|Greg Young and Claudia Tebaldi. (Photo by Carlye Calvin.)|
The unusually mild weather across the Midwest this winter has given forecasters more headaches than usual. Forecasts for next-day temperatures have at times been as much as 20°F (12°C) too cold. One problem has been the statistics packages that help bridge the gap between computer forecast models depicting large-scale features and the plain-language local forecasts that reach the public. Two papers from RAP scientists offer clues to improving the statistical guidance that can so easily go awry. Both were presented at the American Meteorological Societys annual meeting in Orlando, 1418 January.
According to Claudia Tebaldi, a statistician and project scientist in RAP, local forecasts can be improved by drawing on the experience of neighboring stations. Claudias work is based on a refinement of Model Output Statistics (MOS). For years, MOS has been the main tool used by the National Weather Service to predict local temperatures and the likelihood of rain or snow. Through a set of equations, MOS adjusts the forecast-model output for each city by using the long-term weather records at that site. This gives MOS the edge when the weather sticks to the climatological norm, but when conditions are extreme (as in this winters Midwest warmth), the MOS numbers tend to be too conservative.
As part of its recent forecast research, RAP developed a system called Dynamic MOS. It leans more heavily than standard MOS on the last 100 days of weather; thus, it can better reflect any unusual patterns that occur in a given season. Overall, Dynamic MOS tends to be more accurate than its predecessor. Claudia has found a way to improve the system even further through what she calls "trading historical for spatial information." The typical MOS formula for one point uses coefficients that are based on the model output only for that point. Claudia proposes drawing from nearby locations to provide a set of coefficients valid for an entire "neighborhood" as much as 600 miles (1,000 kilometers) across. This strategy, she says, may be more likely to capture the chance of an unusual weather event, such as a rainstorm in a dry climate, that might not show up in a 100-day record at a single site. In fact, just 40 days of weather history appear to be enough when you factor in neighboring locations.
Claudia tested this approach using one- and two-day model forecasts and observations from last summer at about 35 points across the western United States, stretching from Kansas to Nevada. In some cases, large neighborhoods shaved as much as 50% from the forecast error that occurs with single-point MOS. Claudia is now expanding the analysis to other areas and seasons and to longer-range outlooks: "We hope to find that the results we got by randomly choosing [this season and location] are representative of a larger group of circumstances."
Because each computer model has its own set of MOS equations, forecasters often have to juggle competing outlooks for a given city before they issue a prediction. Greg Young, an associate scientist in RAP, has found that a simple average of MOS numbers may perform better than more complicated blending strategies. Greg investigated five different schemes, including multiple linear regression, for blending the outlooks from different models. In looking at high-temperature forecasts for the next day in 18 cities through the summer of 2001, Greg found that a simple average of MOS from three currently used models performed "surprisingly well, and some form of it may perform even better."
Edited by David Hosansky,
Prepared for the Web by Carlye Calvin
Last revised: Fri Feb 22 17:08:40 MST 2001