Modeling Weather or Not

As the severity of natural catastrophes continues to increase and insurers see losses in the billions of dollars, the conversation has turned to risk management and what can be done to mitigate some of this risk.  

While weather modeling has long been in use in the insurance industry, some are now looking to incorporate more longer-term climate-based predictions into their analysis.

What’s more, many believe that artificial intelligence will provide a key element to making climate models more accurately depict real-world risks of weather.

Verisk, a data analytics and risk assessment firm, was involved in a project in Miami Beach to assess future storm surge and the risk of coastal flooding. “We used our catastrophe models to look at the future of sea-level rise, how storm surge would be exacerbated, and how frequently the city would flood,” recalls Peter Sousounis, the firm’s vice president and director of climate change research. “This informed a cost/benefit analysis to understand how much higher sea walls should be built.”

A 2020 study in Nature, the British science journal, called understanding and estimating regional climate change under different anthropogenic emission scenarios “pivotal” to adaptation and mitigation. “However, the high computational complexity of state-of-the-art climate models remains a bottleneck in this endeavor,” the study found.

“Modeling a wide range of processes on a wide range of time scales makes climate models complex, requiring large computational resources and a lot of time,” says lead author of the study Laura Mansfield, a postdoctoral researcher at Stanford University’s Doerr School of Sustainability. “Artificial intelligence [AI] or machine learning [ML] models can be trained from climate simulation data sets. Once built, they can make predictions much more rapidly compared to full climate models to quickly estimate cost and risk associated with a given climate event,” she adds.

General Circulation Climate Models

According to the National Oceanic and Atmospheric Administration, climate models use physical processes to simulate energy and material transfer through the climate system. These general circulation models, or GCMs, the agency says, “use mathematical equations to characterize how energy and matter interact in different parts of the ocean, atmosphere and land.”

“We use dynamical and statistical models to understand how climate will respond to a warming impact,” says Cindy Bruyère, deputy director of the University Corporation for Atmospheric Research’s Scientific Partnership and Service Center in Boulder, Colorado. “For instance, hurricanes are strongly impacted by sea surface temperature. So understanding that link gives you a statistical view into how they might change in the future.

“On the dynamic side are the global models. These are big numerical models run under IPCC [Intergovernmental Panel on Climate Change] climate scenarios. These dynamical models give a full picture of the three-dimensional atmosphere to help us figure out how climate and extreme events will change in the future.”

Bruyère cautions that “expert eyes” are still needed to assess these models. “Because what looks like a hurricane in the real world,” she says, “doesn’t look the same in a climate model.”

Kenneth Kunkel, a research professor of atmospheric sciences at North Carolina State University, agrees that expert advice helps to interpret data in models. “If you take projections from fine-scale models, there’s a lot of variability in space,” Kunkel says. “You have to know how to interpret that not just from one point but every point around it.”

Princeton University’s Geophysical Fluid Dynamics Laboratory describes GCMs, currently the primary tool in predictions, as “a complex mathematical representation of major climate system components (atmosphere, land surface, ocean and sea ice) and their interactions. Earth’s energy balance between the four components is key to long-term climate prediction.”

Kunkel, who also serves as the scientist for climate assessments at the North Carolina Institute for Climate Studies, says spatial resolution of models doesn’t allow for physics at small scales to be simulated.

“GCMs are based on the fundamental laws of physics,” Kunkel says. “That said, they’re imperfect. Tropical cyclones are a good example. Even though the models have physics, it’s limited in the sense that we’re just not yet resolving small-scale physical processes important to the dynamics of a tropical cyclone due to limited computational resources. Thus, small-scale physical processes, such as thunderstorm cells, cannot be simulated directly.”

