Use of Classification Analysis for Grouping Multi-level Rating Factors

Cristina Mano~Elena Rasa, Brésil

The Generalised Linear Models (GLMs), devised in the 70’s together with significant advances in computer hardware caused a near revolution in personal lines pricing. GLMs can be used to estimate price relativities for a number of rating factors with different applications in many lines of insurance business.For some of the rating factors, it is necessary to explicitly define or revise the number of levels used in the analysis. This process of definition takes place at several points during the analysis. For rating factor with few levels, we can combine levels with low data volume, unstable parameter estimates or similar characteristics to get more reliable estimates. For rating factors with many levels without an inherent ordering and with moderate or sparse data, the grouping approach is not as obvious. The problem arises because there are too many rating factor categories to directly include in the statistical model. The procedure for defining groups for car model, an important rating factor in private automobile insurance, considering that there are thousands of different car models, is more complex than for rating factors with few levels. Territory is other important rating factor with a large number of levels, represented by the postal code. For this variable, physical proximity does not imply necessarily similar risk characteristics. The aim of this paper is to give a brief description of the main Classification Techniques that can be used in conjunction of the GLMs to define groups for multi-level rating factors and compare the results obtained with these techniques with alternative approaches as Credibility Theory and Multivariate Spatial Analysis. After presenting the methodologies, the paper turns to practical applications in private automobile insurance.
Date: 30 May - Time: 10:30 to 12:00 - Room: 251
Theme: 9.A. Various topics