By Peter Sousounis Ph.D
Extreme weather events can disrupt local, national, and global economies for months, if not years, to say nothing of their potential to take human life. Based on more than 20 years’ worth of historical loss data, Verisk calculates the global average annual economic loss from extreme weather events be more than $320 billion, as reported in the 2021 edition of our “Global Modeled Catastrophe Losses” report. The ability to anticipate and prepare for extreme weather, not just in terms of a single event from a single peril but also in terms of a series of such events over an entire year, can have widespread benefits to society. And yet, we’re often caught off guard, or at least under-prepared, for many extreme weather events, leading the public to believe they were completely unpredictable—so-called black swans.
The reality is quite different. Extreme weather events may be rare, but we do have the ability to model and forecast them. We know enough to be better prepared. Which begs the question, what type of weather or climate-related event should we consider a “black swan”, and can our predictive capabilities keep pace with a warming planet and the potential convergence of multiple low-probability catastrophes in a single year?
Defining Black Swans
The term “black swan” was brought into popularity this century, even though it was used for many previous centuries to refer to something that was not thought to exist – a black swan. The discovery of an actual black swan in the 17th century on the west coast of Australia by Dutch explorers contradicted a long-held belief in Europe that there was no such thing. It’s hard to say what the consequences of that discovery were, exactly, but the magnitude of the discovery’s impact can be surmised by the fact that a black swan has become synonymous with an event that is entirely unforeseen and rare.
In 2007, Nassim Nicholas Taleb formally identified the three criteria an event must meet to be considered a black swan. In short, a black swan event is 1) extremely rare, 2) has severe consequences, and 3) the event should have been predictable in hindsight (otherwise known as hindcasting). There’s no precise quantification for just how rare the event is, except that the event may be so far out in the tail of a probability distribution that it was unanticipated.
Many weather events that have been labeled black swans don’t actually qualify from a hazard standpoint, since they’re not all that rare. Events typically earn the label black swan when they result in significant and unanticipated disruption almost independent of their actual intensity or their probability. A good example of a true weather-related black swan is Hurricane Katrina in 2005. This hurricane made landfall as a Category 3 storm in Louisiana, with impacts being felt as far away as Mississippi. That area of the U.S. had seen Category 3 hurricanes in the past, and from a hazard standpoint, the hurricane had an exceedance probability of 5%--In other words, there’s a 5% chance of a hurricane of this magnitude happening in any given year (Elsner et al. 2006)—not really an extremely rare event.
Events typically earn the label black swan when they result in significant and unanticipated disruption almost independent of their actual intensity or their probability.
From an insured loss standpoint, however, it was extremely rare. Losses of that magnitude from a storm have a 0.2% chance of happening in any given year. More significantly, it’s estimated that approximately 1,500 people lost their lives because of the storm, far in excess of the average loss of life from hurricanes in this region. The differential between storm intensity and storm impact demonstrates that the New Orleans area was not prepared for this type of storm.
More recent black swan weather events include the winter storm “Uri” in February of 2021 that crippled the Texas power grid for weeks, and led to an estimated economic impact of between $195-295 billion (PerrymanGroup 2021). Although the event is considered by many to be a black swan, lessons from prior cold-snap events in Texas history appear to have gone unheeded. The power grid could have been better prepared, so this event, too, is more aptly labeled a “grey swan” (Doss-Gollin et al., 2021).
An even better example of an event worthy of the black swan label for Texas is Hurricane Harvey in 2017. Harvey’s rainfall (as much as 60 inches in some locations) was noted to have a return period of 2,000 years, or a 0.05% chance of happening in any given year, based on the climate of the late 20th century (Emanuel, 2017). The storm’s economic loss was estimated at $125 billion by the National Oceanic and Atmospheric Administration (NOAA).
How Numerical Weather Prediction Has Lowered the Likelihood of a Black Swan
The atmosphere, which is a fluid, and the weather it creates, can be mathematically predicted. The Navier-Stokes equations to describe the time evolution of fluid flow were developed in the mid-1800s jointly by French engineer and physicist Claude-Louis Navier and Anglo-Irish physicist and mathematician George Gabriel Stokes. The equations are time-consuming to solve, especially over the entire surface of the Earth, and with enough speed for the information to provide a useful forecast.
In the early 1900s, Louis Fry Richardson (Richardson 1922) had the idea of 64,000 people sitting together under one roof with each one performing calculations for their grid point based on weather information that could be gathered from weather balloons, and then passing that information to the adjacent person to calculate in the appropriate direction while getting new information to make their next calculation from someone else. But this human-forecasting assemblyline proved impracticable. It wasn’t until the invention of ENIAC, one of the first analog computers some decades later, that the first numerical weather forecast was generated (Charney, Fjörtoft, and von Neumann, 1950). The accuracy and speed of numerical weather prediction have come a long way since, and in the process have allowed us to better prepare for many extreme weather events.
The ability to simulate an extreme event at all that may occur at some point in the future is important from a preparedness standpoint. It gives us a hint as to the type of black swan events that may loom in our future.
Despite tremendous improvements, numerical weather forecasting is constrained by the chaotic nature of the atmosphere and different circulation patterns from small disturbances somewhere on the globe. In the 1960s, Professor Edward Lorenz discovered that no matter how well we measure the initial conditions, because the atmosphere is chaotic, small errors in those initial conditions will amplify and analysts are bound to get an incorrect forecast to some degree after 11 days (Lorenz 1969). Sometimes the difference can be dramatic, owing to the atmosphere bifurcating to a completely different state. But the numerical forecast will not generate something unrealistic; it won’t generate a blizzard in Miami in July. This constraint, plus the physical foundation of the equations of motion, extend the reliability of long-range forecasts and climate predictions. Climate models will almost never generate something that is not physically possible because the numerical instabilities that could result (e.g., from small round-off errors compounding along the way) have been largely controlled for. For example, it is unlikely that one will forecast an 800 mb tropical cyclone with 250 mph winds hitting New York City, even under the most extreme climate change scenario (the lowest tropical cyclone central pressure ever recorded is 879 mb, and Hurricane Sandy was around 950 mb when it hit the New York area).
