Emerging Regulatory Issues

Study guide: Exam 6U

Emerging Regulatory Issues are the questions U.S. insurance regulators are working through with laws written for an earlier market — chiefly price optimization, predictive models, artificial intelligence and third-party "big data", climate risk and cyber risk. The pattern is consistent: the long-standing rating and solvency tests are applied to the new tool first, and new guidance follows where those tests prove hard to administer.

$\text{A filed rate must be} \begin{cases} \text{not excessive} \\ \text{not inadequate} \\ \text{not unfairly discriminatory} \end{cases}$

Price optimization — adjusting cost-based rates for customers' demand. The NAIC Casualty Actuarial and Statistical Task Force's 2015 white paper held that rates should be cost-based and that adjustments for price elasticity, propensity to shop, individual retention or a propensity to complain are unfairly discriminatory. Maryland (2014) was the first of many states to issue a prohibiting bulletin.
Predictive models in rate filings. GLMs are now routine in personal lines. The Task Force's Regulatory Review of Predictive Models white paper (2020) sets out what a reviewer should obtain: the data and its adjustments, why each variable was selected and how it relates to loss, how the model was validated, and how its output became the filed rates. A variable with no intuitive link to loss needs stronger support. The actuary's side is governed by the ASOPs on risk classification (No. 12), data quality (No. 23) and modeling (No. 56).
AI, big data and unfair discrimination. The NAIC adopted AI principles in 2020 — fair and ethical, accountable, compliant, transparent, secure — and in December 2023 a model bulletin on insurers' use of AI systems, since adopted by many states: a written AI program with board oversight, testing for errors, bias and unfair discrimination, and responsibility for third-party models and data. Colorado's 2021 statute goes further, requiring insurers to show that external consumer data and algorithms do not unfairly discriminate by protected class, with rules phased in by line. The open question is proxy discrimination: whether an outcome correlated with a protected class is unlawful when the class is not an input (Unfair Discrimination).
Climate. Rising weather losses press on solvency and availability together. The NAIC's Climate Risk Disclosure Survey was aligned with the TCFD framework in 2022; catastrophe risk entered the RBC formula (Risk-Based Capital); and ratemaking is being revisited — Florida's public commission has long reviewed hurricane models for residential rate filings, and California in 2024–25 moved to allow catastrophe models and net reinsurance cost in rates, tied to commitments to write in distressed areas.
Cyber. Insurers as holders of data are subject to the NAIC Insurance Data Security Model Law (2017), adopted in many states: an information security program and prompt notice of cybersecurity events. As writers of cyber coverage they face scrutiny of accumulation risk and of "silent" cyber exposure in traditional policies, with market data collected through an annual statement supplement.

A New Third-Party Variable in a Homeowners GLM An insurer files a homeowners GLM that adds a vendor's "household purchasing behaviour" score. It improves the model's lift noticeably. What will a regulator ask, and on what basis?

Answer The legal test is still the statutory one: is the resulting rate unfairly discriminatory? The review will ask:

What is the data? Source, vendor documentation, how it was obtained, and whether it is a consumer report — which would bring in the Fair Credit Reporting Act and its notice duties.

Why does it predict loss? The 2020 white paper expects a rationale; a variable with no intuitive relationship to loss needs strong statistical support, and a regulator may reject lift alone.

Is it a proxy? Its correlation with protected characteristics and with income. In a state with Colorado-style rules, quantitative testing is required; under the NAIC AI bulletin, the insurer must be able to show it tested for unfair discrimination.

Is it a demand variable in disguise? Purchasing behaviour can measure willingness to pay rather than risk. If so, it is price optimization under another name.

Who is responsible? The insurer, not the vendor — the AI bulletin makes third-party models the insurer's governance problem.

A plausible outcome is a request for more support and a proxy analysis before approval, or approval without the variable.

Historical Average Versus Catastrophe Model A homeowners filing has a non-catastrophe loss cost of $\$900$ per policy. Wildfire loss cost is $\$80$ on a 20-year historical average but $\$200$ from a catastrophe model. The permissible loss ratio is $0.65$ . Compare the indicated rates.
Answer
 $\begin{align*} \text{Historical basis} &= \frac{900 + 80}{0.65} \\ &= \$1{,}507.69 \\ \text{Modelled basis} &= \frac{900 + 200}{0.65} \\ &= \$1{,}692.31 \\ \text{Difference} &= \frac{1{,}692.31}{1{,}507.69} - 1 \\ &= 12.2\% \end{align*}$ 
If the hazard is changing, a historical average is a biased estimate of future cost, and a regulator that allows only the historical basis approves a rate about $12\%$ short. Insurers tend to answer persistent inadequacy by writing less rather than by absorbing losses — which is why states facing an availability crisis have traded model-based rates for commitments to keep writing.