Hey there, Austin! So, you may have missed it, but there was a recent weather project calling for multiple inches of Austin-area snow today, and it’s worth talking about.
No, that snow is not going to fall on us. But that prediction is worth talking about. One of the things we’re planning to explore in Weather Watch is the idea of probability in forecasting: the odds of this or that happening, what that probability is based on, how that probability is expressed, and what the general public makes of a forecast.
This “inches of snow” phenomenon is a good example of an ongoing issue: the uses, and misuses, of statistical models.
Our winter tale starts two weekends ago, when Troy Kimmel, an Austin-area forecaster, noticed one of the highly respected weather-modeling systems, the European Centre for Medium-Range Weather Forecasts, spit out an interesting result: 6 inches of snow in Austin! With 10 inches on its way to northern Williamson County!!!
Whoa.
Austin’s weather-geek community had some fun with this.
“Monday’s edition of Fun With Computer Models” was how the TravisCountySevereWx Twitter feed, run by a volunteer group that assists the National Weather Service, summed up a projection that had already dropped since Kimmel noticed it.
Luckily, the “Wilco Snowpocalypse” prediction was so ludicrous that it didn’t set the Internet on fire. But sometimes predictions that are more plausible but based on flawed modeling get all over the Internet, said Troy Kimmel, an Austin forecaster and lecturer at the University of Texas. “Weather porn” became a thing. (Though one you shouldn’t investigate at your work computer.) The proliferation of flawed information has undermined a forecasting profession already smarting from the disconnect between what a forecast means and what the public wants it to mean.
For an example of the frustration over that disconnect, I present a Monday tweet from Atlanta’s WSB-TV weather intern Katie Martin, during the recent American Meteorological Society conference in New Orleans:
Forecasting is still a fairly new science. And as forecasting becomes less art and more science, its practitioners rely on increasingly complex mathematical models. Much of the information created by those models is available to the public. But those models are only as good as the variables programmed into them. Time is a particularly important variable. As Kimmel put it: “We can’t even get tomorrow’s weather right. How can we forecast eight days out?”
Michael Lyttle, one of the three people who run the TravisCountySevereWx Twitter feed, which among other services helps the National Weather Service disseminate accurate information, offers a useful rule of thumb. A week or more out, a forecast can give very general long-term guidance, but not specifics. Five days out, people can start to seriously look at forecast specifics. At one or two days out “we have good probabilities of what is going to happen,” Lyttle told the American-Statesman.
Snow never got to the high-probability stage. By the Monday after the initial projection, the amount of snow supposedly on its way to Austin had dropped to 3 inches. By last Tuesday, the model was predicting chilly rain. You can look outside to judge how the projection ultimately turned out.
But wait — odd projections of the type the European model spit out are apparently not a one-time deal:
