Statistical Prediction of Seasonal East Coast Winter Storm Frequency

Open Access
Publisher Website
Google Scholar
Abstract
A statistical method for predicting seasonal winter storm activity is developed using discriminant analysis. A large pool of plausible predictors is initially considered, which includes sea surface temperature anomalies from the Niño-3.4 region, the tropical Atlantic, and adjacent to the southeastern U.S. coast; land temperature over the southeastern United States; and values of the Southern, North Atlantic, and quasi-biennial oscillation indices. These values are screened using a chi-squared procedure to isolate the most relevant variables and averaging periods. A reduced pool of predictors is then used to develop a series of discriminant functions relating the variables to East Coast winter storm activity categories (e.g., above or below normal). The selection and addition of variables to the discriminant functions is based upon the cross-validated skill score. In all cases, the discriminant forecasting procedure exhibits a high degree of skill relative to a random reference forecast that is constrained to follow the unconditional climatological distribution of storms. For December through February (DJF) storm frequency, only 4 out of 46 seasons are misclassified among three overlapping activity categories (i.e., ≤5, 5–8, and ≥8 storms). For discrete boundaries (≤5, 6–7, and ≥8 storms), about a third of the seasons are misclassified with average seasons being the most problematic. In both cases, two-category errors occur only twice. Gulf sea surface temperatures during the previous storm season is the strongest discriminator of DJF storm activity, with warmer than normal Gulf temperatures associated with active seasons. It is shown that this variable is related to both El Niño and the North Atlantic Oscillation (NAO) phase, as heightened storm activity tends to occur during the positive phase of the NAO and El Niño conditions. Over the period from October–April, forecast skill is weaker, but still high. Using a two-category forecast with overlapping boundaries (≤12 and ≥11 storms), 84% of the active seasons and 71% of the nonactive seasons are predicted. For this longer season, sea surface temperature off the southeast U.S. coast during the preceding summer is the best discriminator. Again warmer than normal temperatures are associated with active storm seasons.