Predictive Modeling: Art or Science?

By John Ammendola

Predictive modeling has made its way into the mainstream insurance arena and given the industry an innovative way to manage risk. From regional carriers to national enterprises, you’ll find many, including Grange Insurance, are discovering modeling to be a useful tool in determining risk and setting rates.

Calculating insurance pricing has always been a complex process — as much an art as a science. Gone are the days when underwriters would gather limited information, manually evaluate applicants, and hope their conclusions were accurate or at least close. Admit it, you remember those days.

Then (thankfully), the banking industry introduced automated underwriting with a computer-generated loan underwriting decision. Using completed loan application information, an automated system retrieved relevant data, such as a borrower’s credit history, and arrived at a logic-based loan decision.

The insurance sector witnessed the results of this automation and jumped on board, customizing the tool for the insurance industry. After seeing success, insurers realized the foundation existed to develop technology further for their benefit. Progress, right?

We certainly thought so at the time, but advanced technology has since afforded the industry a far more innovative tool in the form of predictive modeling. While not a perfect science, when implemented effectively it goes a long way in setting accurate rates and managing risks that help companies like Grange compete in the marketplace. And as predictive modeling evolves into two different kinds of models based on one or more layers of segmentation, the question becomes which is more effective?

The Science

Many midsize to large insurance companies, including Grange, are taking predictive modeling to a new level. The result? Better segmentation, more competitive pricing, financial stability, increased ability to predict loss costs, and, ultimately, the ability to better manage risk and grow business. There’s no doubt the tool can be extremely valuable, but if you’re about to take the modeling leap, which type do you choose: univariate or multivariate?

Univariate analysis, as the name implies, uses one layer of variables (a + b + c = $X) to segment risk into large buckets. While this subsidy model is a step in the right direction and today’s standard for many small companies in the industry, its ability to zero in on specific demographics falls short.

And so along comes multivariate predictive modeling. What’s the difference? In one word: segmentation. The new model exponentially increases the power of predictive modeling by refining targets and homing in on specific segments. The result? The ability to better predict loss costs, including frequency and severity, as well as to price more effectively based on the behaviors of specific groups. So now we can look at, for example, location, vehicles, history, or credit to verify objectively information in a cross-functional analysis. And all that translates into more competitive rates and better business.

Is It a Perfect Science?

Perfect is a strong word, and predicting behavior is extremely complex. The key is to trust your model. Adjusting the variable input without changing the core “predictiveness” is where the art comes in. Test your model at least once a year, and update your variables every six months. It’s a simple formula:

Closer to Predictive Truth = Better Pricing = More Stability Beyond

that, remember nothing can take the place of product management in the underwriting cycle. Models don’t make you better at pricing unless all other factors, including the human element, are also in order. Together, they create a powerful advantage in the marketplace.

The Marketing Edge

Predictive modeling can also affect marketing efforts because of the segmentation it creates. Overall, modeling allows Grange to focus on a desired target market and effectively compete within a certain group or demographic. Identifying this “sweet spot” also helps us target our messaging and make that message more relevant to prospects. Anytime we can increase our understanding of customer behaviors, our communication becomes more meaningful and effective.

We know predictive modeling helps determine:

  • the relationship between the number of products a customer has and increased loyalty
  • customers most at risk for attrition over the next 6 to 12 months
  • which prospects within our geographical region are most likely to become customers
  • which products would be most attractive as introductory offerings for various segments

Taking all of this into consideration, Grange recently pursued a targeted customer retention initiative. Our auto policy retention model scores policies for expected retention probability and can be used to identify at-risk policies for intervention. Those with below-average retention probability scores were targeted for marketing intervention, which included a letter thanking them for their loyalty and communicating the premium-free benefits of their existing policy. A control group remained separate from the intervention to provide a baseline for measuring impact. The selection of the control group policies was random to ensure that differences in retention outcomes resulted only from the intervention itself.

The intervention increased retention rates by 3.45 percentage points, a statistically significant result. Further, because of the financial leverage of retention, “saving” the estimated 206 policies produced $113,000 premium life in the next module, net of direct cost of the intervention. Projecting the effect forward five years, net present value reached $271,000 after direct cost and losses.

Grange also used predictive modeling results to acquire new customers by creating “personas,” or groups of like prospects who share similar characteristics and behaviors. As a result, we can personalize relevant messaging to introduce Grange and our offerings.

The Result

Having seen the results of both univariate and multivariate models, it’s clear the multivariate alternative provides better segmentation and more comprehensive information. When we compare the two, multivariate’s ability to predict consumer behavior and risk is far more concise and capable of pricing risk accurately. The lift is right on par, which shows us we’re ahead of the game.

And whether we’re using multivariate modeling for assessing rates and risk or for marketing purposes — or both — it signals increased ROI and growth for the organization.

For Grange, this is unquestionably how we remain relevant and competitive. Predictive modeling gives us a unique edge and enables us to innovate continually for answers that differentiate Grange in the marketplace.

John Ammendola is president of personal lines and corporate research and development for Grange Insurance. Based in Columbus, Ohio, Grange Insurance has $2 billion in assets and in excess of $1 billion in annual revenue.