Tuned perfect prognosis forecasts of mesoscale snowfall for southern Ontario

Abstract

Listen

Translate article

A procedure for producing site‐specific 1‐ and 2‐day categorical forecasts of 24‐hour accumulated snowfall by statistical forecast methods has been developed and tested for a small area of Ontario adjacent to southern Georgian Bay. A perfect prognosis (“perfect prog,” or PP) method was used, with predictors designed to handle lake‐effect and nonlake‐effect snowfall. Predictors were selected from a basic set of potential predictors by a stepwise multiple discriminant analysis (MDA) procedure done in three stages, where the third stage involved adding functions of predictors already selected in the first two stages to the basic predictor set. The third stage appears to enhance the discriminating power of the original predictor set because the number of “hits” of snowfall forecasts made with independent data was significantly increased and the distribution of forecasts was brought closer to the observed distribution. A two‐step, rule‐based tuning procedure was applied to the PP‐MDA forecasts to help compensate for errors that arise when the PP‐MDA statistical equations are used with numerical weather prediction model data, and for errors that arise from the conservative nature of MDA forecasts. A rule‐based nonparametric statistical classification procedure (CART) was used in the first step. When the rules for tuning forecasts were tested with independent data, CART was found to increase the skill of the tuned forecasts, particularly in the common categories (1,2), and to improve the reliability of category 1 forecasts at a majority of the stations. However, CART was unable to find rules for infrequent and rare snow categories. Step B of the tuning procedure, a semicomputerized manual search for additional rules not seen by CART, was undertaken in an attempt to “do something” about this problem. When tested with independent data, overall improvement was found in the skill of forecasts tuned by two‐step procedure, but it was too small to make an appreciable difference. Several suggestions are made in regard to exploring methods that should result in significantly improved skill of snowfall forecasts for southern Ontario by statistical forecast methods.