Have Data .. Will Mine

I recently recalled a very amusing episode. This occurred way in the past. I had spent the whole of the morning with a general insurance customer discussing claim analytics and claim prediction for automobile insurance. From there I rushed off to meet another client who operated in the life insurance space.

The client was busy with some worksheet data. I asked him what he was up to. He said he has received scores for a new model. I asked him which model is he building now and he told me it was claim prediction. Since I was with a general insurance customer, my mind was still oriented to the general insurance business. Instinctively, I asked him what his definition of claim was. He said with a smirk that claim is when the life assured dies. Then, it hit me that I am sitting in a life insurance business premises. We joked about the fact that we are actually trying to predict the death of a person. We laughed about what the output can be used for. One option was seeing that a person is predicted to die, the company can refuse to take his renewal policy and let it lapse. Imagine the call center interaction --- "Hello sir, since we see you are not likely to live over the next 18 months, we would like to terminate the life insurance policy. Thank you for being a good customer while you were alive."

I hinted that it was a pretty sadist model that he was building. Anyways, we got to the worksheet and I asked him who did this model. His outsourced analytics agency built this model. I asked him to show me which variable was most dominant in generating the claim score. I was not surprised to find that age was the dominant variable. It showed younger customers were less likely to die than older ones.

While the client understood that this was not the right approach, it was amusing that the analytics agency actually built a model for claim prediction. It was a true case of "since I have data, I will build some model". The agency did not question nor advise the client on the right approach to solve the business issue. The client wanted to arrive at expected expenses, including claim over the next couple of years.

In life insurance, the amount of data about the customer is very limited. Claims occur on termination of life (we are not discussing riders here). Length of life depends on quality of life, which in turn depends on various factors such as diet, lifestyle, etc. And not to mention homicides. Information which is not available in the life insurance database. Hence, the approach is to go macro or at a higher level. One should look at mortality of the target market and then draw a proportion of the policy base from the target base. This will give a good estimate of the number claims likely to come in. The analytics agency should have done a forecast of the deaths in a target location. They could have done this either (inside out) by taking the past experience of the life insurer and extrapolating it over the market for forecasting or (outside in) by taking the deaths registered in a target location and then adjusting the forecast for the profile of people buying policies with the insurer.

But instead, we had a typical mentality of a kid who when given a hammer thinks everything is a nail. Since there is data available, a model was built. On this topic, do you know "the salary of a product manager is inversely proportional to the unit price of the product." (findings from one of the models when I was a analytics infant).