Over a pair of blog posts we will be discussing the key concepts behind REASK’s model building philosophy, as well as why we believe they represent a natural step towards a new generation of catastrophe models.
Since the late 90s, the insurance industry has relied on Cat models to provide probabilistic views of the potential risk from natural disasters. While these models have proven very useful, they are still mainly based on a long-term statistical view of risk and poorly account for the impact of climate variability. They also typically lack scalability and therefore tend to adopt a region-by-region / peril-by-peril approach to risk modelling.
We believe that the recent leap in climate data availability should inspire a new generation of Cat models: one that brings together the highest resolution global datasets with state-of-the-art weather models and machine learning (ML) to power a truly global, climate-connected view of risk.
REASK’s climate connected view
From basin-wide and regional activity to track trajectories and intensification rates, all aspects of tropical cyclone (TC) risk modelling are directly dependent on the state of our climate system. Our approach is to use ensembles of global climate simulations to drive all key components of our Global TC model, and therefore reflect the role of climate variability in TC risk.
TC activity driven by global connectors
Understanding and capturing teleconnections in our climate system is fundamental in accurately representing TC activity globally. For example, it is well established that TC activity in various parts of the world is linked to the state of the El-Niño / La-Niña phases (ENSO cycle). One can for instance use the Southern Oscillation Index or Niño 3.4 to quantify the magnitude of these phases. However, while widely used, these indices are traditionally simplistic and manually engineered to be consistent with some human predefined framework.
To extract physical patterns, we have designed an automated Climate Knowledge Discovery Engine (see Clear Path Analysis) which efficiently mines climate, and more broadly environmental data. While humans are only able to study a few parameters or time scales, machines don’t have such limitations. They can mine the large volume of available data and extract what we call climate connectors. Our Discovery Engine uses state-of-the-art techniques from the field of image and pattern recognition to identify key drivers of risk. Figure 1 shows an example of a mixture of Atlantic Meridional Oscillation (AMO) and ENSO related pattern as “discovered” by a neural network.
These climate connectors can then be plugged into Cat models to link risk between regions and/or perils. To illustrate this, Figure 2 shows the distribution of tropical storms between the North Atlantic and the North West Pacific basins, with comparison to a traditional static view of the risk (climatology).
TC event characteristics driven by key climate variables
Beyond activity metrics we also leverage climate simulations to condition TC trajectories and intensification rates. Simulated steering flow, vertical wind shear and sea surface temperatures are all used as inputs to our global machine learning track generator. This allows regional specificities in typical synoptic set up to be accounted for (e.g. Bermuda high, Caribbean shear or ENSO-driven anomalies in easterly flow in the Pacific).
Climate variability
Thanks to this climate-connected framework our models can investigate the impact of climate variability on risk. As an example, Figure 3 illustrates the return period and associated tail risk of tropical cyclones around the Pacific island of Vanuatu depending on the El-Niño / La-Niña phases.
Climate change scenarios
With the ability to connect our view of risk to climate features and variables on the global scale, a natural evolution is to use climate projections under different scenarios as alternative forcings. This will for instance allow comparison of risk distributions from climates representative of the 2020-2050 period (or beyond) to current day baselines in a similar manner to that presented in Figure 3.