The authors first review a system for specifying monthly mean anomalies of midday temperature (T), dew-point (D), and wind speed (W) at a large network of surface stations across the United States. Multiple regression equations containing approximately three terms were derived for each element and month from concurrent fields of 700-mb height anomaly (H) observed over North America and vicinity, plus the previous month's observed local anomaly of T, D, or W, during the 20-year period 1964-1983. Results of testing this forecast system on prognostic values of H prepared twice a month at the National Weather Service from 1973 to 1990 and on observed values of H during the independent period from 1984 to 1990 are presented. The authors pooled the data for all months and 122 stations to compute Heidke skill scores (HSS) for two, three, four, and five classes for each weather element. All scores showed skill that dropped steadily as the number of classes increased. In all cases skill was highest for T and lowest for W, the same result as that obtained in the original derivation. In order to examine the distribution of skill, we computed the HSS separately for each class. The skill scores for the first and last classes were greater than those for the middle class for all elements and months. It is encouraging that forecasts above the 90th percentile, the most critical category for fire potential, consistently showed greater skill than forecasts for any other class. Geographical and seasonal variations of skill based on prognostic height was also examined. The results showed that skill is strongly dependent on location, month, and weather element. Mean annual scores were positive for each element in all parts of the country, while monthly scores were highest in January and lowest in October.Although the forecasts based on prognostic heights were more skillful than climatology for all weather elements, they improved over month-to-month persistence only for T and D. If the height forecasts had been perfect, the skill scores for these elements would have been about two to three times as large as those for prognostic heights and four to five times as large as the persistence scores. When the test period was divided into two subperiods, values of HSS were considerably higher in the later period, despite a drop in persistence scores from the early to the later period. The authors attribute this improvement to increased accuracy of medium-range numerical model predictions during the 1980s.