Towards the next generation of Cat Models – Part 2
3sec gust wind footprint for Typhoon Jebi (2018) using REASK wind field and terrain correction Machine Learning models.

Towards the next generation of Cat Models – Part 2

In a series of blog articles this month we put forward what we believe are necessary steps to build the next generation of Cat models. In part 1 we detailed the need to account for climate variability and embrace a modelling framework that is truly global. Here we discuss two other aspects we believe are fundamental to moving the industry forward: 

  • While observed datasets are the gold standard to train models, they have significant gaps in coverage. High resolution full-physics models offer an alternative with consistent and gap-free global coverage. 
  • There is nothing special about history and excessive calibration towards observed records leads to models that overfit the past. 

The value of high-resolution full-physics numerical weather modelling
Training machine learning (ML) algorithms requires extensive data that captures the range of environmental dynamics to be modelledObservational datasets are a fantastic training source where and when they are available, but they often fail to cover the full range of risks to be modelled. Climate and weather model simulations can act as a good alternative: at REASK we make extensive use of the publicly available global reanalysis datasets (see part 1) but also run deterministic physical models ourselves to train model components that require higher resolution. 

For example, we have built our own database of high-resolution simulations to capture variability in tropical cyclone (TC) wind structure globally by running WRF (full-physics weather model) at 1 km on High-Performance Computing facilities (see InCyc for more details)The InCyc database includes over 20simulations spread across all basins and carefully selected to be representative of the population of historical TCs (Figure 1).

This proprietary database (over 50 Tb of data) is used to build a range of ML tools, that drive different components of our global TC event set and Cat response models. These tools simulate wind field shape and asymmetries [Loridan et al., Monthly Weather Review, 2017along with terrain topography and roughness effects to estimate gust winds (see footprint from Typhoon Jebi above). 

Figure 1: InCyc - REASK database of high resolution full-physics historical storm simulations with global coverage.

There is nothing special about history
We share the view of a growing number in the insurance industry that historical data should not be considered as anything more than a sample from a broad distribution of riskHistory should not be the yard stick by which we measure the accuracy of a Cat model but exists as a single representation of what could have happenedFigure 2 illustrates the significant variability of our tropical cyclone stochastic track set for the post-satellite data (compared to IBTrACS) and highlights how 40 years of history cannot fully characterize the range of climate and weather variability that could have occurred 

Figure 2: Ensemble of 50 year stochastic tracks in Northern Australia

We believe that heavily tuning Cat models to the past (as is widely done to validate return periods of TCs at a set of defined gates for example) is an exercise the industry now needs to leave behind. The location of the historical sample within the modelled risk distribution is a healthy aspect to check, but faithfully following that history should not be the target of Cat models.  

With increased pressure on Cat model users to understand the impact of a warming climate on their risk it is also critical to base our next generation of tools on a transparent and physically driven view – one that can easily be adapted to reflect future risk. Any comparison of current and future risk levels is only valid if both were derived following the exact same process. Over-tuning models to match observed records is not a practice we can apply to future risk.  

Our models are based on an understanding of [and directly driven by] global climate processes without manual tuning and adjustments. To us, this is fundamental to efficiently and seamlessly build global Cat models that provide a consistent view of risk. Carefully fitting models to history is a practice that is only possible where good historical records are available and therefore does not scale well into emerging markets; it also limits model applications in the context of climate change.

If you want to learn more about our Cat modelsplease get in touch. 

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Close Menu