A Report from GIRO Advanced Pricing Techniques (APT) Working Party

Document description

This paper examines in details six overlooked facts of using generalized linear models (GLMs) for non-life pricing and discusses whether the use of GLMs is still fit for purpose in a competitive market. The six facts of GLMs discussed are

1. Either zero or full credibility is given to the data and there is no way to do blending

2. Prediction of a risk depends on data in other completely independent segments

3. Model predictions depend on the mixture of rating factors in the data

4. Maximum likelihood estimate of prediction is lower than mean of prediction distribution

5. Link function could bias the model prediction and significantly change the lower and upper bound of prediction

6. Model diagnostics is only relevant in the segments where the model is used These facts indicate GLMs’ technical limitations in pricing practice, particularly in a competitive market. It is critically important to understand these limitations when apply GLMs in pricing.