A Neural Network Boosted Double Over-Dispersed Poisson Claims Reserving Model

Abstract

We present an actuarial loss reserving technique that takes into account both claim counts and claim amounts. Separate (over-dispersed) Poisson models for the claim counts and the claim amounts are combined by a joint embedding into a neural network architecture. As starting point of the neural network calibration we use exactly these two separate (over-dispersed) Poisson models. Such a nested model can be interpreted as a boosting machine. It allows us for joint modeling and mutual learning of claim counts and claim amounts beyond the two individual (over-dispersed) Poisson models. Moreover, this choice of neural network initialization guarantees stability and accelerates representation learning.