A Sample Size Sensitivity Test for MOS Precipitation Type
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 7 January 2015: 4:00 PM
123 (Phoenix Convention Center - West and North Buildings)
Developing Model Output Statistics (MOS) guidance for rare events such as freezing and frozen precipitation requires a training sample sufficient to obtain stable estimates of the regression parameters. Unfortunately, long samples of training data from a stable model often are not available, which can pose significant challenges to producing accurate and reliable statistical guidance for rare events. About 29 years of reforecast data from the current generation Global Ensemble Forecast System (GEFS) was generated by NOAA/OAR (Hamill et al. 2013). The availability of this dataset provides an opportunity to assess the impact of sample size on the accuracy and skill of MOS guidance. This paper summarizes the development of MOS precipitation type guidance from GEFS reforecasts and presents verification results for varying lengths of training data.
Equations for the conditional probability of freezing, frozen, and liquid precipitation types were developed from GEFS reforecasts for training periods of 1, 2, 3, 5, 10, and 15 years. Equations using predictors derived from the ensemble mean were developed at ~550 reliable METAR sites over the CONUS and Alaska. If a generalized operator or regionalized approach is used, the results from cross-validation suggest that a training period of no less than 2 years and no more than 5 years is sufficient to develop skillful and reliable MOS guidance for precipitation type. If a longer training dataset is available (i.e., 10 years or longer) then a single-station approach for precipitation type is superior.