When I approach attempting to understand catastrophe risk in our warming climate, I've mentioned in my last blog post how we need to be open to all points of view, so I listen to everything. I think it's healthiest to approach the topic from all angles but always try to form your own take on the situation - which in this case probably repeats what others have found out already - but I wanted to show this with an example using some global hurricane landfall data.
An interesting article in Forbes recently looked at the changes in worldwide hurricane landfalls since 1970. The authors commented very reasonably on the data below: "There are a lot of ups and downs in the data, but no obvious trends", which makes a lot of sense to me if you look at the data:
The black bars are the Cat 1-2 landfalls and the grey bar on top of them is the Cat 3-5s. Clearly from eyeballing here I'd completely agree: there's obviously nothing startling to write home about.
Just out of interest I took the data and decided to look at the two classifications (Cat 1-2 and Cat 3-5) separately - there's a reason for this I'll detail below - to see if there's any trend in the individual datasets rather than stacking them atop one another:
We see a weakly declining rate of Cat 1-2 events and an increasing rate of Cat 3-4-5 events. Reading off the trend line on either axis it suggests Cat 1-2 landfalling events dropping from around 11.5 a year in 1970 to around 9.5 a year in 2020. The Cat 3-5 landfalling events changes from around 3 a year to around 7 a year in the same time-frame, so proportionally a bigger change than the decrease in the Cat 1-2s, which admittedly caught my eye.
Now, from a purely US Hurricane loss modelling standpoint, if we're being honest, it's the Cat 3-4-5s that dominate the our loss curves. You can test it for yourself in your local friendly catastrophe model. Assign each event its highest landfall strength by category and calculate an EP curve in descending order of loss. See how many Cat 3-4-5 events are above around the 15-20 year return period: the EP curve is chock-full of them. (That's not to belittle the strength of a Cat 2 and its loss-making power: one of these into Miami or Houston or New York and our balance sheets would still know about it). However, in the bigger picture, we're really all about the Cat 3-4-5 events in catastrophe modelling.
So I wanted to dig into this shift of Cat 3-4-5 events over time. But it still bugged me - is it really worth investigating more into this? That trend in that chart isn't very steep after all. Anyhow, I ploughed on and looked at the data a little differently. Now, I'm no statistician so I'm not going to get too detailed about this but I've done a very simple test on the data that I've summarised in the table below:
We can see here that as we move from the first period (1970-1994) to the second period (1995-2019) the number of tropical systems making landfall has increased by 13%. The Cat 1-2s are basically level, but the Cat 3-5s have increased by about 50% in this time. Now, I want to understand simply whether this might be down to chance.
So, look at this table now that summarises the data in the whole period:
If there's no change in behaviour across the period of the dataset, 32.6% of events in our 1970-2019 period end up as Cat 3-5. Here, we can run some simulations where for each event, it has a a 32.6% chance of it becoming at Cat 3-4-5 event. What I simply want to understand was:
"What are the chances that of the 410 events in the period 1995-2019, 152 of them could become Cat 3-5 storms, assuming that 32.6% of all events becoming Cat 3-5s is steady across the entire 1970-2019 period"?
In other words, I wanted to understand what are the chances that 152 Cat 3-5 events out of the 410 events happened in the 1995-2019 by chance.
Now I realise you could do this with a single equation, but I decided to have more fun and re-simulate the 1995-2019 period 500 times, to see how many of the 410 storms in each simulation actually ended up as Cat 3-5. Remember each event has a 32.6% chance of becoming a Cat 3-5. Here's what I get - and I've ordered the 500 simulations of 410 events in decreasing order of the numbers of Cat 3-5 events in each simulation.
The line across the middle corresponds to the 152 Cat 3-4-5 events we saw in reality. Of the 500 "simulations", I was surprised to see that actually only 16 of them produced more than 152 Cat 3-4-5 events. Or we could put it another way and say that there is 3.2% chance of the Cat 3-4-5 rate happening by chance based on the overall rate of landfalling events worldwide since 1970.
In this very simple study, I'm suggesting the chances of the recent increases in numbers of Cat 3-4-5s being down to chance are non-zero but still actually pretty small.
One caveat here is clearly we have the backdrop of the Atlantic Multidecadal Oscillation contributing to the US Atlantic basin numbers here post-1995, but I should add that this is a global dataset.
The above made me think that there's maybe more value in digging into this behaviour a little more, if you'd like to read on...
