Current status of ENSO prediction skill in coupled ocean–atmosphere models

The overall skill of ENSO prediction in retrospective forecasts made with ten different coupled GCMs is investigated. The coupled GCM datasets of the APCC/CliPAS and DEMETER projects are used for four seasons in the common 22 years from 1980 to 2001. As a baseline, a dynamic-statistical SST forecast and persistence are compared. Our study focuses on the tropical Pacific SST, especially by analyzing the NINO34 index. In coupled models, the accuracy of the simulated variability is related to the accuracy of the simulated mean state. Almost all models have problems in simulating the mean and mean annual cycle of SST, in spite of the positive influence of realistic initial conditions. As a result, the simulation of the interannual SST variability is also far from perfect in most coupled models. With increasing lead time, this discrepancy gets worse. As one measure of forecast skill, the tier-1 multi-model ensemble (MME) forecasts of NINO3.4 SST have an anomaly correlation coefficient of 0.86 at the month 6. This is higher than that of any individual model as well as both forecasts based on persistence and those made with the dynamic-statistical model. The forecast skill of individual models and the MME depends strongly on season, ENSO phase, and ENSO intensity. A stronger El Niño is better predicted. The growth phases of both the warm and cold events are better predicted than the corresponding decaying phases. ENSO-neutral periods are far worse predicted than warm or cold events. The skill of forecasts that start in February or May drops faster than that of forecasts that start in August or November. This behavior, often termed the spring predictability barrier, is in part because predictions starting from February or May contain more events in the decaying phase of ENSO.