University Park, Pa. --- Can a computer program win a weather forecasting contest? Not yet, but a Penn State statistical weather forecasting program did beat the consensus forecast in the 1996-97 National Collegiate Weather Forecasting Contest.
"The computer program came in 20th out of 737 participants," says Dr. Robert Vislocky, research assistant in meteorology. "Not too bad for a quick and dirty prototype program."
Vislocky took his ideas, which he and Dr. J. Michael Fritsch, professor of meteorology, tested at Penn State, and created an automated forecasting system called Advanced Model Output Statistics, AMOS, for the weather contest. AMOS ingests routinely available weather data and, by applying statistical processes for the specific forecast location, predicts precipitation and maximum and minimum temperatures.
Although AMOS was designed for only the cities and variables important for this contest, Vislocky points out that the system can be expanded to forecast for any location and for a vast array of parameters.
The National Collegiate Weather Forecasting Contest is held each year and pits all ranges of forecasters against each other in a 26-week contest. For each two-week block, participants prepare a daily forecast for the same city. The forecasts are due by 7:00 p.m. and cover the period beginning six hours later to 30 hours later. After two weeks, the contest moves on to the next city.
The 1996-97 contest began with 737 participants and ended with 383 active -- those who continued to prepare forecasts. Sixty-seven professional meteorologists, 166 freshmen and sophomores, 324 juniors and seniors and 180 graduate students participated in the contest.
The consensus forecast, consisting of the average of all the human forecasts entered, came in twenty-third, the researchers report in a recent article in the Bulletin of the American Meteorological Society. Two National Weather Service computer programs came in 175th and 266th.
"We are especially gratified to have beaten the consensus," says Vislocky. "Consensus is hard to beat because it represents the cumulative knowledge of all the forecasters in the contest, and this is the first time a computer program has been able to beat it over an extensive verification period.
"Anyone can perform well for one of the cities just from good fortune, but sustaining that kind of accuracy without good forecasting ability is not possible over the length of the contest." While computers play an important part in weather forecasting today, they have mostly focused on dynamic modeling of weather systems, not statistical modeling.
"Not nearly as much money has been devoted to the development of statistical forecast techniques as has gone into other forecast methods," says Vislocky. "As a result, statistical applications in meteorology have not progressed as rapidly. Many other aspects of the weather forecasting process have become more automated. Now, with the development of AMOS, the actual forecasts can benefit as well."
While Vislocky admits that a program like AMOS will almost always have difficulty beating the very best forecasters, especially when there is ample time to prepare the forecast as in the contest, a fairly preliminary statistical program like AMOS did beat 97 percent of all the original entrants and 90 percent of all professional forecasters. If this type of program proves to be as accurate as most humans, there are numerous practical applications.
"It takes a while for a human to do a forecast," notes Vislocky of Penn State. "If users want continuous updates around the clock for thousands of different cities and variables, human forecasters would be hard pressed to satisfy even a small portion of that need. A computer program like AMOS could supply that information in a timely manner for any parameter and any location in the world."
The above post is reprinted from materials provided by Penn State. Note: Materials may be edited for content and length.
Cite This Page: