Austral summer MJO forecast skill in S2S models: decadal shifts and their drivers

The Madden–Julian Oscillation (MJO) is a key driver of global subseasonal-to-seasonal (S2S) climate variability, initiating teleconnections that affect weather patterns worldwide. Improving understanding of the factors that modulate MJO predictability is therefore critical for advancing S2S forecasting systems. Using a multi-model framework, we evaluate changes in MJO prediction skill between two periods (1981–1998 and 1999–2018) during austral summer (December–February) and examine the processes underpinning these differences. Our analysis reveals a pronounced decadal variation in MJO forecast skill, with high-skill years in 1981–1998 showing prediction lead times of around 10 d longer (based on the bivariate correlation of the Real-Time Multivariate MJO (RMM) index) than in 1999–2018, while low-skill years show little change. This asymmetric reduction coincides with stronger MJO amplitude in the earlier period, despite relatively stable model mean-state biases in tropical sea surface temperatures (SSTs) and lower-tropospheric moisture. Key findings include: (1) persistent moisture biases across both periods, yet higher skill in 1981–1998, suggesting that model systematic errors alone cannot explain the differences; (2) a stronger relationship between Quasi-Biennial Oscillation (QBO) and MJO forecast skill in the first period, independent of stratospheric resolution in the models; and (3) weakened coupling between the MJO and large-scale climate modes, including the QBO, El Niño–Southern Oscillation (ENSO), and Indian Ocean Dipole (IOD), in 1999–2018, indicating reduced dynamical support for prediction. These results suggest that decadal variations in MJO forecast skill are strongly influenced by changes in the background dynamical environment.