|
Activity Number:
|
403
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Wednesday, August 9, 2006 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statisticians in Defense and National Security
|
| Abstract - #307577 |
|
Title:
|
Local Bayesian Model Averaging for Calibrated Weather Forecast Probabilities
|
|
Author(s):
|
Eric Grimit*+
|
|
Companies:
|
University of Washington
|
|
Address:
|
Box 351640, Seattle, WA, 98055,
|
|
Keywords:
|
probabilistic weather forecast ; mesoscale ensemble ; constrained optimization ; continuous ranked probability score
|
|
Abstract:
|
Bayesian model averaging (BMA) is a standard approach to statistical inference recently used to generate calibrated probability distributions for weather variables, given an ensemble of dynamical model forecasts. In previous work, a set of BMA parameters is estimated using available observation data within the entire model domain. This leads to overestimation of the forecast uncertainty in regions where the errors are consistently smaller than average and underestimation where the errors tend to be larger. This paper presents results for BMA applied with separate parameter sets for each model grid point. The local BMA parameters are estimated with training data from a neighborhood around each point. The neighborhoods are defined through optimization of the continuous ranked probability score with constraints based on both geophysical and model characteristics.
|
