Deductible in Non-Life Insurance Risk Prediction

Specialeforsvar: Simon huge

Titel: Deductible in Non-Life Insurance Risk Prediction

Abstract: When predicting non-life insurance risk using GLM, deductible stands out among other policy characteristics. It truncates the risk the insurer is liable for, and it limits the data available to losses with values above the the deductibles. The GLM framework can’t accommodate distributions with non-zero truncation, making it theoretically incorrect to use on truncated data. This thesis begins with an introduction to the relevant background theory, the establishment of the metrics to assess the quality of prediction, and the presentation of the data to be used. Following that, the thesis will explore two approaches to handling deductible. The first approach is to ignore misassumption and go ahead with truncated data, using deductible as model covariate(s). This approach is easy to employ in existing GLM software, while providing decent predictive power. The second approach is to augment the data to restore the losses with values below the deductibles, and then model   using GLM, now without misassumption. This approach isn’t implemented in most existing GLM software, and is rather complex and computationally intensive to employ. The predictive power varies depends heavily on the specific procedure of data augmentation. The exploration of these two approaches leads to the development of several methods. One of the methods developed, sampling iterative augmentation, which repeats data augmentation multiple times, performs better than any other method investigated in this thesis.

Vejleder: Jostein Paulsen
Censor:   Mette Havning