Technical break-through in the CatFocus® European windstorm model: Advanced risk assessment and validation of windstorm return periods using a dynamical, multi-model storm set.
Major exposure, limited data
Seven of the 40 most costly worldwide insurance catastrophic events1 occurring between 1970 and 2008 were European winter storms, including for example Lothar (1999) and Erwin (2005). These storms have the power to cause extensive property and environmental damage and loss of life (the estimated industry loss from Lothar was 5.77 billion euro2).
As there are only approximately 50 years of quality storm observation data, assumers of European catastrophe property risk use catastrophe models to generate extended event sets by either statistically or dynamically perturbing real storm data. Dynamical models, despite immense computing power requirements, represent a major improvement in that they are physics-based, generating physically realistic storms that are not constrained by the attributes/parameters of historical windstorm events.
Access to dynamical model data, computing power, and the scientific skill required to interpret and calibrate the data to historical events, make analyses that use this type of information problematic. PartnerRe has now successfully done such work and the results provide invaluable input into our European windstorm catastrophe model.
PartnerRe breakthrough – the first to use a multi-model storm set for risk analysis and validity checks on European windstorm loss curves
PartnerRe has an experienced research team that has been developing catastrophe models in-house for over 15 years, and which has long worked and continues to work in close cooperation with the scientific community. In 2006, our team developed the CatFocus® European Windstorm Model, a numerical weather prediction (NWP) based, dynamical model for European windstorm. This model delivers a comprehensive3, high resolution, historical storm set (example in figure 1) and has been steadily enhanced and updated.
Figure 1: Maximum wind gust for storm Lothar (1999) from the PartnerRe historical storm set.
In a recent exercise, the first of its kind, we were able to utilize the data from twenty-two studies involving 13 different dynamical European regional climate models (RCM). Each model delivers an ‘alternative’ view of history by providing an individual, high resolution, 40-50 year storm set. These regional models are the most advanced from the academic community and yet each varies in the way that it represents the physical processes of windstorm events – facts that make these models an excellent quantitative indicator of return period uncertainty (figure 2). Before the data could be used for risk assessment, the climate models were each calibrated and validated to PartnerRe’s historical storm set (to read more on this topic, see Haylock (2011)4).
Figure 2: Europe-wide5 loss curves for the regional climate model (RCM) storm sets after calibration to the PartnerRe historical event set. PartnerRe’s CatFocus® historical modeled losses are given by the blue points. This plot shows that although there is considerable variation in the individual RCM loss curves, there is a good fit between the mean of the RCM curves (the dotted line) and the historical event set (solid black line). Source: Haylock (2011)4, adapted.
Internal consistency – the sign of a good model
For a catastrophe model to be reliable it should, over a large domain and given a spatially uniform portfolio, display broad agreement (fall within the statistical limits of uncertainty) between its historical and stochastic (generated) event sets. This is known as ‘internal consistency’ and signifies a sufficient volume of quality data, good understanding of the complex physical processes at play and the identification and correction of modeling biases.
One of the main areas of modeling uncertainty that remains today, even within the most advanced catastrophe models, is assigning frequency (or return periods) to modeled losses. Different methods are deployed for this across models and all involve considerable subjectivity, making the need for internal consistency tests on a model’s regional loss curves (loss against return period) all the more crucial. If a model generates a loss curve that lies outside the uncertainty range of the historical storms on which it was based, the model is not internally consistent.
A good model will be able to display internally consistent loss curves. Such a test can be performed by comparing the return period of loss of the stochastic events with a statistical measure of uncertainty of the historical losses. In this study this was done using a Generalized Pareto distribution (GPD) fitted to the historical events (figure 3).
Figure 3: GPD loss curve (black line) from the CatFocus® European windstorm model based on the model’s high resolution, dynamically modeled, historical event set (blue points). The RCM-derived stochastic storms (red points) fall well within the 95% confidence interval of the historical data, indicating internal consistency.
The results of the test were positive, that is, the Europe-wide loss curve from the calibrated regional climate model storm set, is well within statistical uncertainty of the GPD fitted to the historical event set of the CatFocus® European windstorm model, thus validating our own model’s methodologies and loss curves.
Application to small regional portfolios
The extra storm sets from the regional climate models can be used to improve the risk analysis of small regional portfolios; such portfolios usually lack historical events on which to base analysis, the locations having experienced very few (possibly zero) storms over the few decades since quantitative observation began. Although the internal consistency checks described in the previous section cannot be applied for such portfolios due to the lack of historical event data, the enhanced risk analysis and positive ‘internal consistency’ results from other tested regions indicates good overall model reliability.
1 natural and man-made
3 132 historical storm events
4 European extra-tropical storm damage risk from a multi-model ensemble of dynamically-downscaled global climate models M.R.Haylock; Nat. Hazards Earth Syst.Sci., 11, 2847-2857, 2011
5 For a population based portfolio
Dr. Tom Fletcher, VP Data Analytics, North America Life, shares his perspective on the component strands necessary to build and maintain ethical standards in predictive […]