By CYRIL TUOHY, managing editor of Risk & Insurance®
On December 1,
Managing Editor Cyril Tuohy interviewed Bayes Corp. Chairman Jim Sewell and President Nigel Cooper in Philadelphia about the advantages of Bayesian algorithms. First developed by Thomas Bayes in the 18th century and found in a host of modern applications, Bayesian algorithms help to analyze probabilities rather than frequencies or physical properties.
Q: As a mathematician, what attracted you to Bayesian logic to begin with?
Jim Sewell: It was a stroke of luck. Back in 1979, I'd been responsible for producing the inserts to send mail out at USAA. The marketing director said, "Can you get me the best 50,000 names for this personal insurance product?"
I said, "The best 50,000? No, I can't do that," and I stuck that little note up on my desk and it was up there for a year.
I had an occasion to invoke Bayes' formula, and, well, that was the answer to that question.
Q: Back then, were insurers applying Bayesian logic for any purposes?
Nigel Cooper: No. We started the company in 1992. We were the bastard child.
The classical statisticians would look at Bayesian logic and say its kind of intuitive stuff. There's a lot of logic in the search engines. What we're doing is, we're really keeping it very simple. We're taking all the data and coming up with the next probability.
The biggest thing is talking to the client and getting the data. Bayesian logic is very forgiving. In the real world, data is screwed up, it's dirty.
Q: Is that an advantage for people who deal with insurance
The output of the model is really about decision support. It's helping someone do their job better. One of the big insurance companies had 400 people looking at subrogation. We got it down to 250 people within 6 months. The way we did it was that, in most systems where claims adjusters are looking at incidents, 20 percent of the people do 80 percent of the effective work.
We took that sample of the 20 percent of the successful claims, and we score all the claims as if only that 20 percent were looking at it.
Q: If Bayesian probability models are as valuable as you say they are, why use regression models, which focus on the relationship of variables, in the first place?
Mainly because people have been trained to use certain things. Regression models are not bad. But the Bayesian models tend to be more flexible and work well with the data you have available. Bayesian models can also give you an answer relatively quickly.
Q: It sounds like the flexibility of the Bayesian model is analogous to the difference between the Web platform and the mainframe, is that right?
NC: That's right. Even when you have a sample, you might have the head of risk biasing the sample by saying, "I want you to look at all the women with disabilities or all the men out on workers' comp leave," so the actual results might be biased.
One of our biggest jobs is to find out what problem the client is really trying to solve.
Q: It sounds like there's some softening toward the Bayesian model and people are more accepting of it.
Oh, yes, they are. It started out with Bayes' Theorem, which is simple high school statistics, but it is embellished now to an entire field of statistics. It's very complex, and it is perfectly consistent with classical statistics.
One of the things we've had to do is get away from the idea of modeling and scoring to what's the information? What decisions do you need to make? How can we get that in front of you versus getting into how the model works.
You don't need to know how the clock works. You just need to know that the time is accurate.
Q: I would imagine that this whole idea of being able to move quickly is much more valuable than a be-all and end-all folks saying we'll come back to you
in three months.
We've been able to say, "We'll increase your subrogations, and of the hundreds of millions of dollars that come in each year, we're going to be able to increase that."
We don't want to get paid for analytics. We want to get paid for success.
December 3, 2009
Copyright 2009© LRP Publications