Analog Probabilistic Precipitation Forecasts Using GEFS Reforecasts and Climatology-Calibrated Precipitation Analyses*

Abstract

Abstract Analog postprocessing methods have previously been applied using precipitation reforecasts and analyses to improve probabilistic forecast skill and reliability. A modification to a previously documented analog procedure is described here that produces highly skillful, statistically reliable precipitation forecast guidance at ° grid spacing. These experimental probabilistic forecast products are available via the web in near–real time. The main changes to the previously documented analog algorithm were as follows: (i) use of a shorter duration (2002–13), but smaller grid spacing, higher-quality time series of precipitation analyses for training and forecast verification (i.e., the Climatology-Calibrated Precipitation Analysis); (ii) increased training sample size using data from 19 supplemental locations, chosen for their similar precipitation analysis climatologies and terrain characteristics; (iii) selection of analog dates for a particular grid point based on the similarity of forecast characteristics at that grid point rather than similarity in a neighborhood around that grid point; (iv) using an analog rather than a rank-analog approach; (v) varying the number of analogs used to estimate probabilities from a smaller number (50) for shorter-lead forecasts to a larger number (200) for longer-lead events; and (vi) spatial Savitzky–Golay smoothing of the probability fields. Special procedures were also applied near coasts and country boundaries to deal with data unavailability outside of the United States while smoothing. The resulting forecasts are much more skillful and reliable than raw ensemble guidance across a range of event thresholds. The forecasts are not nearly as sharp, however. The use of the supplemental locations is shown to especially improve the skill of short-term forecasts during the winter.