Rank Histograms and Correlations
Caren Marzban, University of Washington, Seattle, WA; and R. Wang, F. Kong, and S. Leyton
The rank histogram has become a standard tool for the verification of ensemble forecasts. However, by virtue of being a summary measure it can be misleading. For example, it is well-known that reliable forecasts can lead to a non-flat rank histogram, and unreliable ensembles can lead to flat rank histograms. Here, it is shown that the rank histogram is also affected by two types of correlations - temporal, and between ensemble members. It is shown how these correlations can be filtered out, thereby rendering the rank histogram itself more reliable. The method is illustrated on temperature and wind speed reforecasts for 90 stations across the US. It is found that even after controlling for correlations the forecasts are still generally unreliable. The spatial pattern of the rank histograms is examined for the purpose of "explaining" the results.
Supplementary URL: http://faculty.washington.edu/marzban/rankhist.pdf
Session 5, Probability forecasting
Wednesday, 20 January 2010, 1:30 PM-2:30 PM, B305
Browse or search entire meeting
AMS Home Page