Subseasonal forecasts of the MJO and convectively coupled equatorial waves in ACCESS‐S2 and GraphCast: A multi‐year analysis

Abstract

The Madden–Julian Oscillation (MJO) and convectively coupled equatorial waves are fundamental drivers of tropical precipitation variability at subseasonal‐to‐seasonal (S2S) time‐scales, yet their accurate prediction remains challenging for S2S forecasting systems. This study evaluates the predictive skill of two contrasting approaches—the dynamical Australian Community Climate and Earth‐System Simulator–Seasonal version 2 (ACCESS‐S2) and the machine‐learning (ML)‐based GraphCast—in forecasting the MJO, equatorial Rossby (ER), Kelvin, and mixed Rossby–gravity (MRG) waves across the Tropics. Using 38 years of hindcast data (1981–2018) with 42‐day lead times, we apply wavenumber–frequency filtering techniques to precipitation and assess forecast performance against observed precipitation estimates from the Multi‐Source Weighted‐Ensemble Precipitation (MSWEP). Although the evaluation period overlaps with GraphCast's training data (1979–2021) and may favor its performance, it enables assessment across decades of climate variability and highlights the core strengths and limitations of each modeling approach. Our results reveal distinct complementary capabilities between the two models: ACCESS‐S2 demonstrates better skill for longer‐time‐scale phenomena, maintaining meaningful correlations () for MJO forecasts up to 21 days over the Maritime Continent during boreal summer (April–September) and the Indian Ocean during boreal winter (October–March) and between 9 and 16 days for ER waves across different tropical basins. In contrast, GraphCast shows better performance for shorter‐time‐scale wave prediction, achieving 6–9 days of skill, defined as , for Kelvin waves and 4–8 days for MRG waves. However, both models systematically underestimate precipitation amplitude across all wave types, particularly over the Maritime Continent and Pacific regions. Phase‐speed biases, in contrast, remain small (below 20%), suggesting that forecast degradation is driven primarily by loss of wave coherence rather than propagation errors. These findings highlight complementary capabilities that could inform hybrid ensemble forecasting strategies for improved S2S equatorial wave and tropical precipitation predictions.