Utvärdering av IBNR-modellering för ett Litet Skadeförsäkringsbolag

This thesis explores and compares several methods for estimating Incurred But Not Reported (IBNR) claims, with a focus on their practical application in small insurance portfolios. Using data from a Swedish construction insurance provider, the study evaluates four reserving techniques: Chain ladder, Mack model, Bornhuetter-Ferguson and Tweedie Generalized Linear Models (GLM) with a stochastic simulation being applied to the latter. The analysis is conducted using both yearly and quarterly claim development data, allowing for an assessment of how model behavior changes with data granularity. While all models yield similar estimates for well-developed accident years, they diverge significantly for more recent and immature periods. In the quarterly setting, chain ladder and Mack overreact to limited data, resulting in highly inflated and volatile reserve estimates. Bornhuetter-Ferguson, GLM, and Tweedie demonstrate greater stability by relying on additional underlying assumptions. However, the volatility introduced by quarterly granularity -- particularly in the presence of sparse claims data -- led to unreliable outputs across all methods except for Bornhuetter-Ferguson which produced more stable estimates. As a result, the study favors yearly modeling for small portfolios, where stability and interpretability are crucial. The findings suggest that while methods such as chain ladder and Mack provide a useful baseline, other approaches may be better suited to the challenges faced by smaller insurers, particularly those that reduce volatility, like Bornhuetter-Ferguson, or provide distributional insight, like Tweedie.