Evaluation of Mjo Predictive Skill In Multi-physics and Multi-model Global Ensembles

Abstract

The Madden-Julian Oscillation (MJO) is the primary driver of intraseasonal (30-90 day) variability in the tropics, and has been linked to various weather phenomena across the globe. Therefore, it is believed that improved MJO forecasts may allow for enhanced predictive skill in both the tropics and mid-latitudes for the subseasonal range (weeks 3-4). Here, month-long hindcasts from both the Climate Forecast System version 2 (CFSv2), and the Flow-following Icosahedral Model coupled with the icosahedral Hybrid Coordinate Ocean Model (FIM-iHYCOM), are evaluated in terms of MJO performance. For each model, the hindcasts consist of 4 ensemble members run once per week over the period 1999-2010. It is found that both FIM-iHYCOM (using the Grell-Freitas deep convective scheme, hereafter “FIM-CGF”) and CFSv2 have skillful MJO forecasts out to ~18-20 days, and that their corresponding multi-model mean was skillful out to ~22 days – even when accounting for ensemble size. In contrast, running FIM-iHYCOM with SAS deep convection (hereafter “FIM-SAS”) yielded significantly worse MJO forecasts; consequently, combining FIM-SAS with CFSv2 realized little to no improvement in forecast skill. This suggests that ensemble mean forecasts can only add value when the individual ensemble members are of similar skill. Moreover, a multi-model ensemble (i.e., CFSv2 combined with FIM-iHYCOM) is preferable to a multi-physics ensemble (combining FIM-CGF and FIM-SAS): in the latter, the poor performance of FIM-SAS actually caused the multi-physics ensemble mean to have less skill and more error than FIM-CGF alone. Both multi-model and multi-physics ensembles substantially improve the spread/error relationship, although all remain under-dispersive.