Förbättra Försäkringspremie med a Priori Skattning och Skadehistorik

Accurate claim frequency estimates are central in non-life insurance modelling. Industry practice relies on a prior rating obtained from a GLM that uses static features such as building area, construction type, and age at inception.However, a policy’s actual claim history may deviate systematically from this prior,motivating the use of experience rating techniques-methods that refine the prior by learning from past claims. This thesis investigates how experience rating can be used to refine the prior claim frequency estimate,using a policy's own claim history.Three approaches are implemented and explored: (i) a Bühlmann-Straub credibility model with Poisson claim counts, (ii) a GLM that uses previous claim counts and prior estimates as covariates,(iii) an observation-driven dynamic Poisson–gamma state–space model, which is a models the mean of claim frequency by an AR-gamma model. The analysis is conducted on a dataset of annual records for commercial policies, each record paired with the insurer’s prior claim frequency estimate. All three approaches outperformed the baseline prior estimate.The dynamic model achieves the best overall performance, the GLM ranks second,and the Bühlmann-Straub model comes third.Performance is also examined for specific subsets of the data.The credibility model excels for extreme outliers,the GLM performs best for policies with long, moderately risky histories,and the dynamic model remains consistently strong across most scenarios.In summary, these outcomes confirm that integrating individual claim experience can effectively improve frequency predictions and, in turn, enhances risk assessment for individual policies.