Mathias Graf says natcat models still depend on quality data

12 Jul 2016

The Head Cat Research and Development, Zurich Insurance, presented “Common misperceptions of models: the uncertainty of tail risk" at the Catastrophe Knowledge Exchange event at the Swiss Re Centre for Global Dialogue.

Click here to find out more about the Catastrophe Knowledge Exchange event.

Mathias Graf led the session by showing that for a return period up to 20 years, the risk can be validated by claims data, the risk up to a return period of around 100-200 years can be validated by scenarios (at least for meteorological perils), but tail events above a return period of 200 years are mainly driven by model assumptions and the considered uncertainties. The group had a look at the uncertainties involved in the whole risk estimation process including exposure data, hazard model, vulnerability mode, financial model and model selection uncertainties. Among the questions addressed were: (i) how can the uncertainties be validated? (ii) how can the uncertainties be reduced? (iii) how should the uncertainties be treated and communicated?

Read the text version of Mathias Graf’s video interview below:

It's hard to say if the insurance industry would be ready for another country but I definitely think in the last ten years the cat models significantly improved. We're much better in estimating the risks and therefore I think we have a much better handle on the risk itself but it is also dangerous to over rely actually on the existing cat models. It's crucial that we develop our own view of risk to validate and calibrate the models. Additionally we should have an accumulation control as a second check and also to cover regions where we don't have any cat models.

As the natcat models are continuously improved especially on the hazard side, it has new scientific insights coming in there. The resolution is increasing. They are utilizing the computational power but my point of view the most room for improvement will be the vulnerability models, and the financial models, and also the uncertainties considered within the models. They are only as good as a claim data that it flows in. Luckily as a primary insurance we have a lot of claims data that help us actually for the validation and calibration of the model, and also derive our own view of risk.

What is challenging about natcat modeling is in Zurich we license our vendor models on one hand side. On the hand side you have the dedicated research team which we are trying to validate and calibrate the models to come up with our best view of risk. Our main challenges are definitely the lack of transparency in the models, and also the closeness of the platform, and also given that to be able to have a perfect risk model also a challenge a model is only as the input data so we're trying to also improve our exposure date. We define a standard to measure the quality and also have a standard format."

Click below to read the themes from the first Catastrophe Knowledge Exchange:

Learn about risk profiling with Swiss Re's CatNet®