A methodology for assessing ensemble experiments

Abstract

Climate simulations and forecast experiments of increasingly large ensemble size are being performed to assess the predictive skill of a dynamic model on seasonal and longer timescales. Especially in the cases of ensemble climate simulation or forecast forced by observed or predicted sea surface temperatures, the model is expected to maximize potential predictability due to boundary forcing and to minimize internal variability generated from dynamic instability. In the light of small predictive skill in extratropics from boundary forcing, one must evaluate skill of the ensemble mean quantity against intersample variability or spread of the individual ensemble member. On the other hand, certain dominant signals in climate variability, such as E1 Niño‐Southern Oscillation, have been documented. Predictability for these major signals is the hope of seasonal and climate forecasting using a dynamic model. It may be unrealistic to anticipate a model being able to simulate or forecast the full spectra of climate variability. The question is how to evaluate a model's performance in capturing the dominant climate signals in ensemble experiments with increasingly large sample size. These issues have motivated us to develop a compact methodology for assessing climate experiments with large ensemble size. This method treats the ensemble mean as signal and intersample variability as spread or noise in a common framework. Hence not only dominant signals from boundary forcing can be isolated, but also sensitivity of these signals to the forcing can be assessed. Other potential applications of the method to climate simulation and forecasting are also discussed.