The Oil Drum | National Hurricane Center and the Likelihood of Hurricanes

67 comments on National Hurricane Center and the Likelihood of Hurricanes

Comments can no longer be added to this story.

| Show without comments | PDF version

Halfin on March 14, 2006 - 10:44pm

Two comments. First, here is the corresponding forecast for 2006:

http://hurricane.atmos.colostate.edu/Forecasts/2005/dec2005/

It's up to 81% this year! That's after 68% and 69% the past two years. So I guess they're learning...

Second, it's probably not right to assume that the hurricanes are independent. I don't remember the whole history, but Katrina and Rita had similar tracks (at least at the end) and came in pretty close to each other. It might well be that once conditions arise that are favorable for a hurricane to hit, that we could be vulnerable to more than one.

Here's one odd thing from their forecasts. They are all like this, but this is the one for 2006:

Entire U.S. coastline - 81% (average for last century is 52%)
U.S. East Coast Including Peninsula Florida - 64% (average for last century is 31%)
Gulf Coast from the Florida Panhandle westward to Brownsville - 47% (average for last century is 30%)

The thing is, the probabilities don't add up, even though the subregions are geographically distinct. The probability for the whole coastline is 81%, the probability for the western part is 47% and for the eastern part is 64%. Those two parts add up to 111%.

What does this mean? Presumably, that the odds are not independent. The chances are not zero for hurricanes to hit both west of Florida and in the eastern part. I wonder if this could shed any light on their imputed probability for more than one hit, if we did some statistical mojo with these numbers?

Comments can no longer be added to this story.

plucky underdog on March 14, 2006 - 11:33pm

That's easy

Letmebefirstletmebefirstletmebefirst

p(X) = probability of event X
E = Hurricane in East, W = Hurricane in West
~X means NOT event X
& means logical OR
| means logical OR

p(E) = 0.64 so p(~E) = 0.36
p(W) = 0.47 so p(~W) = 0.53

so p(~E & ~W) = 0.36 * 0.53 = 0.19

so p(~(~E & ~W)) = p(E|W) = 1 - 0.19 = 0.81

QED. Cigar?

Comments can no longer be added to this story.

plucky underdog on March 14, 2006 - 11:34pm

Errata

& means logical AND

Oops

Comments can no longer be added to this story.

Halfin on March 15, 2006 - 1:29am

Thanks, Plucky, that does make sense! So it really does add up properly.

Comments can no longer be added to this story.

Stuart Staniford on March 15, 2006 - 12:52am

At an unconditional level, the independence assumption has to be wrong because we know there are things like sea surface temperatures in the tropical atlantic, the ENSO index, etc, that have a lot to say about whether there are lots of bad hurricanes or not.

However, those variables are available for NHC to incorporate into their predictions. So if you want to argue that the hurricanes are highly non-independent even when conditioned on the NHC forecast, that suggests there might be good features that NHC is not incorporating into their forecasts.

Comments can no longer be added to this story.

plucky underdog on March 15, 2006 - 1:22am

Correct - but surely the NHC's statisticians aren't that naive? Or maybe we've just learned an important lesson about the quality of Gubmint stats. Or maybe those probabilities are based on a much more sophisticated model, and the numbers just came out that way.

Poisson process, anyone? Or is there a system recovery time between major storms in a small province like Florida? Or did they just run a zillion simulations with random initial conditions and report the frequency of outcomes?

Comments can no longer be added to this story.

Dan Ur on March 16, 2006 - 3:10am

Having read the discussion sections for most of the major and some of the minor storms for the past 3 years over at NHC's website. You get the picture that there are about 5 to 10 computer models that they use almost all the time to predict the path of a storm then they use a few human guesses ( either a mix of the models, or a best guess from the group that have been doing this for years ). They have in the past, used human guessing to get things right or model's averaged to get things right. But they admit that they do not know all the factors and repeat a lot of the time that their models are NOT prefect.

The window of 3 to 6 hours ahead can be shaky at best, and anything longer than that has error factors. Last season those error factors came to full light, when they could not predict the outcomes of several late season storms and you could tell from reading them there was a massive amount of frustration going on.

