Machine learning (ML) methods are famously disrupting every field they get applied to, and many believe we are just at the beginning of a major revolution that will spare no industry. At the heart of the concept is the assumption that, if presented with enough relevant examples, ML algorithms can learn to predict a given outcome without any need for a modeller to explicitly specify any rule.
Catastrophe modelling, which is centred on the study of rare events, presents a couple of challenges for such methods:
- There exists very few well documented historical examples for algorithms to learn from
- Models need to behave outside of the scope covered during the learning phase
In the past year Risk Frontiers’ researchers have collaborated with experts in the ML field to identify ways to best apply these methods to catastrophe modelling. This collaboration led to the development of a new generation of tropical cyclone wind models (see below) and the formulation of a new framework to account for peril correlation.
Example study: Tropical Cyclone surface wind field modelling
- Loridan, T. Crompton, R., and E. Dubossarsky: “A machine learning approach to modelling tropical cyclone wind field uncertainty”, Monthly Weather Review, May 2017, in press.
- Talk at data Science Sydney meetup, 27 September 2016: https://www.youtube.com/watch?v=ZB6Od_89TWA
Tropical Cyclone (TC) risk assessment models rely on very large ensembles to simulate the track trajectories, intensities and spatial distributions of damaging winds from severe events. Given the constraints in computing time associated with the generation of such ensembles only very simple formulations have been used to date. The talk first introduces the concepts behind catastrophe modelling using TC risk models as an example and provides an overview of the modelling techniques common to the field. In a second part the potential for machine learning to improve common practice is explored and some key challenges are discussed.