Role of initial error growth in the extended range prediction skill of Madden-Julian Oscillation (MJO)

In the seamless forecast paradigm, it is hypothesized that the reduction in initial error in the dynamical model forecast would help to reduce forecast error in the extended range lead time up to 2–3 weeks. This hypothesis is tested in a version of an operational extended range forecast model based on National Centre for Environmental Prediction (US) Climate Forecast System version 2. Forecast skills are assessed to understand the role of initial errors on the prediction skill for Madden-Julian Oscillation (MJO). Retrospective forecasts are categorized in two groups. One group defines the lowest initial day’s error and the other with the highest initial day’s errors. Then, the error growth for these two categories is analyzed for the strong MJO events during May to September. The initial errors of MJO forecast are categorized and defined using the multivariate MJO index introduced by Wheeler and Hendon (Mon Wea Rev 132(8):1917–1932, 2004). The probability distribution of bivariate root mean square error (BVRMSE) and error growth evolution is used as metrics. The results showed that the initial error does not show any significant difference in the amplitude after a lead time of 7–10 days, and the error growth remains the same for both sets of runs. It is also found that the errors originate from the events with the initial phase in the western Pacific and the Indian Ocean. The study advocates the importance of better representation of MJO phases over the ocean in the model to improve the MJO prediction rather than simply focusing on improving the initial conditions.