The computer models available to them are only as good as the data streams they were designed around. Things are changing and they know it is happening, but they just can't tell you for sure what and how!! This makes their job even harder, because so many people now depend on them getting it right, when they the NHC know they can't get it right if the ocean keeps changing and they don't have new modeling programs that predict those changes.

Its like us asking Dave and Stuart to predict the Oil movement in a basin between March 16th at 5 pm and March 18th at 8 pm and tell us where the next blowout will occur before it happens.

Models and programs of a changing world are only so good as the "Past" data used to make them.

Comments can no longer be added to this story.

Gets IT on March 15, 2006 - 6:16am

http://www.fema.gov/hazus/dl_rat.shtm

FEMA has developed a companion software tool to HAZUS-MH called the HAZUS-MH Risk Assessment Tool (RAT) to help you produce your risk assessment outputs for earthquakes, floods, and hurricanes. This tool was developed as a third-party model to support HAZUS-MH and is used to display the outputs from the HAZUS-MH risk assessment in an easy-to-use format. The RAT pulls natural hazard data, inventory data, and loss estimate data into pre-formatted summary tables and text. These summaries can support the presentation of data to decision-makers and other stakeholders and in your mitigation plan.

bullet Risk Assessment Tool (zip, 19.3 MB)

Comments can no longer be added to this story.

Halfin on March 16, 2006 - 3:36pm

I worked out the corresponding probabilities for 2006 using the Poisson distribution as Stuart described, and came up with the following probabilities for this season based on their 81% estimate:

p(1 or more) = 81%
p(2 or more) = 49%
p(3 or more) = 23%
p(4 or more) = 9%

So I think they would have to say again that the probability of this year being as bad as the previous year would be quite low, even with these larger probabilities.

Just out of curiosity, what would it have to be to get the probability with 4 or more to be at least 50%? This corresponds to a "lambda" of 3.672 which means a probability of 1 or more big storm hitting of 97.5%! I don't know if even the most stringent advocate of a linkage between global warming and hurricanes would say we are in that territory yet. So I suspect that any way you look at it, we had a substantial bit of bad luck these past two years, and are unlikely to be that unlucky again this year.

Comments can no longer be added to this story.

Stuart Staniford on March 16, 2006 - 4:14pm

This analysis has made it clear to me that there's a real mystery here in 2004,2005. While I stand by my general sense that the Webster/Emmanuel/Hansen "global warming is driving SSTs higher and storms are going to dissipate ever more power" idea is correct as an explanation of the last few decades, it also seems that these last two years are anomalous (in a highly significant way) relative to the trend in terms of what hit the US.

One possibility, as you suggest, is that they are just exceptionally bad luck, and 2006 and 2007 will mostly return to the trend of gradual worsening that's been going on for decades (in which case the NHC's 81% is a reasonable looking estimate). The other possibility is that we have hit some kind of non-linearity in the way the climate responds to increasing global warming forcing, and we are in a new (and much worse) regime. In that case, our best estimate of lambda should be 3.something and the probability of 1 or more storms should be 90%+. I don't think we have enough information to argue one way or the other very convincingly at this point.

If we are crossing some kind of threshold, there may be some flickering back and forth for a while. On the other hand, if we were to have another season or two like 04/05, I don't think that would leave much room for doubt. It would also help to have some kind of physical explanation for what switch had flipped (I have no idea at present, though changes in the Atlantic circulation seem like an interesting place to start looking given the Bryden et al paper).

Comments can no longer be added to this story.

67 comments on National Hurricane Center and the Likelihood of Hurricanes

Comments can no longer be added to this story.

| Show without comments | PDF version

“Of all races in an advanced stage of civilization, the American is the least accessible to long views… Always and everywhere in a hurry to get rich, he does not give a thought to remote consequences; he sees only present advantages… He does not remember, he does not feel, he lives in a materialist dream.”

—Moiseide Ostrogorski (1902, 302-303)