Long‐lead ENSO predictability from CMIP5 decadal hindcasts

The limits for the predictability of the El Niño–Southern Oscillation (ENSO) have been long discussed. Even when ENSO prediction skill is expected to be limited, questions remain as to which are the controlling factors. The role of atmospheric noise for ENSO initiation, the growth of initial errors and inadequate models have been identified as key elements (e.g., Chen et al. 2004; Chen and Cane 2008; Jin et al. 2008; Guilyardi et al. 2009). Past studies have used retrospective forecasts from seasonal prediction systems to evaluate the predictability of ENSO for up to 24 months (e.g., Chen et al. 2004; Chen and Cane 2008; Ludescher et al. 2014). More recently, Wittenberg et al. (2014) have shown, using a 4000-years control run and a set of reforecasts from the model GFDL CM2.1, that in this setting, free from external forcings, the potential predictability is lost after the 3–4 year range. For several decades it has been recognised that ensemble prediction is fundamental for an adequate representation of the probabilistic nature of forecast information (e.g., Tippett and Barnston 2008 and references therein). More recently, it has been acknowledged that multi-model ensembles provide a more accurate representation of the forecast uncertainty than single model ensembles, resulting in more skillful prediction systems (e.g., Tippett and Barnston 2008; Kirtman et al. 2014). Abstract Using decadal prediction experiments from the WCRP/CMIP5 suite that were initialized every year from 1960-onward, we explore long-lead predictability of ENSO events. Both deterministic and probabilistic skill metrics are used to assess the ability of these decadal prediction systems to reproduce ENSO variability as represented by the NINO3.4 index (EN3.4). Several individual systems as well as the multi-model mean can predict ENSO events 3–4 years in advance, though not for every event during the hindcast period. This long-lead skill is beyond the previously documented predictability limits of initialized prediction systems. As part of the analysis, skill in reproducing the annual cycle of EN3.4, and the annual cycle of its interannual variability is examined. Most of the prediction systems reproduce the seasonal cycle of EN3.4, but are less able to capture the timing and magnitude of the variability. However, for the prediction systems used here, the fidelity of annual cycle characteristics does not appear to be related to the system’s ability to predict ENSO events. In addition, the performance of the multi-model ensemble mean is explored and compared to the multi-model mean based solely on the most skillful systems; the latter is found to yield better results for the deterministic metrics. Finally, an analysis of the near-surface temperature and precipitation teleconnections reveals that the ability of the systems to detect ENSO events far in