AltInst: insurance followup

A while back I publically wondered here why insurance companies use
such coarse risk-assessment categories, why they don't use machine
learning to the extent of, say, credit-card companies. And speculated
that it might be due to regulatory constraints.

Talking to a fellow who consulted on claims modeling for a car insurance
company provided some insight. The idea is to use known independent
variables like history, age, car type, etc. to predict certain
parameters allowing one to estimate expected claim value.

One might, for example, try a memory-based approach (basically a fancy
name for averaging values in the query point's neighborhood). But
this is not an option, because it's a black box, not model-based. You
want a model that you can not only get by regulators, explain to
managers, check against actuaries' intuitions, and even run by hand,
not to mention explaining them to customers.

So not only is model-based learning needed, but the model has to be
simple. Insurance companies prefer binary axis-oriented discriminators,
not anything smooth or multi-variable. So these folks ran a decision-
tree learning algorithm to get a whole ton of simple rules. A good
number were novel, and retrospectively saved a lot of money. A few
made it past the actuaries and salemen to the lawyers. I don't know
whether any were implemented.

Interesting story about organizational learning...

-- 
     Eli Brandt  |  eli+@cs.cmu.edu  |  http://www.cs.cmu.edu/~eli/
[To drop AltInst, tell:  majordomo@cco.caltech.edu  to: unsubscribe altinst]