This master thesis analyses and describes Natural Catastrophes loss dependence structures between different sub-portfolios. Large natural catastrophe events are very rare but can lead to extreme losses. The available historical loss information is typically scarce, which is why it is difficult to analyse and proof structural dependencies based on real loss data. Probabilistic Natural Hazard Models provide not only estimations for single portfolio losses and the corresponding probabilities, they also allow to analyse the dependence structures.
Such models consist of scenario catalogues which represent the hazard probability distributions for specific natural perils (e.g. windstorm). Based on portfolio data and damage information for the hazard intensities at all locations, scenario losses are simulated for the portfolio exposed to the peril. An aggregation of losses from different countries but from the same events is the data basis for the analysis. In the investigated model, windstorm losses for ERGO portfolios in Germany and Poland are used.
An internal project of ERGO makes it essential to examine the dependence structure of occurrence losses. Therefore, we sample from the Occurrence Event distribution to simulate dependent losses for both countries. In order to examine the two dimensional dependence structure, different copulas are fitted to the sampled data and their goodness-of-fit values are compared. Additionally, the difference to the dependence structure for annual losses is explored.
At the end, a method to simulate dependent losses using the estimated copulas is developed and the performances of the copulas are compared and tested on the data with the method.
Keywords: Probabilistic Natural Hazard Models, Event Loss Table, Occurrence Losses, Annual Losses, Occurrence Exceeding Probability, Dependence, Copula
«
This master thesis analyses and describes Natural Catastrophes loss dependence structures between different sub-portfolios. Large natural catastrophe events are very rare but can lead to extreme losses. The available historical loss information is typically scarce, which is why it is difficult to analyse and proof structural dependencies based on real loss data. Probabilistic Natural Hazard Models provide not only estimations for single portfolio losses and the corresponding probabilities, they...
»