A Quick Look at Sea Surface Temperatures (SSTs)
With this upward shift in intensity in mind, I wanted to do something - again very straightforward - to see how the "fuel" for intense hurricane activity has changed since 1970. Without a warm sea, you're not getting hurricanes. So all I've done is drawn two boxes that many hurricanes pass through en route to potential landfalls in Florida (left) and Japan (right) to see how the temperature of the sea surface has changed within these boxes. They're shown below. Only the sea grid-points are used in the calculation and we've used the OISST v2 dataset.
What I've done in the chart below is show the percentage of day each year between July and November where the maximum temperature somewhere within these boxes exceeds a specific temperature. I chose maximum temperature here to get a feel for changes in the extremes, nothing more.
I've started by picking a temperature that occurs fairly frequently. It's slightly different for Japan (28c) and Florida (27c) and the chart below shows how frequently the maximum temperature in each respective box exceeds this value each year as we move from 1982-2019 (the time extent of this dataset).
You can see how for both territories there is the hint of an upward shift in the frequency of these relatively common maximum SSTs: nothing really significant here however, so no real indication here that seas are getting notably warmer.
What happens if we change the SST criteria and we move more towards the sort of temperatures that might sustain a stronger category of storm? We show below the same chart as above but with a slightly higher threshold on the SSTs in each box:
I see something interesting happening here. Now we're looking at a slightly higher SST cut-off we can see the slope of the best-fit lines is steeper and we see the frequency of exceedance of this higher maximum SST within each box clearly increasing over time. As we've said before, we've got the Atlantic Multidecadal Oscillation in the Florida data moving into a warm phase here after 1995 to explain some of the shift, but we've also got the same trend in the data for Japan.
Does this suggest that we're seeing an increasing tendency towards more frequent high maxima at the higher end of the range?
To help understand this, we can look at this data one more way. The chart below shows the percentage daily (Jul-Nov) frequency of maximum sea surface temperature (in 0.5c bins) for the first half (1982-2000) versus the second half (2001-2019) of the SST dataset in the Florida (red) and Japan (blue) boxes.
The solid lines here show 1982-2000 and the dashed lines 2001-2019, with the colours for the two regions. The shape of the SST frequency curve is similar but even in this small shift in time of 20 years, the data seems to have translated along the x-axis to higher temperatures.
If we take an example from the Japan curves, the percentage of days between July and November when the daily maximum SST in the box is either 30.5 or 31c increases from 17% of all the day to 28% - we're seeing about 65% increase in the days when the SST maximum is this high.
The most obvious and notable thing - highlighted in the shaded area - is how much more frequently we are seeing the sort of maximum SSTs in the tail of the distribution that can support stronger hurricanes in the boxes close to landfall in Japan and Florida/southeast US.
Clearly this is a slightly myopic view and there's not just SST that drives hurricanes: vertical wind shear, Saharan dust and larger scale climatic variations such as El Nino/La Nina and their changes in a warmer climate to name but a few could obviously have an influence but as I mentioned earlier: no warm seas, no hurricanes. However the one thing in addtional we should note here is that whatever shift in SSTs we might be seeing here has already happened.
So there might something here that merits a bit more of a deep dive. Let's put it in some bullet points to attempt to summarise all the above - if you're still with me. And if you're not, please email me!
All the thoughts below relate to a time-frame of the last 40-50 years per the above datasets:
The frequency of Cat 1-5 hurricanes globally hasn't really changed much
The frequency of occurrence of SSTs able to support hurricanes hasn't really changed much: at least in our Florida and Japan example
The frequency of Cat 3-5 hurricanes seems to be increasing and possibly not down to chance
The frequency of occurrence of the higher SSTs able to support higher-end hurricanes seems to be increasing: at least in our Florida and Japan example
So, do you see what I'm seeing? I find it fascinating that we're seeing increased frequency of the warmer end of the SST spectrum that might point to why we might be seeing a few more strong tropical systems, but less of a signal when we spread our analysis to a wider group of sea-surface temperatures and all hurricane categories.
This is only really a day or so's worth of work and would like to look at it a bit more but thought it was worth sharing for comment and further discussion, especially when trying to relate this type of analysis to catastrophe risk, where we're probably more interested in the tail events.
If this conjecture sounds sensible to you (and it may not!) then it's all the more important that we as an industry understand how fluid present-day hurricane risk might be given that there are what seem to be some suggestions of its fluidity right in front of our eyes.