Ztráta informace při agregaci dat v modelech závazků neživotního pojištění

This thesis is focused on comparison of two models for calculation of IBNER reserve. The first model is based on aggregated data and it is worldwide commonly used. We focus on Frequency-Severity model, which is derived from chain-ladder method. The idea is to study separately frequency as well as severity of claims a and predict relative one-year change of reserve for open claims. The second model is based on non agregated micro data, using data as granular as possible. This proposed "micro" model is based and predict relative one-year change of reserve usage of generalized linear models (GLM). Both models are compared using mean squared error (MSE). Bootstrap random resampling is used for generating of plenty of „new“ datasets to evaluate both models on many different datasets. For practical implementation a dataset of MTPL claims is used for all calculations. It was proved, that aggregate model provides significantly better results than micro model does.