Podcast: Predictive models hold promise for property-casualty insurersOctober 2 2018
The Society of Actuaries’ Listen At Your Own Risk podcast series explores today’s most interesting questions of risk. Recently Gennady Stolyarov II, FSA, ACAS, MAAA, who is the Lead Actuary for Property and Casualty Insurance, in Nevada’s Division of Insurance, joined host Andy Ferris to discuss how predictive modeling works and how its thoughtful application can help private property and casualty insurers achieve their goals while improving consumer outcomes and the availability of insurance – as well as some challenges involved in predictive modeling and approaches which can improve the quality of predictive models. LISTEN TO THE FULL PODCAST
What are some examples of predictive modeling applications in property-casualty insurance?
Predictive modeling has been used in a variety of applications from underwriting and rating of insurance policies, to marketing and claims handling over the past two decades. The rate of development and adoption of predictive models is certainly increasing.
The majority of these models are still focused on credit-based insurance scoring. They consider information in the credit histories of individuals and develop scoring algorithms that seek to forecast on an aggregate statistical basis the risk of insurance loss in a personal line of insurance, generally auto or home insurance.
An emerging field is usage-based automobile insurance. We have certainly seen a great proliferation of models that seek to measure driving behavior, for instance, the speed at which one drives, the times of day at which one drives, braking and acceleration behavior. Original usage-based insurance models were based on devices that could be plugged into the onboard diagnostic port in a vehicle and then relay data back to the insurer. Some newer usage-based insurance models are compatible with a person’s mobile phone, so an individual can just download an app and have their driving behavior measured that way.
There are also certain location-based models that use geographical and/or demographic characteristics of an area to develop more granular territorial rating for auto and home insurance, and then there are catastrophe and other peril-specific models that seek to estimate the risk of earthquakes or wildfires or wind and hail types of losses. For instance, Nevada does not have any hurricanes, so hurricane models are not deployed here; but in other states, hurricane models may contribute considerably to the pricing of a home insurance policy.
Another type of model that has emerged recently is the so-called price optimization model. A price optimization model determines the extent to which a selected relativity or rating factor that an insurer uses moves toward an actuarially indicated relativity or rating factor because an insurer isn’t always going to adopt the exact indicated factor. It’s important to note from a regulatory standpoint that these models are still bound by the same requirements that apply to any other predictive models in that they may only consider characteristics related to the risk of insurance loss. So in my review of a price optimization model, I would seek to ensure that that model remains risk-related and that it does not consider characteristics like price elasticity of demand or the tendency of a consumer to complain or the tendency of a consumer to shop for insurance which are not related to how likely a consumer is to experience a loss or file a claim in a particular line of business.
QUESTION: What are some of the biggest ways predictive analytics benefit the property-casualty arena?
Gennady Stolyarov II:
The most beneficial possibilities for predictive modeling are still on the horizon, and we may see these possibilities come to fruition within the next several years or within the next decade.
The most significant promise is to improve the availability of insurance to individuals and businesses for whom manual underwriting processes may have been considered too costly in the past. For instance, small businesses have historically had difficulties finding workers’ compensation insurance and have needed to resort to the more expensive assigned-risk market, even if the actual operations of the small businesses are quite low-risk—like for a predominately internet-focused or office-based startup company. If a predictive model can reduce the time and expense of underwriting or rating these policies by generating a risk score that reasonably relates to the prospective risk of insurance loss, then this could improve the appetite of certain insurers in writing these small business risks.
Similarly, predictive models that focus on claims could come to recognize signs of straightforward and uncontroversial claims, which are the vast majority of claims, and fast-track those claims for expeditious payment. Various insure-tech firms are already attempting to do this with their claim-handling predictive algorithms. Some predictive models that focus on the actual activities that contribute to the risk of insurance loss, such as driving behavior with usage-based insurance or the operation of a drone with by-the-hour drone insurance policies, can eventually result in a tighter connection between insurance pricing and the behaviors that cause or mitigate loss.
Historically, as we know, the insurance industry has relied on large numbers of what I would call proxy attributes, some of which are connected in only a circumstantial or general manner with the risk of insurance loss. The use of such proxy attributes has generated much controversy because the public does not intuitively understand, for instance, how information in consumer credit reports relates to driving behavior or home maintenance. So my hope is that the next generation of predictive models could actually result in a paradigm shift away from the increasing complexity of these proxy attributes and towards simpler, more direct structures that consider the most direct drivers of insurance losses or other consumer behaviors that insurers consider important to measure and anticipate – for instance, indicators of a claim that could lead it to be processed very quickly and efficiently versus indicators that might lead an insurer to seek to investigate it in greater depth.
With developments in autonomous and semi-autonomous vehicles, smart home technologies, and many other types of connected devices, insurers will be able to consider data that are much more closely related to these consumer behaviors. And while the accurate and reasonable consideration of such data may help many consumers reach outcomes that everyone finds logical and equitable, issues such as privacy, consent, and data ownership remain crucial to reflect upon and keep in mind throughout the process of developing these predictive models of the future.