Moreover, if we do see something model-generated that is unusual, it is important to pay attention to it—to see whether the model continues to generate a similar feature either with slightly different initial conditions or with subsequent forecasts, or whether other models produce the same unusual result. The practice of model consensus or consistency adds to our confidence that the feature will occur—that small errors in initial conditions won’t affect that result, however unusual.
The limitations that a chaotic atmosphere place on day-to-day weather forecasts do not preclude the ability for models to numerically simulate the weather extremes that could occur—at some point in time. This last point is important—perhaps as important as forecasting exactly when an event will occur.
The ability to simulate an extreme event at all that may occur at some point in the future is important from a preparedness standpoint. It gives us a hint as to the type of black swan events that may loom in our future. Both long-term resilience measures and real-time emergency response procedures have to be designed to anticipate future catastrophic weather events—especially as they may be affected by climate change.
A Dynamical Basis for Modeling the Unprecedented
Climate change is making extreme events more intense and more frequent—in particular, heavy precipitation (Kirchmeier-Young and Zhang 2020) and more recently, wildfires (Abatzoglou and Williams 2016). It is expected that events of unprecedented intensity will likely continue to occur (e.g., Hurricane Harvey-like) leading to potentially high losses, so historical hindsight may not be sufficient.
The combined impact from hurricanes and wildfire in the U.S. in both 2020 and 2017 was not necessarily by happenstance. Large-scale atmospheric circulations can provide a conducive environment for such correlated extreme events to occur.
Necessary foresight can come from climate models, and especially from those that simulate ultra-low probability ways in which the climate itself could change dramatically, such as from the Atlantic Meridional Overturning Circulation (a system of ocean currents) completely shutting down (Bellomo et al. 2021). While capabilities to provide that foresight may currently be hampered by model resolution, model biases, and model errors, these models, because of their physical underpinnings, can certainly give advanced notice to what might be in store as the climate continues to change—in terms of simulating cruder versions of events, which may be considered unprecedented when compared to model-generated versions of similar events from past climates.
For example, a model-simulated version of Hurricane Harvey for the current climate may only generate 50 inches of rainfall because of above-noted limitations, but if a future climate version were to yield 60 inches, that would be an indication that events causing even more rainfall are possible. Combined with scientific insight that can be enhanced through machine learning (Prabhat et al. 2021), physical downscaling the coarse model output to capture smaller-scale event features, and Monte Carlo simulations to generate probabilities, it may be possible to become aware of weather events currently deemed impossible—think of Hurricane Harvey-scale event impacting New York City, a European 2003 (or 1540) heat wave in contemporary China, or whether a tornado outbreak of the scale that hit the U.S. in 2011 could impact Europe in the future.
Moreover, because of their global perspective, such models could hint at what kinds of combinations of such extreme events would be possible in a region or around the world in the same year. The combined impact from hurricanes and wildfire in the U.S. in both 2020 and 2017 was not necessarily by happenstance. Large-scale atmospheric circulations, the kind that general circulation models (GCMs) examine, can provide a conducive environment for such correlated extreme events to occur.
For both 2017 and 2020, for example, a moderate La Niña was in place; that weather phenomenon provided a large high-pressure ridge over the western U.S., which accelerates drying of vegetation and creating (wild)fire fuel load. It also enhances down-sloping winds across California, which can further promote drying and spread a fire once it has started. A La Niña also reduces wind shear over the Caribbean Main Development Region, which allows strong hurricanes to develop.
So, even though seasonal forecast models at the time did not tell us that there would be record-breaking wildfires and record-breaking hurricane activity in the U.S. in 2017 and 2020, perhaps we could have realized—based on past data—what the models were projecting for the summer and fall of 2017 and 2020, and we should have anticipated the elevated activity. Exacerbating this natural weather phenomenon, climate change has been increasing sea surface temperatures, which provide the primary fuel for hurricanes to develop, and it has been steadily increasing the dryness in the Western U.S. (Williams et al., 2020).
How Catastrophe Models Can Help Us Anticipate Black Swans
We’ve already demonstrated the ability of models to help us understand the probabilities and potential impacts of single, catastrophic weather event. But what about the prospect of multiple, exceedingly rare catastrophes in the same year?
Catastrophe models typically provide probabilistic views of individual perils, so the probability of a year with a 500-year flood loss is not correlated in any way to the probability of a 500-year wildfire loss or a 500-year hurricane loss. But it is likely that the probability of all such levels of loss occurring in each peril in the same year is greater than simply the product of the individual peril-probabilities, which in this case would be .002^3 = 8 x 10-9, or about 1 in 100 million.
More exact probability could be estimated by using a very high resolution GCM or strategies similar to what Verisk is developing (Sousounis et al. 2021), which will create globally correlated catalogs of events such that a year of hurricane events in a hurricane model catalog would pair up with a year of wildfire events in a wildfire model catalog, which would in turn pair up with years of other peril model catalogs’ events. Through the combination of physics-based equations to numerically and physically model extreme weather and climate patterns; machine learning techniques to downscale information from those models to better capture extreme intensities; and Monte Carlo simulations to better quantify the probabilities of occurrence—all in a globally spatially consistent way—catastrophe models will be better able to capture black swan hurricanes, floods, wildfires, and other atmospheric perils, making these events more foreseeable, which will lead to better awareness and preparedness.