Predictive Analytics: Regulatory Review
By Rachel Hemphill
Insurance regulators must review data, methods, and models from diverse companies and diverse product lines, in an industry—and a social, economic, and technological environment—that is constantly changing and presenting new challenges. To be effective, regulators must be familiar with the latest modeling methods used by companies for insurance rating and underwriting. The continuing evolution of methods employed by insurance companies has increased the importance of regulators understanding several potential pitfalls. Three such problem areas are proxy variables, spurious correlations, and price optimization. In each case, a major reason for heightened regulatory concern has been the incorporation of Big Data into insurers’ predictive analytics.
Proxy Variables
Proxy variables occur when a variable is highly correlated with another variable that cannot be used directly. Proxy variables are problematic when they are proxies for a variable that is itself prohibited. That is, the proxy variable is in essence a “stand-in” for a prohibited variable, and so raises the same concerns as the prohibited variable. Big Data increases regulatory concern about proxy variables due to the vast increase in the number of variables available for incorporation in the model. For example, data on grocery store purchases could conceivably be used to construct a proxy variable for ethnicity. This increase in variables considered has also led to an increased use of methods, such as dimensionality reduction techniques, that result in variables that may not have an intuitive interpretation or a clear relationship to loss. It becomes more difficult for the regulator, and even the company actuary, to understand what the predictor variable represents and whether it is a proxy for a prohibited variable.
Spurious Correlations
The widely used rule for measuring the significance of a variable is that the p-value must be 0.05 or smaller. In the Casualty Actuarial Society monograph, Generalized Linear Models for Insurance Rating (page 9), the authors explain, regarding the common 0.05 threshold, that “Since in a typical insurance modeling project we are testing many variables, this threshold may be too high to protect against the possibility of spurious effects making it into the model.” Spurious correlations exist when the historical correlation between two variables is random or coincidental. In these cases, one variable cannot reliably be used to inform a projection of the other variable going forward. For example, over the past year the number of California Department of Insurance rate regulation actuaries has increased, as has California average rainfall. Unfortunately, however, we cannot expect to influence future California rainfall by hiring additional actuaries. While the authors of Generalized Linear Models believe that a requirement stricter than the 0.05 threshold may be necessary to avoid spurious correlations, regulators have actually seen filings moving in the opposite direction, attributing significance to p-values of 0.10, 0.15, or even 0.20.
The question of spurious correlations has become more relevant as technical specialization has increased, creating distance between subject expertise and methodological expertise. For advanced models and predictive analytics, regulators have seen insurers increasingly rely on models that are outside the specific area of expertise of the company actuary, supplied by third-party vendors or actuarial consultants. This separates the individual with the deep understanding of the book of business—the company actuary—from the individual with the deep understanding of the modeling methods. When this is the case, company actuaries should still have at least a basic understanding of the model, evaluate whether it is appropriate for the intended application, and follow the other related guidance described in ASOP No. 38—Using Models Outside the Actuary’s Area of Expertise (Property and Casualty). In addition, it is prudent for the company to ensure that the vendor or consultant will be available to answer later regulatory questions regarding the model, or the model review may reach an impasse.
Price Optimization
While definitions vary, price optimization generally means setting prices based not solely on the cost of the individual risk transfer but also, in part, on the price that a customer or group of customers will tolerate. Why are regulators concerned about price optimization? One simple reason is that free-market forces are not a reliable defense against unfairly discriminatory insurance pricing, due to asymmetric information, for example, or because legal requirements can result in inelastic demand. Twenty jurisdictions have issued bulletins on price optimization; actuaries should be very familiar with the bulletins for their jurisdictions. If a jurisdiction has not yet issued a bulletin regarding price optimization, this does not mean that price optimization is allowed in that jurisdiction. Indeed, many jurisdictions that have issued bulletins considered them not to be introducing a new requirement, but simply reiterating an existing requirement, that rates not be unfairly discriminatory.
How will regulators know if an insurer is using price optimization? We expect the filing actuary to follow ASOP No. 41—Actuarial Communications, Section 3.2, which states:
In the actuarial report, the actuary should state the actuarial findings, and identify the methods, procedures, assumptions, and data used by the actuary with sufficient clarity that another actuary qualified in the same practice area could make an objective appraisal of the reasonableness of the actuary’s work as presented in the actuarial report. [Emphasis added.]
In following ASOP No. 41, Section 3.2, a company actuary would presumably need to provide, as part of the filing, sufficient documentation to review the reasonableness of both the indications and where and why selections differ from indications. The logic of this section would seem to require that all methodologies or modeling processes used in determining the final rates should be clearly documented in the actuarial communication for any filing. This is especially important if the selections are based in part on some form of price optimization, given that the use of price optimization has often not been explicitly documented.
Unfortunately, some have proposed that the solution to the inherent inequity is to shift the burden to the consumer, who is then expected to address imbalances in information, for example by doing exhaustive research on the prices and coverages provided by the various companies. Of course, expecting consumers to understand and address all of the drivers of their own insurance costs is a high demand. This doesn’t just require that the consumers know that they may be charged a counterintuitive penalty for their loyalty to their insurer and should therefore shop around. It also requires that they understand, for example, the effect that their credit score may have on their insurance costs. Further, for consumers to understand their insurance costs, they need to understand complex variables, such as the kurtosis of the average rainfall in a neighborhood where they are considering purchasing a home. The reasons for such variables may be known to an actuary, but will not likely be known to the consumer.
To say that an insured should avoid a loyalty surcharge by shopping around presumes that the insured has an abundance of time and information to research and procure alternative insurance, and that the time cost of that research to the insured should be ignored. In fact, it may be a rational choice for an insured not to invest tens of hours of research to find the best price; yet, the higher price charged to a loyal customer still does not become cost-based with respect to the insurer’s expected losses and expenses simply because it is too costly of an investment for the consumer to avoid the penalty.
