Insurance fraud detection in automobile insurance

We will focus on creating a series of time dependent models for detecting fraudulent claims depending on the level of dynamic information available, and fine tuning these models before testing them with live data and putting the retained models into production. Our objective is to better filter our actual label-claims to form a better control group on which we can train robust classifiers that will detect fraud. During the next months we are going to filter our data by 1) identifying business rules that trigger an automatic classification, and 2) delete variables that cannot be used to detect fraud. We will then apply robust classifiers to our new filtered data. During the last months of the project we will 1) identify relevant events threshold in a claim’s life that will trigger a prediction from the system, and 2) proceed to algorithm selection, fine-tuning, live tests, and putting the retained models into production

Raphael Zerbato
Superviseur universitaire: 
Georges Dionne;François Bellavance
Partner University: