Comparative Skill of Two Analog Seasonal Temperature Prediction Systems: Objective Selection of Predictors

View More View Less
  • 1 Climate Analysis Center, NMC/NWS/NOAA, Washington, D.C
  • 2 Institute of Global Climate and Ecology, Russian Academy of Sciences, Moscow, Russia
© Get Permissions
Full access

Abstract

Analog prediction systems developed in the United States and the former Soviet Union are compared for U.S. seasonal temperature prediction. Of primary interest is the viability of the Russian “optimization” concept for a priori selection of U.S. seasonal analog forecast predictors. Optimization is a specific technique for choosing predictor variables for analog matching on a forecast-by-forecast basis. Validation of this procedure would lead to more efficient design of analog prediction models and the elimination of some subjectivity in the process that inevitably results in overstatements in realizable skill. The procedure's effectiveness was tested using predictor and predictand datasets from the U.S. system in a cross-validation framework. Skills of different models were assessed on the basis of 40 seasonal forecasts at 92 U.S. stations.

The Russian system (called GRAN for “Group Analog”) was first run without optimization using the a posteriors selected predictors used in the U.S. system. A version of the U.S. system (without use of antianalogs) that is conceptually very similar to GRAN without optimization was run for comparison in this calibration step. The results reveal that these systems perform in a nearly identical manner when predictor and predictand datasets are the same. Next GRAN forecasts were made using all available predictors and then using only predictors selected via optimization. The results not only show that objective a priori predictor selection by optimization is just as effective (in terms of skill) as subjective a posteriors selection but also suggest it may produce superior results in summer forecasts.

Abstract

Analog prediction systems developed in the United States and the former Soviet Union are compared for U.S. seasonal temperature prediction. Of primary interest is the viability of the Russian “optimization” concept for a priori selection of U.S. seasonal analog forecast predictors. Optimization is a specific technique for choosing predictor variables for analog matching on a forecast-by-forecast basis. Validation of this procedure would lead to more efficient design of analog prediction models and the elimination of some subjectivity in the process that inevitably results in overstatements in realizable skill. The procedure's effectiveness was tested using predictor and predictand datasets from the U.S. system in a cross-validation framework. Skills of different models were assessed on the basis of 40 seasonal forecasts at 92 U.S. stations.

The Russian system (called GRAN for “Group Analog”) was first run without optimization using the a posteriors selected predictors used in the U.S. system. A version of the U.S. system (without use of antianalogs) that is conceptually very similar to GRAN without optimization was run for comparison in this calibration step. The results reveal that these systems perform in a nearly identical manner when predictor and predictand datasets are the same. Next GRAN forecasts were made using all available predictors and then using only predictors selected via optimization. The results not only show that objective a priori predictor selection by optimization is just as effective (in terms of skill) as subjective a posteriors selection but also suggest it may produce superior results in summer forecasts.

Save