What Are The Chances?

Among the regular entrants in the r-blogger universe, Arthur Charpentier is usually a math-heavy fanboy in his posts. Today is quite different. His main area of interest is actuarial matters. As mentioned here more than once, getting to be FSA is no mean feat. As I've also mentioned, I don't consider actuarial work product to be reliable. When there is stable historical data, which everybody uses just because the risk under analysis doesn't/hasn't changed from then to now, there's not much to it. The fact remains, as it always has been, that predicting off the end of time series data is at your own risk. Or your employer's.

The problem, of course, is that major losses are due to rare events, for which there is little to no historical data, in the financial sense. For each insurance sub-genre (life, P&C, auto, etc.), there is some data from which a trend (or, more to the point, an inflection) may be gleaned. Climate change has, and will continue to have, major influence over nearly every sub-genre of insurance, but P&C most heavily. IMHO. If one wishes to underwrite, it is far more important to be a subject matter expert than an actuary. That's just not debatable.

So, in all, here is the punchline, typos included (his native language, it seems, is French; not that this is a good excuse :) ):

So I want to show that the upper bound of the AUC is actually quite low ! So it's not a modeling issue, it is a fondamental issue in insurance !

For those in the insurance business, I suppose the attitude is; 'it's better to make the wrong decision with bad data than to make the right decision with qualitative analysis'. McNamara and Cheney and The Manchurian President are examples of decision making using neither.