Why natural catastrophe models are still missing climate change

Climate change keeps surprising the risk modellers, and it’s not a new problem.

“You take an event like Houston, like Hurricane Harvey—was there a catastrophe model that captured an event like that sufficiently? My understanding is no, there wasn’t.”

These were the words of Daniel Stander, special advisor at the United Nations Development Programme (UNDP), talking about the failure of current natural catastrophe models to capture the true impact of climate change as it intensifies and increases the frequency of severe global weather patterns.

“There wasn’t anything that looked like Harvey—not just a little, but closely represented the kind of storm that lingered over Houston for weeks,” he added.

Stander was speaking as part of an Intelligent Insurer panel titled “Climate change: adapting models to better understand the systemic threat”.

Other members of the panel were Bhaskar Chattaraj, partner, head of modelling and R&D at TigerRisk; Sébastien Piguet, co-founder and head of underwriting at Descartes Underwriting; and Meghan Purdy, senior product manager at Jupiter Intelligence; with Claire Churchard, deputy editor of Intelligent Insurer, as moderator.

Purdy agreed that current models are missing a trick when it comes to defining the true climate change signal. Her firm produces metrics for the recent past (using the Intergovernmental Panel on Climate Change historical baseline period), present day and looking forward. She said that when she speaks to insurers a lot of what they want to look at is just comparing the baseline to the present day to see how much the climate might have already changed. They then want to look five, 10, 50 years out for what’s going to happen in the future, she said, but this approach needs to change.

“One example is Houston, which experienced three so-called 500-year events between 2015 and 2017. So either they were profoundly unlucky or the definition of a 500-year event needs to change. The old definitions we are using are just not correct any more,” Purdy said.

The new models

Creating accurate modelling that incorporates the mercurial changes of global warming is the Holy Grail for the insurance industry. But if current models are “severely undercooked”, as famously stated by Richard Trubshaw, founding partner of MAP, is it time to dump existing models and start from scratch?

Chattaraj, who has been building and using models for a long time, starting his career at AIR, is critical of how existing models have been adapted so far.

He said that cat models, which handle a lot of the climate change data, are based on historical data and weather parameters that have been observed.

“How are they being adapted, that’s where my big complaint is. The modellers are not building something different on climate change. They’re looking at history and perturbing it.”

In practice this means adding more extreme storms to the model, he said, and putting more storms where there were no storms before, because that’s roughly what the climate scientists are saying.

“It’s not that trivial—you can’t just perturb a model and say ‘it’s a model of the new world’,” Chattaraj explained.

“When you model a new world you have to take new assumptions and make a forward-looking model. They are taking a backward-looking model, just turning a few knobs and saying ‘that will capture climate change for catastrophes’.”

He added that apart from extreme temperature and precipitation, there’s not much confidence in modelling other perils, so people are not sure how they will be affected. “There’s a lot of uncertainty and we have to work towards understanding and reducing it.”

Types of models

Piguet said that his firm is a young company, which launched products about two-and-a-half years ago.

“We had to start with all-new models. I’m not saying that we adapted old models from the industry—we had to start from scratch as a company. We are developing a model which by design allows us to take into account the impact of climate change,” he explained.

Piguet said that the firm is leveraging a lot of scientific literature, and still making use of current models which have been designed by scientists all over the globe.

“In many cases we are using famous models in the scientific community, which are not used much in the insurance industry,” he said.

“But it’s important to distinguish between current models used by scientists and current models used by the industry.”

Purdy agreed, saying that there’s a spectrum between climate models on one side and cat models on the other. “My company is trying to do something in the middle, but I agree that cat models tend to be more historically focused, they’re not thinking about what the storms of the future will look like because they’re looking backwards to understand what they might be.”

How can what you get from a cat model versus what you can get from a climate model be mixed?

Purdy said that modellers can’t take a pure climate model approach because there are not enough simulations of future climate to blanket the world’s coastlines with all the possible storms that could happen.

“But we can use climate models for things such as sea surface temperatures, wind shear conditions and other factors that are indirectly affecting the strength and the storm tracks that we see in the future.

“We have the climate models telling us that, and then the types of events that we layer on top of the climate models are stochastic, and then we can increase the wind intensity and the precipitation intensity based on what the background conditions are telling us.

“That’s helping us mix those two approaches and speak the language of insurance and the cat models while also bringing a climate dimension into it,” she explained.

So, when it comes to new models, throwing the baby out with the bathwater is not the answer.

“There’s so much good knowledge and understanding that has been built on since the 1980s when cat modelling began,” said Stander.

“I don’t think we can chuck all that knowledge and understanding away. But what I heard from Bhaskar and Meghan is that maybe we need to fundamentally change our approach and not start with backward-looking models and tweak them, but start by building forward-looking models.”

This is a complex process and Chattaraj suggested that this is where artificial intelligence (AI) and machine learning could be helpful.

“Scientists and climate modellers have produced only physical models in the past,” he said. “The time has come to combine the physical, AI and machine learning models because there’s not enough data and the physical models are not doing well. So we need to go the other route.”