Diagnosis of skill variability as a basis for discriminating use of CPC long-lead seasonal forecasts

In 1990 the predominant view was that Climate Prediction Center seasonal forecasts were not skillful enough to be of economic benefit. At that time the second author argued in BAMS that there were known circumstances for which these forecasts could be expected to have appreciable skill and that some users could use this knowledge to their considerable advantage. Today the prevailing view seems to be quite different, namely that even marginally skillful sets of these forecasts have economic benefit, but may be equally erroneous. Again we argue that some subsets of these forecasts will have considerable skill and that some informed, discriminating users will be able to use CPC's products effectively. The information these users need to discriminate, i.e. when to use or not use a forecast, consists of the performance characteristics, including skill, of the forecasts for those parameters, places, and times of year of interest to them. The presentations of skill information on CPC's web site are uninformative in this context. In particular, all locations are combined for each lead time in bulk measures and presented as time series and grand averages of the bulk measure for the lead, with no discrimination by season, situation, or location. In this form the skill analysis is practically useless to the customer.

We have recombined the skills in ways that will be more informative to not only potential customers of the forecasts, but CPC as well. We first display the bulk skills by lead time (which in itself turns out to be informative) and then provide the same graphs stratified by time of year, major ENSO episodes, and both. Where possible we combine consecutive leads with similar skills to achieve more confident estimates. For those cases where consequential skill is identified we examine the geographical distribution of skill. We will demonstrate that a substantial amount of useful skill information can be extracted from the 9-year CPC long-lead forecast history. Some of this information should be considered by CPC as a basis for possible modifications to forecast practices. We are confident even more valuable information for potential users can be extracted by consideration of other performance measures and attributes, including those of the corresponding probability forecasts.