Before U.S. landfall, NHC issued 45 main forecasts for Irma, plus a bunch of intermediate ones. Their first was way back on Aug. 30, when Irma was just off the coast of Africa. By most metrics, the forecast was a good one. The idea of a major hurricane was there, the track fell within all the cones.
As shown above, the first day or two of the forecast tended to be nearly perfect. By days three, four and five (not shown, but here), some larger error is seen. Overall, the error is within or better than historical expectations, but it is more pronounced near the critical turn toward Florida.
Despite the barrage of data in the lead-up to a storm, these graphics should remind us that it's hard to forecast a hurricane more than a few days out.
The cone of uncertainty is not determined by the forecast uncertainty for that particular storm. Instead, it represents an average track error over the past five years. As forecasts improve each year, the cone shrinks. During the 2017 hurricane season, the cone size is about 52 miles for a 24-hour forecast. That grows to 90 miles by day three and 243 miles by day five.
The distance from Miami to the mainland Florida point of landfall in Irma is about 90 miles. The forecasts that became extremely concerning for Miami were generally in the three- to four-day period, when an error of 90 to 180 miles might be anticipated. What we saw is a several-day-out forecast that is working, and much improved even compared with the days of Katrina.
The European model performed very well in forecasting Hurricane Irma.
For most of the storm, the Euro's track error was the best at all hours when you compare it with other models. In some cases, the Euro was beating the American model's three-day error at day five. It was also one of the first to show the storm making landfall on the north Cuba coast.
But it also had plenty of moments where it did not perform well. As Irma was making its final approach to Florida, it could actually be argued that the Euro did not have the best forecast. It (and in some ways, the official NHC forecast) remained stubbornly west of the eventual track. U.S. models performed better.
A unique challenge in Irma's forecast was the shape and orientation of the Florida peninsula. On a global scale, the peninsula is tiny. It only spans 140 miles at its widest point, and it's closer to 100 miles in most areas. Given what we know about the cone of uncertainty, that means we were never going to have much confidence in the forecast until the final days before landfall.
These factors can shift a Miami threat to Tampa quite quickly, even though the cities seem distant. Hurricane Charley (2004) was initially expected to hit Tampa before it destroyed Punta Gorda. Oblique approaches mean a fraction of a difference in angle can cause landfall to shift dozens or even hundreds of miles up the shore.
It's not just a Florida problem, though. We don't have to look too far into the past to find a high-stakes tropical forecast that turned on a dime over a few miles at the beginning of the forecast. Hurricane Joaquin briefly looked like it was going to blast into the East Coast. There was model consensus. Then it didn't happen. Cases related to crippling megalopolis blizzards and rumors of such are all too common as well.
The United States needs a "moon shot" program to dramatically improve weather prediction.— Ryan Maue (@RyanMaue) September 9, 2017
Given the massive economic and social ties to accurate weather forecasts during severe events, perhaps we need to demand even more out of modeling. If down-to-the-mile is critical to know, we need to get down to the mile, whatever it takes.
While I tend to think a lot of these problems can be helped along by private industry — IBM, for one, has entered the ring with Deep Thunder — it is also time we more seriously considered extreme weather the national security threat it is. Such a shift in thinking may help increase the attention and funding needed to get to the next level of forecasting.
Data used in this post come from the National Hurricane Center.