Predictive modeling has forever changed the way insurance policies are priced. The tool allows insurers to design more sophisticated models that tap more detailed data sets to refine precisely how much each customer should be charged.
Casualty actuaries got an overview of how far the revolution will change insurance pricing at the session,”The Revolution and Evolution of Predictive Modeling,” presented at the recent Casualty Actuarial Society (CAS) Spring Meeting in Phoenix.
Claudine Modlin, senior consultant at Towers Watson, laid out how predictive analytics has advanced pricing in the past decade. Steven Armstrong, a CAS fellow, detailed ways the tools could improve insurance operations beyond the pricing function.
At the end of the 20th century, Modlin said, insurers were bound to mainframe computers and highly aggregated data sets. It was easy for a company to understand its competitors’ plans. And rating plans were finalized based on the collective judgment of underwriters and actuaries, with little data-driven guidance in how and where to deviate from the expected costs.
Today, insurers use predictive analytic tools to hunt through gigabytes of data to find sometimes non-intuitive variables that hold clues to a customer’s riskiness and purchasing behavior. Generalized linear models (GLMs) have become the global industry standard for pricing segmentation. This is due in large part to the multivariate framework, multiplicative nature of rating plans and high transparency in the results.
“As an industry, we have really learned a lot,” Modlin said. “We have advanced our toolkit.”
Next Great Loss Predictor
The use of insurance credit scores was a great loss predictor in the past two decades, and the search for the next great loss predictor is ongoing. As insurers follow the information revolution, they are improving the quality and accessibility of their internal data, investigating third-party data sources, and investing in more computing power to harness the information. This has led some companies to investigate thousands of predictors — including what other policies an insured has, whether they pay their bills on time, and various characteristics of the area in which the risk is located, etc.
Modelers employ a variety of techniques to cull the list of potential predictors. The process of variable reduction involves business judgment but is frequently supplemented with statistical data mining techniques, such as principle components analysis or classification and regression trees.
Companies looking to refine GLMs further pay attention to identifying interaction variables and to mining GLM residuals to improve the pricing of high-dimension variables (e.g., territory and vehicle groups).
In auto, insurers are starting to use telematics — gathering information about driver behavior from a device in the vehicle. Information will flow in, virtually moment by moment, Modlin said. “Do you slam on the brakes? Do you peel around corners?”
As much of the industry has refined loss cost estimating, the use of science to understand customer demand lags. GLMs can help. The challenge is to capture customer attributes and price-related information (e.g., quote offered at new business or price change offered at renewal) that will provide insights into customer elasticity.
The next evolutionary stage for pricing sophistication is for companies to learn to integrate their cost estimates with knowledge of customer behavior. This can involve scenario testing rate changes and measuring the effect on key performance indicators, taking the effect of customer behavior into account. Scenario testing involves price optimization techniques that systematically integrate cost and demand to indicate an optimal set of prices that meets or exceeds corporate objectives for profitable growth, while staying within company constraints.
Predictive model use doesn’t have to end with ratemaking, Armstrong said. Models can help in other areas. And actuaries can explain how the models work and what potential they contain.
“I want you to think beyond pricing” and help solve business problems, he said.
For example, predictive models could help underwriters work more efficiently. For instance in auto insurance, young drivers receiving good student discounts have to regularly turn in copies of their grades. Predictive modeling could show that some types of students don’t need to perpetually update, while others would.
Armstrong said models also could help underwriters determine which workers’ compensation risks should be tapped for a premium audit.
Predictive modeling also could help marketing by researching what mix of social media grows the customer base or what brand attributes drive new business. The concept isn’t new to marketers, but the actuarial skill set can enhance understanding of the work.
And claims departments “swim” in a vast, vast pool of data that only awaits discovery — claims diaries, records on attorney involvement, and information on service providers and adjusters, Armstrong said.
The list of areas where actuaries could help insurers quantify and understand their operations seems limitless, he added.
“Wherever there is data, there is opportunity.”