Assessment of prediction skill of intraseasonal variation from dynamical, statistical, and combined models

We consider intraseasonal variation (ISV) prediction by statistical and dynamical models. For the fair comparison, the real-time multivariate Madden-Julian Oscillation (MJO) (RMM) index for the boreal winter is used as a predictand. The statistical prediction results are compared by reassessing the multi linear regression (MLR), wavelet, and singular spectrum analysis (SSA) model. The correlation score for RMM1 (RMM2) falls away to 0.5 between 16-17 (15-16) days for MLR, 7-8 (9-10) days for wavelet, and 8-9 (9-10) days for SSA model. As both wavelet and SSA model have a discontinuity at the boundary of data, the skill of the real-time forecast shows a steep decrease at the beginning of the forecasts. To examine the skill of dynamical prediction, serial integration is performed with Seoul National University AGCM and CGCM over the entire boreal winter period. The ocean-atmosphere coupling acts to improve the simulation ability of MJO variability, the eastward propagation, and the phase relationship between convection and SST. The skill score of RMM1 (RMM2) falls out to 0.5 at 18-19 (22-23), 15-16 (17-18), and 16-17 (15-16) for CGCM, AGCM, and MLR. This result demonstrates that dynamical prediction does not lag statistical prediction in skill and is even better when ocean-atmosphere coupling is included. The dependency of prediction skill on the initial phase and amplitude of the MJO is investigated. The score is better when the MJO is initialized during an active period than during a quiescent period for both systems. Based on different characteristics of prediction skill for each phase and amplitude and for individual models, predictions are combined using available information extracted using the better of the two predictions. By simple selection, the prediction skill is clearly improved in strong MJO cases. Using another combination process based on Bayesian concepts, two independent predictions are combined by minimizing the forecast error that is known from historical information. It shows the superior to both of the predictions over the entire forecast lead days.