GCMs and catastrophe models can simulate the climate out to 2050 and beyond, says Sousounis, who warns of vendors who overpromise and underdeliver. He says it’s important to know both the biases and capabilities of high-resolution models. “There are people who generate future climate information and advertise high accuracy and precision in terms of resolution,” Sousounis says. “Predicting that floods are going to increase by x percent in a particular portion of the country by 2050 oversells the science by far.”

“If people are going to make those claims, they have to attach uncertainty bars like, ‘We expect this to happen. We have this kind of confidence.’ That’s very important for the insurance industry.”

AI, ML Mimic Real-World Perils

According to Argonne National Laboratory, in Illinois, climate models, no matter how “computationally intensive,” contain uncertainty. To improve climate simulations, scientists look to AI to dramatically improve understanding and representation of the real world. The lab’s website claims AI offers the potential to dramatically increase the accuracy of predictions down to the scales of interest to scientists and stakeholders.

Bruyère notes that AI has been used to predict the number of tropical storms. “These models work really well in the current climate,” she says. “The uncertainty for future storms is still very high because we do not yet fully understand the impact climate change could have on extreme weather events. That said, AI gives us an advantage, as it is built on science and not on backward-looking observations.

“I’ve seen them [AI models] used to explain visually what would happen to properties if sea levels rose or surges occurred. That visual impact is used a lot in social science to communicate on a level that’s real.”

The bigger uncertainties, she says, are about GHG concentrations and shared socioeconomic pathways (projected global changes up to 2100). In other words, she says, the amount of GHG that humans will add to the system is one of the biggest uncertainties.

According to a Cornell University study, observational data delivered through ML can correct statistical biases in high-resolution models. A 2022 study published in Neural Information Processing Systems’ “Proceedings” calls statistics for extreme events “difficult to predict,” especially on a regional scale.

“Extreme weather events have a very low probability of occurrence,” says lead author Antoine Blanchard, senior scientist at Verisk. “To estimate the likelihood of very-low-probability events accurately requires a tremendous volume of data. For example, to estimate the probability of another Hurricane Harvey, we would need data containing multiple Harvey-like storms, but this data doesn’t exist. This question becomes even harder at the regional scale because the physical processes at play are difficult to simulate from historical data.

“The next frontier in cat modeling is to provide views of risk that incorporate events unlike those we’ve seen before and in which all perils are simulated coherently and consistently across the globe. This cannot be achieved with GCMs alone because of the biases. But augment a GCM with observational data and ML, and immediately a new paradigm emerges to usher in a new era for cat modeling.”

Climate Risk Key to Strategic Planning

According to the 2022 position paper in Nature Climate Change, organizations should include climate-related risks in strategic plans. The paper calls forward-looking information about climate risk “critical” for decision makers and stakeholders.

“The current market for climate risk predictions suffers from a fundamental problem,” says Mark Roulston, a consultant at the Smith Institute and the paper’s lead author. “Companies are providing predictions of climate risks years or decades ahead but expect to be paid up front for these forecasts. In contrast, providers of short-range weather forecasts can be evaluated on their track records. If the customer is unhappy with the accuracy, they can change providers. It will be many years before the accuracy of a climate risk forecast is known. You don’t need to be an economist to appreciate the problem.”

The 2022 position paper describes climate change as “a multi-disciplinary problem,” noting that forecasting climate, including GHG concentrations, requires expertise in physical science, economics, policy and innovation to avoid an information gap. IPCC predictions rely on socioeconomic scenarios, states the article, which endorses prediction markets, such as those used in betting, to synthesize information into probability forecasts.

GCMs bridge the gap between underwriting and risk. “Global models are one of our tools to show what’s going to happen and how our weather systems are going to react to changes in our climate,” Bruyère says. “They’re not perfect, but they give a view of what our future will look like because of human-induced changes.”

Adds Sousounis: “Using that information to look at potential futures to create a catastrophe model gets back to whether the client is interested in the view for 2050 to give insurance industry, local governments, urban planners the confidence to design infrastructure that’s more resilient and less susceptible to damage,” he says.