Accuracy of 24- and 48-hour forecasts of Haines' Index

摘要

The University of Wisconsin-Madison produces Webaccessible, 24- and 48-hour forecasts of the Haines Index (a tool used to measure the atmospheric potential for large wildfire development) for most of North America using its nonhydrostatic modeling system. The authors examined the accuracy of these forecasts using data from 1999 and 2000. Measures used include root-mean-square error (RMSE), mean error (ME), percent misforecast (PM), correlation between error and observed Index value, and Kuiper skill score (K). Average values of these measures indicate that the model forecasts are typically too low (RMSE = 0.74, ME = -0.01 for 24-hour forecasts), about a third (PM=38 percent) of the point forecasts for the year were misforecast, but that the forecasts are more accurate than random forecasts (K= 0.48). The correlation between error and observed value shows that, in general, the model's predictions are a bit too extreme, pushing the Index values to the ends of the spectrum. Statistics for 48-hour forecasts were similar to those for 24-hour forecasts, indicating that errors arose primarily during the first 24-hours of forecast simulations. Monthly values of the accuracy statistics show winter forecasts to be most accurate and summer forecasts to beleast accurate.