Koopman-derived MJO index extends skillful prediction to 46 days versus 11 days for RMM

The study applies a data-driven Koopman operator approximation to reanalysis fields of circulation and convection. This yields an index whose phase evolution is far less noisy than the Real-time Multivariate MJO index while preserving the same eight-phase structure and amplitude signals. Because the Koopman modes are linear in a lifted space, the index admits stable autoregressive forecasts that maintain coherence out to 46 days before correlation drops below 0.5.

Standard RMM indices rely on empirical orthogonal functions that mix propagating and standing signals, injecting high-frequency noise that limits linear predictability. The Koopman construction filters these components explicitly, producing a lower-dimensional representation whose spectrum is dominated by the 30- to 60-day band. This methodological shift explains the tripling of forecast horizon without any change in the downstream prediction model.

Operational S2S systems still reach 35 days with the conventional RMM, indicating that the Koopman index currently serves best as a diagnostic complement rather than a replacement. Its smoother trajectory could nevertheless improve verification metrics and guide targeted model development for subseasonal tropical variability that drives global teleconnections.

Next steps include testing the index inside coupled ensemble systems and assessing whether its reduced noise improves downstream predictions of mid-latitude blocking and monsoon onset.