Wednesday, December 16, 2020

Hurricane Rates Fitted to the Poisson Distribution, as Evidence of Climate Change

This 2020 Atlantic hurricane season set the all-time record (back to 1851) for named storms (30), and had the 2nd most number of hurricanes (13) and major hurricanes (6, 7-way tie for 2nd) (both of these records held by 2005).

One year does not make a trend - so the larger question is whether hurricanes/named storms are a Poisson process, and can we can use the Poisson distribution to determine whether recent years are exhibiting an increased trend. This has been shown mathematically, and the criteria certainly seems to fit tropical systems, just like it does with shark attacks:

The Poisson distribution arises in connection with Poisson processes. It applies to various phenomena of discrete properties (that is, those that may happen 0, 1, 2, 3, ... times during a given period of time or in a given area) whenever the probability of the phenomenon happening is constant in time or space.

The key part here is "probability of the phenomenon happening is constant in time" - is the probability of a hurricane/named storm increasing over time? 

In each chart, I'll present two comparisons - one over the entire HURDAT history of tropical cyclones in the Atlantic (1851 to now), and another just in the satellite era, which is 1967 to now. While NOAA undergoes extensive reanalysis projects to fill in past years prior to satellites, it is still very possible that more storms are named today due to increases in data and technology, including satellites:


To account for this, I'll do both 1851-Now and 1967-Now (satellite era), across:
  • Named storms
  • Hurricanes
  • Major Hurricanes (Category 3 or Higher)
  • Accumulated Cyclone Energy - ACE (more on this later)

To fit the best distribution parameters, I used the Kolmogorov-Smirnov test to minimize the distance between what actually happened and the theoretical distribution fit. I then take 1 minus this figure to assess, as a percentage, how well the distribution fits (so the closer to 1, the better). Similarly, I've previously used the exponential distribution to disprove the notion of "momentum" in college football, and assess how likely it is to skip a TV timeout in college basketball. For transparency, all data, charts, and distribution fits are here.

In graph form, it's hard to tease out trends - so we'll do so using the statistical distributions.


Named Storms

Goodness of fit: 89.1%
Goodness of fit: 87.7%

The Poisson approach seems to fit, especially over the entire history in HURDAT. The deviations we're interested in are to the right - in the sat era, years with 18 or more named storms. There are 8 such instances since 1967 - with 7 of them (87.5%) coming in the past 25 years (the most recent half of the sat era). The most extreme spikes, at 28 and 30 named storms, are 2005 and 2020 - both in the past 15 years.

Hurricanes

Goodness of fit: 93.11%

Goodness of fit: 93.05%

The Poisson fit works even better for hurricanes. In the sat era, years with 12 or more hurricanes deviate from the distribution. There are 4 such instances since 1967 - with 3 of them (75%) coming in the past 25 years (all 3 in the last 15 years). The most extreme spikes, at 13 and 15 storms, once again are 2005 and 2020.

Major Hurricanes (Category 3 or Higher)

Goodness of fit: 93.5%

Goodness of fit: 86.7%

The Poisson fit works best for major hurricanes over the entire database. But that's partly because in the sat era there have been so many deviations - years with 5 or more major hurricanes. There are 10 such instances since 1967 - with 9 of them (90%) coming in the past 25 years. As described previously, 2005 has the all-time record at 7, with 2020 tying for second at 6. 

Accumulated Cyclone Energy - ACE

Named storms assess the maximum intensity a storm achieves - but it doesn't illustrate how strong the storm is over time, i.e. the total energy generated over the storm's entire existence. Accumulated cyclone energy, or ACE, is the better wholistic measure to assess total strength:

The calculation takes a tropical cyclone's maximum sustained winds every six hours and multiplies it by itself to generate the values. These values are then added together which become a total for a storm...



ACE necessitates a different distribution fit as well (log-normal has been shown to be the best fit for ACE).

Goodness of fit: 96.7%
Goodness of fit: 91.2%

This log-normal fit works very well over the entire HURDAT database. Looking at years in the sat era that have an ACE > 1 standard deviation above the mean gives us a cutoff of around 165 - which reflects the 3 blue peaks in the second chart above, and represents the top ~15% of data. There are 11 such instances since 1967 - with 10 of them (91%) coming in the past 25 years. The wrinkle here is 2020 was not close to a record-breaking year (ACE of 180.3) - this represents the 90.6 percentile, and ranks #6 in the satellite era - behind 2005, 1995, 2004, 2017, and 1998.

Conclusion

As with any conclusion in statistics, correlation does not imply causation - but especially in the messy reality of real-world data, it's often the best we can do. The vast majority of adverse events (the outliers to the right of all of these charts) have happened in the past 25 years - indicating that at the very least, abnormally active Atlantic tropical seasons are more common now than before, regardless of what is causing the increase.

Arguments Against Climate Change Being the Cause of Increased Rates of Storms
Arguments For Climate Change Being the Cause of Increased Rates of Storms
Additional Data On Climate Change, Presented As Is

No comments:

Post a Comment