MJO Prediction Skill Using IITM Extended Range Prediction System and Comparison with ECMWF S2S

Eastward propagating Madden–Julian Oscillation (MJO) is a dominant mode of the intraseasonal variability and hence a potential source of intraseasonal predictability. Therefore, advancing MJO prediction using state-of-the-art dynamical model is of utmost importance for improving intraseasonal prediction. The prediction skill and predictability of MJO are assessed using 44 members ensemble hindcast (16 years data; 2001–2016) of CFSv2 based extended range prediction (ERP) system of IITM as well as 10 member ensemble hindcast (16 years data; 2001–2016) of ECMWF S2S dataset. The MJO is diagnosed using a newly developed Extended Empirical Orthogonal Function (EEOF) analysis. Near equatorial (15° S–15° N) model anomaly fields are projected onto the leading pair of observed eigen modes. The leading pair of observed eigen modes are obtained based on the EEOF analysis of the combined field of zonal wind at 200 hPa (U200), zonal wind at 850 hPa (U850) and velocity potential at 200 hPa (chi200). Model forecasted principal components (PCs) are quantitatively compared with observed PCs using bivariate correlation coefficient and root mean square error (RMSE). We find that MJO could be predicted up to around 22 days (around 31 days) for IITM ERP system (ECMWF S2S dataset) as measured by anomaly correlation coefficient remains larger than 0.5 and RMSE remains lower than 1.4. This prediction skill is quite low compared to potential predictability, which is estimated as more than 40 days both for IITM-ERP and ECMWF system. MJO prediction skill varies with initial MJO phase, particularly at the longer lead. This variation is more significant for the ECMWF system. Model (both for IITM-ERP and ECMWF) predicted amplitude drops at a faster rate and phase propagation speed for almost all initial phase is slower and amplitude is weaker compared to the observation. It could be concluded that even the state-of-the-art models [IITM-ERP (basically NCEP CFSv2) and ECMWF] are also not free from systematic errors/biases. Hence, there is an enormous space for improving MJO prediction skill by reducing these errors/biases in the dynamical model and error in the initial condition.