A Framework for Diagnosing Seasonal Prediction through Canonical Event Analysis

Abstract

ydrologic extremes in the form of flood and drought have large impacts on society that can be reduced through preparations made possible by seasonal prediction. However, the skill of seasonal predictions from global climate models is uncertain, which severely limits their practical use. In the past, the skill assessment has been limited to a single temporal or spatial resolution for a short hindcast period, which is prone to sampling errors, and noise that leads to uncertainty. In this work a framework that uses ?canonical? forecast events, or averages in space?time, to provide a more certain assessment of when and where models are skillful is developed. This framework is demonstrated by using NCEP?s Climate Forecast System, version 2, hindcast dataset for precipitation and temperature over the contiguous United States (CONUS). As part of the canonical event analyses, the probabilistic predictability metric (PPM) is used to define spatial and seasonal variability of forecast skill and its attribution to El Niño?Southern Oscillation (ENSO) over the CONUS. The PPM indicates that there are clear seasonal and spatial patterns of model skill that provide a better understanding of when and where to have confidence in model predictions as compared to a skill metric based on a single temporal and spatial scale. Furthermore, the canonical event analysis also facilitates the attribution of spatiotemporal variations of precipitation predictive skill to the antecedent ENSO conditions. This work illustrates the importance of using canonical event analysis to diagnose seasonal predictions and discusses its extensions for model development.