Extended predictability skill of ENSO using a multi-model dataset with an ensemble model output statistical approach

Predictability skill of El Niño Southern Oscillation (ENSO) have been extensively studied. A breakthrough was reached in the 1980s with the stunning 1-year lead predictability skill using dynamical forecast for limited tropical Pacific regions. Recently, Ding et al. (2018) had evaluated the predictability skill in the tropical Pacific Ocean with the North American Multi-Model Ensemble (NMME) and the Coupled Model Intercomparison Project 5 (CMIP5) datasets. Their results show a modest predictability skill improvement, which reveals the complexity of the problem, even using a computationally expensive multi model-ensemble approach framework. Recently, Newman and Sardeshmukh (2017) show that comparable results can be obtained with a less computational expensive approach, using a linear inverse model (LIM), but they treated the NMME model-ensemble average as the best of the NMME product, which is the standard metric reported by other researchers. Using an Ensemble Model Output Post-processing (EMOS), we showed that NMME is of the limited use without a proper post-processing approach. We hypothesize that EMOS outperforms both LIM and NMME ensemble average, which would happen because of explicit use of the ensemble spread and multi-model non-exchangeable approach, something that is not properly consider by previous studies (e.g., Ding et al., 2018). If EMOS, as shown here, outperform standard metrics, the approach can be used to explore what are the potential sources of the climate variability that limit predictability skill of ENSO at a global scale.