Predictability and prediction skill of the boreal summer intraseasonal oscillation in the Intraseasonal Variability Hindcast Experiment

Boreal summer intraseasonal oscillation (BSISO) is one of the dominant modes of intraseasonal variability of the tropical climate system, which has fundamental impacts on regional summer monsoons, tropical storms, and extra-tropical climate variations. Due to its distinctive characteristics, a specific metric for characterizing observed BSISO evolution and assessing numerical models’ simulations has previously been proposed (Lee et al. in Clim Dyn 40:493–509, 2013). However, the current dynamical model’s prediction skill and predictability have not been investigated in a multi-model framework. Using six coupled models in the Intraseasonal Variability Hindcast Experiment project, the predictability estimates and prediction skill of BSISO are examined. The BSISO predictability is estimated by the forecast lead day when mean forecast error becomes as large as the mean signal under the perfect model assumption. Applying the signal-to-error ratio method and using ensemble-mean approach, we found that the multi-model mean BSISO predictability estimate and prediction skill with strong initial amplitude (about 10 % higher than the mean initial amplitude) are about 45 and 22 days, respectively, which are comparable with the corresponding counterparts for Madden–Julian Oscillation during boreal winter (Neena et al. in J Clim 27:4531–4543, 2014a). The significantly lower BSISO prediction skill compared with its predictability indicates considerable room for improvement of the dynamical BSISO prediction. The estimated predictability limit is independent on its initial amplitude, but the models’ prediction skills for strong initial amplitude is 6 days higher than the corresponding skill with the weak initial condition (about 15 % less than mean initial amplitude), suggesting the importance of using accurate initial conditions. The BSISO predictability and prediction skill are phase and season-dependent, but the degree of dependency varies with the models. It is important to note that the estimation of prediction skill depends on the methods that generate initial ensembles. Our analysis indicates that a better dispersion of ensemble members can considerably improve the ensemble mean prediction skills.