Tornados, Statistics and Better Modeling

On the heels of an uptick in tornado activity, we talk with FSU's James Elsner.

Chris Westfall
Chris Westfall

Thanks for being a subscriber to the free version of Risk Market News. Consider becoming a paid subscriber for access to all updates.


Catastrophe Q&A:

Florida State University's James Elsner

Tornadoes that tore apart areas of the Midwest and Southern US over the weekend reveal the limitations of forecasting and modeling severe convective storms.

Dr. Dr. James Elsner, the Earl and Sophia Shaw Professor and Chair of Geography at Florida State University, says that insurers and modeling firms should rethink their approach to the tornado risk and shift more resources to statistical analysis.


Risk Market News: What aspects of last week's tornado do you feel made it a significant event in terms of either expected losses or casualties?

Dr. James Elsner: The unusually warm air out ahead of the parent storm system is probably the most impressive.

That directly contributes to the amount of convective available potential energy (CAPE), which really is the fuel for these super cell thunderstorms use to produce tornadoes. I also think the unusually warm air influences, indirectly, the strength of the jet stream, which leads to the rotation. The aspect that's the most impressive was the amount of warm air at this time of the year.

RMN: Why is the atmosphere producing more of this warm air fuel?

Dr. Elsner: There's a lot of weather going on in the background. But we're all warming the planet, so when we get these setups of increasingly warm air coming in from the Gulf of Mexico. I think it's a fairly simple connection to make between the extra heat and from climate change to fueling the super cells.

RMN: You have done some research that focusing on the tie between population density, tornado strength and losses. Could you describe what you were talking about in that research and how it ties into what we saw last week?

Dr. Elsner: You have to think of it statistically. The research is looking at things across many events and trying to understand what are the risks.

If you look at the number of casualties from tornadoes it's going to depend on how strong the tornado, or the group of tornadoes, are and the number of people it affects. Population and power, the two P's, are the two components of losses when it comes to human lives. Trying to understand how the two factors of people and power statistically interact.

Statistically, we are seeing a shift  toward the Southeast. Historically, places like Oklahoma and Texas got the most tornadoes and there were relatively few tornadoes in Tennessee and Mississippi. But that has changed a bit. We still see more tornadoes on average over Oklahoma, but we've seen relatively more tornadoes in places like Tennessee. And when you compare the population of Oklahoma versus the population of Tennessee, there's close to three times as many folks that live in Tennessee than in Oklahoma. You need to change the P in that equation.

RMN: What do these changes mean from a modeling perspective?

Dr. Elsner: It's not clear if there's a connection between the shift toward the Southeast and climate change, but it does represent a change in demographics as tornadoes occur farther East tend to be more of a nocturnal event. And nighttime tornadoes tend to be more dangerous. Modeling companies need to grapple with and think about those issues, including the nocturnal nature of the storms further East and the increasing population density relative to the traditional Great Plains events.

RMN: Are there any recent tornado or severe connective storm modeling advancements that you think are especially promising?

Dr. Elsner: I think there's two different themes here.

One is the dynamical models that are used for weather prediction. And I continue to get better. The resolution continues to get better. They really still can't resolve a tornado, it's just too small of scale. But they can start to resolve the parent mesocyclones that are responsible for some of the strongest tornado. I'd expect to continue improvements in these kind of dynamical models, although there's probably some limitations in the convective feedback and leading to numerical instability in those models. It starts to really become a hurdle that's difficult to overcome.

For risk modeling, you can start to think about how you might be able to use some of these high resolution, super cell resolving models to simulate hypothetical storms using some kind of climate model as an initial condition.

Also, people are becoming more familiar with more nuanced statistical models. Statisticians are very good at working with unreliable data. They do it all the time. Meteorologists need to learn how to work with things like Bayesian models where you can set priors and you can do conditioning and random effects. These are the kind of tools that are going to be more useful going forward to understand how things are changed or have changed and simulate future scenarios that are worthwhile for the industry.

RMN: Is a greater focus on statistical modeling simply something that's part of the education process for industry professionals?

Dr. Elsner: I think it's the education process.  I came from a classic meteorology background and we don't learn statistics in any meaningful way. We learn how to maybe analyze data, do some statistical analysis with extremely low frequency (ELF) data but we don't really spend any effort on influential statistics or even classical statistics.

You get a lot of folks who are very good at understanding the phenomena of hazards, whether it's tornadoes or hurricanes or wildfires, but they're not really good at looking at historical data in any meaningful way. And I think the social scientists are much better at this and that's the only mathematical tools that's available.

You can't really do the black box. There's folks that talk about artificial intelligence and yes, you can use greedy algorithms and the black box, but that's not really understanding causation. It's not thinking through the problem in a meaningful causal way. And I think we'll see more of that in the future.


Reinsuranceinsurancecatastrophe modelingnewsletter