructed teaching examples, not data from any real carrier.
  • Case Study 2's carrier and scenario — a clearly-labeled composite of a real, documented industry and regulatory concern (proxy discrimination in predictive models); the carrier, the disparity, and the remediation are constructed to teach the pattern, with no fabricated statistic.

If you read only one thing: read the NAIC's published materials on artificial intelligence and big data in insurance alongside Colorado's SB21-169 and its regulation. Together they are the clearest real-world statement of the rules now forming around the exact tension this chapter teaches — that a model's accuracy is never a substitute for its fairness and defensibility, and that "the algorithm chose it" is not an answer a regulator will accept.