Non-traditional Approaches to Predictive Analysis

Property/casualty insurance actuaries should not be afraid to try alternative approaches to data mining as part of the predictive modeling process, lawyer and economist told the Casualty Actuarial Society (CAS) Ratemaking Seminar in Las Vegas.

Ayres, author of the book “Super Crunchers,” noted in a keynote address that non-traditional approaches to predictive analysis, such as neural networks, might have a role for actuaries, subject to regulatory constraints. “You should be thinking about trying alternative approaches, even if your central approach is general linear regression,” he said at the seminar. “Every once in a while I would try a neural network and see if your traditional approach is robust to alternative specifications.”

He went on to explain that as the size of datasets has increased, neural networks may be able to estimate many more parameters than traditionally accommodated by linear regression. He also cited the example of Epagogix Ltd, a UK-based company specializing in artificial intelligence, and its ability to forecast the box office success of movies by using a neural network model.

According to Ayers, a studio gave Epagogix the scripts for nine movies and asked the company to make their predictions on the box office revenues before a single frame was shot. Independently the studio also made their predictions. While Epagogix wasn’t perfect, it made accurate predictions on about six of the nine movie scripts—twice the accuracy of the studio.

“If Epagogix can be successful at number-crunching on a very high degree of difficulty question with relatively little data, it shouldn’t be surprising that people in this room can do a much better job on trying to score out some insurance risk when we have much larger data sets,” Ayers said.

Ayres observed that a chic approach in certain number-crunching areas is to draw on the power of the collective wisdom of crowds to make a prediction. However, he suggested that true wisdom lies in mining a company’s historical data.

Read more…