There is significant potential for machine learning technologies to enhance the re/insurance industry’s approach to modelling wildfire risk, which has failed to adequately predict the growing frequency of large losses in recent years, according to Swiss Re.
The reinsurer stressed the possible advantages in utilising real-time data to complement probabilistic wildfire models as part of a report on building resilience to wildfire in Canada.
Wildfire losses in Canada have surged since the early 2000s, Swiss Re noted, with re/insurance industry costs totalling almost CAD $5 billion (USD $3.7 billion) from 2003 to 2017.
The 2016 Fort McMurray fire in Alberta cost re/insurers $3.8 billion alone, making it the costliest natural disaster in Canada’s history.
Swiss Re explained that the rise in wildfire losses was partly due to an increase in average temperatures due to climate change, and partly due to increasing human activity in wildland-urban interface areas in Canada that are prone to wildfires.
While a significant amount of research has been conducted in the operational wildfire management sphere, wildfire risk modelling in insurance has lagged behind risk modelling in other insurance-relevant spheres related to natural catastrophes, such as earthquakes, storms and floods.
While risk modellers such as AIR, CoreLogic, Eqecat and RMS have started to develop full-scale probabilistic models for regions with high exposure to wildfire, the complex interplay of factors that affect wildfire activity has meant that providing accurate predictions remains challenging.
Swiss Re has responded to this challenge by collaborating with MIT to enhance wildfire models by using machine learning to analyse years of satellite images and plot the spatial and temporal relationships of the atmosphere and biosphere as well as human interactions within the natural system.
Machine learning-based models have a major advantage over probabilistic simulation models, Swiss Re claims, because they can detect inter-relationships between underlying features that cannot be parametrized, such as vegetation zones and lighting, several months in advance.
“Wildfire risk estimations may become more robust if existing probabilistic simulation models with emphasis on vulnerabilities and exposures are combined with machine learning-based hazard models,” Swiss Re stated in the report. “This could significantly improve the ability of insurance companies to assess the wildfire risk their customers face.”
The accurate pricing of wildfire risk by insurers will play a key role in incentivising appropriate risk prevention and reduction by individual policyholders as well as entire communities, Swiss Re concluded.
In this way, re/insurers can be valuable partners to governments in fostering wildfire resilience, the reinsurer said, adding that there is an alignment of interest with those in urban planning, public safety standard setting, and loss prevention planning at the municipal, provincial and federal levels.