Aller au contenu. | Aller à la navigation

Études et Formations
Service à la société

Abstracts Session 3


Julie Huyghe (ULB): Performance curves and Autocalibration 


Performance curves (Lorenz Curve & Concentration Curve) are used to compare the performances of different models. It appears that the fact that these curves coincide is equivalent to the important concept of autocalibration (Gneiting, 2010; Kruger & Ziegel 2019). This allows us to build a testing procedure for autocalibration (paper in progress). An autocalibrated predictor is desirable in any setting where a global balance is desirable, that is, where it is important that the sum of estimates does not deviate too much from the sum of actual observations at both the entire data base level and more locally, in meaningful classes.
Moreover, under auto-calibration, forecast dominance reduces to convex order and the Gini index as well as the ICC (integrated concentration curve) gives strictly consistent scoring rules, which is not the case in general.

Charlotte Jamotton (UCLouvain): Insurance analytics with clustering techniques


The k-means algorithm and its variants are well-known clustering techniques. In actuarial applications, these partitioning methods can be used to detect clusters of policies with similar characteristics. The resulting partition then provides an actuarial framework on the basis of which a map of dominant risks can be drawn up. Despite being a well-studied area, clustering has not been the subject of much focused actuarial research. The aim of this research paper is to adapt well-established clustering methods to complex insurance datasets that contain both categorical and numerical variables. To that end, we develop a novel approach based on Burt distance. The first part of this paper starts with a review of the k-means algorithm to lay the foundations for our Burt framework. The second part extends the scope of application of the mini-batch and fuzzy versions of the k-means to insurance data. The last part shifts the focus to spectral clustering, a technique based on graph theory that allows for non-convex cluster shapes. For large-scale datasets, we propose a prior reduction of the data using our Burt-adapted k-means to work around the computational complexity.