Notwithstanding that millions tune in to the long-running reality TV show America's Next Top Model, the real modeling action is not in Hollywood. Instead, it is on computer mainframes churning out annual simulations of 100,000 years or more of catastrophes such as hurricanes, earthquakes and terrorist attacks. Such analysis drew the attention of the Maryland Court of Appeals in its seminal opinion last Wednesday in People's Insurance Counsel Division v. Allstate Insurance Co. (attached), which affirmed the appropriateness of modeling in an insurer's decision to issue, or not, homeowners' insurance policies.
The facts in Allstate were relatively simple. In 2006 Allstate determined that it would no longer write homeowners' policies on Maryland properties within one mile of the Atlantic Ocean. It subsequently extended that decision to completely exclude from new policies five Maryland counties, and portions of an additional six counties (all identified by zip code). It relied on a model developed by Applied Insurance Research, Inc. (AIR), which showed that the hurricane losses Allstate would suffer in the identified zip code areas were too high. Dutifully Allstate filed the appropriate papers with the Maryland Insurance Administration. The Administration found nothing exceptional about the application. The People's Insurance Counsel Division (PICD) (a part of the Office of the Attorney General) did, however, and requested a hearing. It lost before the Commissioner of Insurance, then before the Circuit Court and again before the Court of Special Appeals (see our post).
PICD then appealed to Maryland's highest court and argued before the Court of Appeals that Allstate had failed to meet its burden of showing that its decision was not "arbitrary, capricious or unfairly discriminatory." See Md. Ins. Code § 27-501(a)(1). Following from that, PICD further argued that the designation of areas by zip code did not have an objective basis and therefore was arbitrary and unreasonable. See Md. Ins. Code § 19-107(a). Allstate's proofs consisted primarily of computer modeling evidence, which the Commissioner found sufficient.
Much of the opinion is directed to the parsing of Maryland's Insurance Code and its legislative history to determine whether § 27-501 even applied (the Court of Special Appeals had found it did not, and the Court of Appeals reversed that portion of the decision). We leave it to the insurance blogosphere to address that further.
What is of interest to this readership is how modeling came into the decision and where modeling stands as a result.
In the proceeding Allstate offered a model that simulates hurricanes from genesis to decay and the damages that would be suffered. Basically, AIR modelers "developed mathematical functions that describe the interaction between buildings and their contents and the local intensity to which they are exposed." PICD at 7. Allstate established with expert evidence that catastrophe risk is not diversified ("adding additional catastrophe risk does not reduce overall risk because of pooling but actually increases the overall risk") and that historical loss data is incomplete and outdated "making it difficult to estimate losses." PICD at 7. Accordingly, "it has become standard practice for insurance companies to use catastrophe models to anticipate the likelihood and severity of potential future catastrophes before they occur." PICD at 5-6.
The advantages of modeling are substantial;
(1) It was able to capture the effects on catastrophic loss distribution of changes over time in population patterns, building codes, amounts insured, and construction costs;
(2) It provides a complete picture of the probable distribution of losses rather than just estimates of probable maximum losses;
(3) Because simulation models can be tested more easily than other approaches, it leads to greater stability in estimating expected annual losses;
(4) It provides a means to determine the impact of new scientific information; and
(5) It provides a framework for performing sensitivity analyses and “what if” studies. PICD at 6
As the Court noted, "By using computer models, they can get 100,000 years of simulated loss experience, which is good not just for State-wide pricing but also for loss characteristics related to hurricanes down to the ZIP Code level." PICD at 7.
PICD retained an actuary to rebut Allstate's proofs; he testified with respect to "actuarial science." He was hampered, perhaps fatally, when the Commissioner refused to allow him "to express any opinion with respect to the model that formed the basis of Allstate's amended filing." PICD at 11. We were not there but the Court of Appeals paints a picture of a non-committal expert. He offered that the decision to not write new policies was unreasonable "'because there is no showing that it is reasonable.'" And he "declined to choose" the method Allstate should have chosen to reduce its risk. PICD at 11.