Regulatory Resources
The regulatory review process for predictive analytics would be significantly improved by giving regulators readier access to targeted, well-organized, and comprehensive continuing education, which can equip the regulators with general knowledge of predictive analytics topics. Beyond this, though, regulators also need to discuss these issues from a regulatory perspective with other regulators. A valuable resource for regulatory actuaries is multi-state coordinated training and education efforts through the NAIC and NAIC’s Casualty Actuarial and Statistical Task Force (CASTF). The NAIC is currently assisting regulators on these topics through the following:
Rachel Hemphill, MAAA, FCAS, FSA, Ph.D., is chief systems actuary at the California Department of Insurance.
This article is solely the opinion of its author. It does not express the official policy of the American Academy of Actuaries; nor does it necessarily reflect the opinions of the Academy’s individual officers, members, or staff.
CPC Capitol Hill Briefing on Flood Insurance
 As Congress considers legislation to revise and reauthorize the National Flood Insurance Program (NFIP), the Casualty Practice Council (CPC) presented a June 26 Capitol Hill briefing on flood insurance that was well attended by congressional and federal agency staff.
Presenters included Stu Mathewson and Nancy Watkins, members of the CPC’s Flood Insurance Work Group, with Senior Casualty Fellow Jim MacGinnitie moderating. Reviewing the work group’s flood insurance monograph released in April, The National Flood Insurance Program: Challenges and Solutions, the presenters underscored the Academy’s support for expanding private flood insurance coverage. The briefing was attended by more than 40 people, including congressional staff and representatives of the Government Accountability Office and Congressional Budget Office.
Presenters also expressed concern that none of the pending legislative proposals include a serious plan to repay or forgive the NFIP’s past and future debts attributable to infrequent megastorms, and they reminded attendees that rising sea levels are expected to cause an increase in the frequency of coastal flooding and in the severity of megastorms.
Essential Elements Paper on Technology and Auto Insurance
 The Academy released a new Essential Elements, “Auto Insurance in the 21st Century,” which explores insurance issues surrounding autonomous vehicles, ride-sharing services, and distracted driving. The Essential Elements series is designed to make actuarial analyses of public policy issues clearer to general audiences.
CPC Presents on Predictive Modeling at NAIC Insurance Summit
 In response to a request by state regulators and in an effort to share an actuarial perspective, the CPC presented a full-day program on predictive modeling at the NAIC’s Insurance Summit on May 25 in Kansas City, Mo.
The five presentations, coordinated by Roosevelt Mosley, chairperson of the Automobile Insurance Committee, covered seven basic questions:
-
What is predictive modeling?
-
Where does predictive modeling occur?
-
Who does predictive modeling?
-
Why is predictive modeling done?
-
When does predictive modeling happen?
-
How is predictive modeling done?
-
How much predictive modeling is enough?
The presentations also covered the process of developing predictive models, data sources, considerations for insurance companies in developing rate filings, and regulatory concerns about Big Data and predictive modeling. The program concluded with a public policy discussion led by Senior Casualty Fellow Jim MacGinnitie.
Participating in the panel discussion was Rachel Hemphill, chief systems actuary in the California Department of Insurance Office of Principle-Based Reserving (see lead story, above).
“The balanced and informative manner with which the presentations were provided will undoubtedly help our various jurisdictions develop best practices for regulatory review of rate filings that make use of complex predictive models,” said Michael McKenney, actuarial supervisor in the Pennsylvania Insurance Department’s property and casualty bureau, who chairs NAIC’s Casualty Actuarial and Statistical Task Force (CASTF).
Actuaries Climate Index Updated
 Actuarial organizations, including the Academy, that represent the profession in the United States and Canada reported in late June that the Actuaries Climate Index (ACI) composite value for fall 2016 was 2.07—the highest seasonal level recorded for the two countries combined.
“We have now seen four of the last five seasons having an Actuaries Climate Index value over 1.50, compared to the 30-year reference period, which had no index values above 1.00,” said Doug Collins, chairperson of the ACI Climate Change Committee.
P/C Issues on Agenda for 2017 Annual Meeting and Public Policy Forum
breakout-session topics of interest to P/C actuaries will be included:
-
Flood Insurance: The Changing Picture—The National Flood Insurance Program (NFIP) will expire on Sept. 30 unless Congress reauthorizes it. What changes can we expect? What does the future look like for insurers, regulators, actuaries, and consumers? What is the new role for private insurers in the flood insurance market? How might rising sea levels affect the NFIP? (Estimated continuing education (CE) credit: 1.8)
-
Auto Insurance in the 21st Century—Speakers will provide an intriguing look at new and emerging technologies and how they will impact—and already are affecting—automobile insurance. Included will be a discussion of autonomous vehicles, shared vehicles, and distracted driving and how these new developments will impact consumers, regulators, and insurers. (Estimated CE: 1.8)
-
2018 Preview: A Survey of the State & Federal Landscapes—Speakers will address expected hot topics for next year in Congress, the state legislatures, and state regulatory agencies. Topics to be discussed may include: climate change, cyber risk, risk-based capital, travel insurance, Big Data, international reserving standards, and more. (Estimated CE: 1.8)
This year’s agenda also features a keynote address by acclaimed journalist Bob Woodward of the Washington Post and Watergate fame, and an interactive plenary discussion featuring nationally syndicated columnist and PBS NewsHour political analyst Mark Shields and ABC News contributor and veteran Republican campaign consultant Alex Castellanos. Early registration rates are available through Sept. 20—register today.
|
Phone: 202-223-8196