A Comparison between Bayesian Model Averaging and Heteroscedastic Censored Logistic Regression Using 2012 GEFS Precipitation Reforecasts over the US Middle-Atlantic Region

Open Access
Yang, Xingchen
Graduate Program:
Civil Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
March 30, 2016
Committee Members:
  • Alfonso Ignacio Mejia, Thesis Advisor
  • Steven J Greybush, Thesis Advisor
  • Xiaofeng Liu, Thesis Advisor
  • ensemble forecast
  • postprocessing
The potential of Bayesian model averaging (BMA) and heteroscedastic censored logistic regression (HCLR) to postprocess precipitation ensembles from the National Oceanic and Atmospheric Administration’s (NOAA’s) National Centers for Environmental Prediction (NCEP) 11-member Global Ensemble Forecast System Reforecast version 2 (GEFSRv2) dataset is investigated. To implement BMA, we select two different modeling scenarios: exchangeable and non-exchangeable members. We term the BMA postprocessing with exchangeable members BMAx. To compare the postprocessors, as part of our experimental setting, we use 24-h precipitation accumulations and lead times of 24- to 120-h. As the study area, we select the middle Atlantic region (MAR) of the United States (US). In contrast with previous postprocessing studies, we consider here a wider range of forecasting conditions (e.g., the effect of spatial pooling, training length, lead time, precipitation threshold, and seasonality) when evaluating BMA and HCLR. Additionally, BMA and HCLR have not yet been compared against each under a common and consistent experimental setting. To implement BMA and BMAx, we use a sliding window of 25 days and train each GEFSRv2 cell separately, as opposed to using spatial pooling. These training conditions were selected by carefully examining the skill of forecasts associated with different window lengths and number of cells. To compare and verify the postprocessors, we use the BSS, CRPSS, and reliability diagrams, conditioned upon the forecast lead time, precipitation threshold, and season. Overall, we find that HCLR tends to outperform BMA/BMAx but the differences among the postprocessors are not as significant. Also, BMA and BMAx behave similarly across lead times and seasons, iv thereby indicating that the GEFSRv2 members remain indistinguishable across lead times. The improved performance of HCLR over that of BMA seems related to the ability of HCLR to include the ensemble variance as a predictor. In the future, an alternative approach could be to combine HCLR with BMA to take advantage of their relative strengths.