Professional Insight

Actuaries Plunge Into the Big Data Pool

The story of big data — the oceans of data that modern technology generates — in some ways presents a modern cliffhanger. Now that we have access to all of that data, what can we do with it? Actuaries heard a variety of answers at the 2015 CAS Spring Meeting in Colorado Springs, where a trio of experts shared how to glean useful information from a world awash in selfies, tweets and status updates. Casualty actuaries now have the ability to tap into data and stretch beyond their traditional roles of pricing and reserving to help claims adjusters and marketers to do their jobs more effectively.

Philip Borba, a senior consultant at the consulting firm Milliman, showed how to find gems of insight in the standard claims report. Adjusters’ notes, he said, contain useful information that actuaries can use in predictive models to pick out which claims are most likely to turn contentious.

Borba, like most researchers in this area, breaks a standard narrative report into “n-grams.” N-grams are a single word or a short series of words. For example, the phrase “tested positive for amphetimines and marijuana” would yield several n-grams, including “tested,” “positive,” “tested positive” and so on.

Borba looked at 6,949 accident reports from the National Motor Vehicle Crash Causation Survey, a National Highway Traffic Safety Administration project that analyzed crashes. In the study, researchers wrote reports after visiting accident sites just after the crash occurred.

The 6,949 reports generated 13.3 million n-grams. Harnessing computer power to build a model, Borba looked for information on the use of cell phones and driving under the influence of medications. He found that narratives held important information on how often these two were linked to traffic accidents.

A separate study created a predictive model for workers’ compensation claims files, in which n-grams held clues to find “jumper” claims — ones likely to blow up. For example, an adjuster’s note 14 days after injury might note the claimant is “scheduled for an arthroscopic surgery,” Borba said.

The insurer “might not get that bill for eight weeks,” he said. Plucking that information from an adjuster’s narrative could help set a claims reserve more accurately and faster. Textual data like n-grams influenced almost a quarter of the model Borba built.

A second researcher, Douglas Wing, FCAS, an assistant vice president at Insurance Services Office, showed how insurers use computers to study visual information.

Computers see photos differently than we do. For people, a picture is a signal that helps them remember what an object is, Wing said. To a computer, it’s a large number of pixels or a unique jigsaw of colored polygons.

In a process called image segmentation, the computer turns a picture into a series of polygons. A photo of a tree is changed into thousands of polygons, one for each leaf, one for each branch. Another process, feature extraction, lets the computer find common shapes — eyes or ears, for example.

The process in essence turns the computer into a super-sophisticated set of eyes, which Wing says insurers are beginning to take advantage of.

In homeowners insurance, for example, photographs can show the area of a home’s roof, along with its pitch and type of roof — all helpful in underwriting a policy or settling a claim. It’s expensive and dangerous to measure a roof by hand, Wing said, particularly in winter.

“You don’t have a lot of people volunteering to go on the roof in the cold and in the snow,” he remarked. But a computer can read a flyover photograph, identifying roof lines, chimneys and vents, all of which underwriters are interested in. After a disaster, Wing said, a computer can compare before-and-after photos to see which homes are damaged and what the insurers’ overall exposure is likely to be.

Auto insurers can use the technology as well, Wing said. Claims on many damaged cars can be adjusted with photos alone. A computer analyzing a damaged vehicle could settle about 40 percent of claims within a day, Wing said. Often the insured could take the picture, reducing the time and cost of settling a claim.

Computers already store millions of license plates, as vehicles scour parking lots, photographing every plate. The practice began as a way for repossession dealers to find cars. Insurers could use the same information, Wing said, to recover stolen autos or validate that a car is garaged where the insurance policy says it should be.

“This is already starting to happen,” Wing said. “We need to start leveraging it.”

Roosevelt Mosley, FCAS, a principal at Pinnacle Actuarial Resources, described how Twitter yields valuable information on insurance marketing. Social media outlets like Twitter, Facebook and LinkedIn provide a candid window into the conversations and opinions of millions of people, he said. “Current and potential customers are sharing sometimes intimate details of their life with the world,” Mosley said. Insurers have the opportunity to observe and react.

Companies can respond to online cris de coeur or passionate public protests as part of their customer service practice. They can listen, tapping into customer sentiments. They can monitor and pick up on broad market trends.

Data mining, Mosley said, is a “virtual focus group.” A company can put an ad online and then see how consumers like it.

Mosley contrasts two GEICO ad campaigns centered on visual puns: a camel proclaiming Wednesday as “Hump Day” and a pig that does strange and yet not strange things like riding in a car shrieking “whee!” or draping itself in a blanket at a football game. The camel ad campaign garnered a tremendous social response when it was launched, Mosley said, but interest leveled off quickly, even as the company tried to reignite it. The ads with the pig “had a much longer run,” he said, even though the ad provoked extreme responses. In fact, Mosley believes the pig ads have endured because they are divisive.

Mosley has also used social research to understand usage-based auto insurance, in which companies use a telematics device to monitor driving patterns. The question: How much of a discount do customers want before they think installing a device is worthwhile? It was no surprise, he said, to learn that people with higher discounts were more satisfied with the program. More surprising, however, was that the size of the discount was not as important as the size of the actual discount compared with the discount customers thought they should get.

Monitoring social media has the advantage of being unfiltered, Mosley said. “People are sharing raw emotional responses on both the good and the bad side,” he said.

On the other hand, that means insurers have to take care to understand what is driving the strong feelings; this is a process that “can get really tricky,” he said, but one that is worthwhile.

“Instead of having to guess what your policyholders want or what your customers are thinking,” he said, “sometimes you just have to do a little digging and you can find out.”


James P. Lynch, FCAS, is chief actuary and director of research and information services for the Insurance Information Institute in New York.