Weather Nerd

In which I compare the National Hurricane Center to Nate Silver

Alternative title: “Why is the trend your friend?”

In comments on my previous post, a commenter quoted my statement that the most significant thing about the overnight computer model runs “isn’t the exact landfall point, which is subject to huge errors this far out. The point is the trend. The trend is your friend. And the rightward trend is continuing.” He then asked a very reasonable question:

This confuses me. It seems to say that the best estimate of the track of the storm is not the current projection, but, because the last few projections have trended east, the best estimate now is somewhere east of the current projection. But… if trend works like that, wouldn’t the models take such trend into account already? Making the current projection itself the best estimate? (Sort of an Efficient Market Hypothesis kind of thingie.)

Below is my response:

Fair question, Pseudolus. The “trend is your friend” concept is admittedly unscientific, or at least, not totally scientific. And the National Hurricane Center certainly doesn’t subscribe to it. They base their forecasts primarily on the latest computer model consensus, with a bias, if anything, toward looking back at the previous model consensus (because they’re conservative and don’t want to vacillate wildly back and forth in the event of the models “ping-ponging” between solutions, as sometimes happens), not at looking forward toward an anticipated future model consensus.

But we armchair storm-watchers (and plenty of actual meteorologists, too) see the “trend is your friend” scenario play out over and over again, frequently enough that do we put some stock in it, while certainly not ignoring the present forecasts. To understand why, let me make an analogy to polling, which PajamasMedia/InstaPundit’s politically minded readership probably has more familiarity with than meteorological computer models.

Think of each individual model run as analogous to an individual poll during the presidential election season, and a model as analogous to a pollster. Some models are more reliable than others — AVN or the European model are sort of like Gallup, while another model, let’s say NOGAPS, might be more like Zogby. Also, some individual model “runs” (like individual polls by a particular pollster) may be outliers: just as you might get a wonky Gallup poll showing a 5-point deviation from the other polls, you might get an anomalous GFS run here and there (which is why the NHC is conservative about immediately shifting its forecasts when the models shift).

The NHC’s forecast is sort of like a Nate Silver analysis of the polls on FiveThirtyEight. It takes all the various models, including “models of models” that synthesize the other models together to come up with a consensus (like the or RealClearPolitics “polls of polls,” or FiveThirtyEight’s regression analysis), and then it sprinkles in some subjective/qualitative analysis, maybe even a little bit of “gut” feeling, and comes up with a forecast that’s loosely based on — but not slavishly duplicative of — the underlying models.

Where does the “trend is your friend” concept fit in? Well, suppose you’re a political pundit, and you’re looking at the same poll data that Nate Silver and his model are looking at, but you think you’re noticing something he isn’t adequately calling attention to: that, hypothetically, Obama’s lead over Perry has been steadily shrinking in the Gallup poll and others, from 8 points to 6 points to 3 points to 1 point. You look at this and think, gee, that “trend” could indicate some underlying pattern, and if the trend is likely to continue, Perry should really be considered the favorite at this point, even though Obama is nominally still ahead.

However, if you’re just looking at the latest snapshot polls (like the current model runs), Obama’s up by 1. FiveThirtyEight’s regression analysis, meanwhile (like an NHC forecast), probably has Obama up 2 or 3, because the prior polls have a lingering impact on the analysis, and hypothetical future polls aren’t factored in at all. [NOTE: This may not literally be true. Silver’s regression model may account for “trends.” I’m not sure. And even if the model doesn’t, Silver himself certainly would, in his write-up. But hey, it’s a loose analogy.] Yet those hypothetical future polls weigh heavily on your mind, because you’re looking at it from a more “gut” or “common sense” perspective, and trends that drastic are often a sign of things to come.

So… who’s right in this hypothetical analogous scenario? Is Obama up 2 or 3, or is he up 1, or is the “trend your friend” and Perry’s really the favorite? Certainly, trends can and sometimes do reverse. But they often don’t. I can’t back this up with mathematical proof, but it’s my anecdotal observation that, with storms of this type/track, when the trend is this pronounced, it more often than not continues, and rarely outright reverses. That’s no guarantee, of course — far from it — but then, neither is any individual model run or NHC forecast a guarantee. In my mind, the “trend” is something to weigh along with the other data: you don’t assume it’s right, but you also don’t just ignore it. So, just as you might well think Perry is the favorite in the scenario I’ve described, I view further eastward shifting as a slight “favorite” right now.

That said, all it takes is one or two polls showing Obama suddenly up 5 again to totally change the picture… and that’s why the forthcoming runs of the AVN and Euro models, due out in the next few hours, are so critical. If the eastward trend stops or reverses, then we have a whole new situation.

By the way, Irene is now a major hurricane, as expected, and is forecast to be a Category Four within 24 hours. After that, intensity will likely “fluctuate” for a few days, due to eyewall replacement cycles, before weakening begins as wind shear increases and the water below the storm cools. As for track, there’s really nothing new to report. As I alluded above, we’re waiting on the next round of computer model runs. Stay tuned.