In a post-hearing submission PICD argued that "Allstate was required to produce valid statistical data demonstrating the probability of a hurricane sufficiently strong to cause catastrophic damage actually making landfall in Maryland and that it failed to do so." PICD at 23. The statistical standard was based on dicta in an earlier Court of Special Appeals decision, Crumlish v Ins. Comm'r, 520 A.2d 738 (1987), which the Commisioner and the Court distinguished. First, Crumlish's requirement for statistical evidence was not a universal requirement. PICD at 25. More significant was the "catchall" exception added to § 27-501 which established a "standard approved by the Commissioner that is based on factors that adversely affect the losses or expenses of the insurer under its approved rating plan and for which statistical validation is unavailable or is unduly burdensome." PICD at 25.
"That is what the Commissioner did in this case." PICD at 25. In other words, the Commissioner found Allstate's evidence met its burden of demonstrating that its use of modeling as the basis to stop writing policies in certain areas was reasonably related to its business and economic purposes and was not discriminatory.
The dissent would have adopted the Crumlish dicta and required Allstate to offer statistical evidence concerning the landfall of destructive hurricanes in Maryland. PICD, dissent at 5. Such an assessment was either to be based on the historical record (an impossibility as no hurricane had ever made landfall in Maryland) or "climate science" (which one would think would include modeling). PICD, dissent at 9, 10. According to the dissent, all Allstate provided was a computation of the "relative risk" of a hurricane landfall in Maryland as once in 25,000 years based on the worst 5% of hurricanes that made landfall in North Carolina, Virginia, and Delaware. Allstate justified its decision based on hypothetical hurricanes, i.e., a model. PICD, dissent at 7.
The Court properly rejected this distinction. The use of probabilistic catastrophe risk modeling came of age following the destruction caused by Hurricane Andrew in South Florida in 1992. As stated by modeler RMS in its 2008 A Guide to Catastrophe Modeling (p6): "It became clear that a probabilistic approach to loss analysis was the most appropriate way to manage catastrophe risk. Hurricane Andrew illustrated that the actuarial approach to managing catastrophe risk was insufficient; a more sophisticated modeling approach was needed." Another modeling firm, EQECat, put it this way: "The main concern for all users is the uncertainties in the models. Some time ago, the only way to estimate a probable loss was to trust few statistical studies of past losses from some historical events and or on the experience of the underwriter. The uncertainty in these models was quite large as confirmed once a new event [such as Hurricane Andrew] took place. The main problem is that there is not enough historical data, and the standard actuarial techniques of loss estimation are inappropriate for catastrophe losses."
One of the purposes of catastrophe modeling is to assist the user (often an insurer) in avoiding the alliteratively named "risk of ruin." If all the industry is using a tool that can minimize the risk of run, it would ill-behoove a court to take away that tool. In Allstate the Maryland Court of Appeals agreed.
Nevertheless, if one is looking for guidance on how modeling will be received in the courts, there is one significant question left unresolved by this decision: how will competing models be treated? PCID's expert seems to have been completely out of his league. Whatever his actuarial credentials, if the issue is modeling then a modeling expert is needed. And at the very least the AIR model was subject to challenge. In a review published just this month, Assessing US Hurricane Risk: Do the Models Make Sense?, AIR takes on its competition, RMS, and states: "with this latest round of updates, we [modelers] find ourselves more divergent in our views of risk than ever." (p5) As one example of this divergence, "Catastrophe modeling companies have vastly different views on what influence sea surface temperatures (SSTs) in the Atlantic Ocean have on U.S. hurricane landfall risk." (p12). If AIR is correct, perhaps application of the RMS model would have altered the list of excluded zip codes. More fundamentally, does the uncertainty established by competing models (and that is inherent in modeling) impose an unavoidable and unacceptable arbitrariness in application? That is for another day. For the moment, modeling companies and those who use them likely will proceed full speed ahead.
Post scriptum - Climate change seems to have been a subject not to be discussed. As noted by the dissent, if Allstate was worried about the science of climate change, it didn't bring it up. Nevertheless, the dissent did bring it up and asserted that meteorological change occasioned by climate change could be a legitimate basis for Allstate's decision. The modeling firms think otherwise. Eqecat's CEO Bill Keogh has stated because of the uncertainty associated with climate change's effect on hurricanes, " it has no role in catastrophe risk modeling."
Peoples Insurance Counsel Division v Allstate Insurance Company.pdf (78.07 kb)