Predicting insurance faud by machine learning procedures

Specialeforsvar ved Jesper Sonne

Titel: Predicting insurance fraud by machine learning procedures

Abstract:  In this thesis we consider insurance customers with a contents policy, who wrongfully seeks compensation from their insurance company. These customers are referred to as "frauds". The available data comes from a Danish non-life insurer. More specifically, we investigate if these frauds share any common properties and if it is possible to identify these frauds from information regarding the customer and thereby predict frauds. For this identification and prediction work we utilize a Generalized Linear Model and a machine learning procedure called Random Forest. The random forest procedure is our main focus and the idea is to find a model, that could potentially be used by a Danish insurance company. To utilize the random forest procedure, we need to introduce other methods such as decision trees, bagging and the random subspace method. An important ’a priori’ assumption is that data does not tell the full story, i.e. the number of frauds might be higher than observed according to the Danish insurance organisation Insurance & Pension (Forsikring & Pension)

Vejleder: Olivier Wintenberger
Censor:   Pierre Pinson, DTU