Modelling wildfires

March 8 2019 by Nick Ferguson

As bushfire season ravages Australia once again, it is clear that climate change is increasing the hazard and making it more difficult for insurers to price the risk. New solutions are needed — and machine-learning technology may be able to provide them.

While it is impossible to connect any single event to climate change, Australia’s summers are unequivocally getting longer and drier, extending the bushfire season and pushing it into new areas. Climate Council, an Australian non-profit, found in a 2016 study that the economic cost of bushfires to New South Wales and the Australian Capital Territory will likely double by 2050.

And areas normally unaffected by fire are starting to burn. In November, the north-eastern state of Queensland declared a catastrophic fire warning for the first time in its history — at a time of year that is normally considered the wet season.

“We have not experienced these conditions before,” said the local fire commissioner at the time. “It is unprecedented.”

Machine-learning techniques that make use of big data and artificial intelligence may offer the ability to do a better job of predicting wildfire hazard, according to Swiss Re, and therefore help insurers to gain a better view of their exposures, which in turn is beneficial for pricing, reserving and portfolio management.

“Furthermore, the accurate pricing of wildfire risk by insurers plays a key role in incentivising appropriate risk prevention and reduction by individual policyholders as well as entire communities,” said Monica Ningen, chief executive of Swiss Re in Canada, where wildfire losses have risen sharply during the past few decades. “Insurers can be valuable partners to different stakeholders, from policyholders to governments at all levels, to foster wildfire resilience.”

Risk solutions providers including AIR, CoreLogic, Eqecat and RMS are already developing full-scale probabilistic wildfire models in high-risk places such as California and Australia, but fires are complex and modelling them is extremely challenging. There is limited historical data and a staggering quantity of inputs needed to create a model, while the environment itself is constantly changing. Predicting how an individual event will play out is almost impossible.

However, it is possible to predict the conditions that will cause wildfires to ignite and become bigger. This knowledge can then be complemented with real-time data to understand wildfire risk better.

Swiss Re has teamed up with MIT to use “deep learning” to predict annualised wildfire loss, analysing several years and months of satellite images to learn spatial and temporal relationships of the atmosphere and biosphere as well as human interactions within the natural system.

Deep learning is part of a broader family of machine-learning methods based on learning data representations, as opposed to task-specific algorithms.

“Machine learning-based models have a major advantage over probabilistic simulation models because they can detect inter-relationships between underlying features that cannot be parametrised (eg, vegetation zones and lightning or surface temperature and prior burned area several months in advance),” said Swiss Re in a recent report.

Such predictions could be extremely valuable to insurers and wider communities as they try to address the risks they face from wildfire because one thing is for sure: they aren’t going away.

Climate change is creating bigger liabilities all over the world, as highlighted in a recent report by Clyde & Co. The number of registered weather-related loss events has tripled since the 1980s and annual average inflation-adjusted insurance losses have risen to US$55 billion from US$10 billion.

Failure to understand the risks of climate change is partly to blame, with people and businesses continuing to move into areas that are exposed to risks — whether from wildfires or other events that are likely to be made worse by climate change, such as wind storms or floods.

It is to be hoped that better models and a better understanding of the risks will put a more accurate price on such behaviours and result in more informed decisions.

MORE FROM: Insights