Empirical Extended-Range Prediction of the Madden–Julian Oscillation

Abstract

An empirical model that predicts the evolution of the Madden?Julian oscillation (MJO) in outgoing longwave radiation (OLR) and 200-mb streamfunction is developed. The model is based on the assumption that the MJO can be well represented by a pair of empirical orthogonal functions (EOFs) of OLR and three EOFs of streamfunction. With an eye toward using this model in real time, these EOFs are determined with data only subjected to filtering that can be applied in near?real time. Stepwise lag regression is used to develop the model on 11 winters of dependent data. The predictands are the leading two principal components (PCs) of OLR and the leading three PCs of streamfunction. The model is validated with five winters of independent data and is also compared to dynamic extended range forecasts (DERFs) made with the National Centers for Environmental Prediction?s Medium Range Forecast (MRF) model. Skillful forecasts of the MJO in OLR and streamfunction with the empirical model are achieved out to about 15 days. Initial skill arises from autocorrelation of the PCs. Subsequent skill beyond about 1 week arises primarily from the cross correlation with the other PCs that define the MJO. Inclusion of PCs not associated with the MJO as predictors appears not to reliably improve skill. Skill is found to be substantially better when the MJO is active at the initial condition than when it is inactive. The empirical forecasts are also found to be more skillful than DERF from the MRF for lead times longer than about 1 week. Furthermore, skill of DERF from the MRF is found to be better when the MJO is quiescent than when it is active at the initial condition. It is suggested that significant improvement of tropical DERF could be achieved by improvement of the representation of the MJO in the dynamic forecast model.