Does statistical model perform at par with computationally expensive general circulation model for decadal prediction?

Abstract

Decadal predictions have gained immense importance over the last decade because of their use in near-term adaption planning. Computationally expensive coupled model intercomparison project phase 5 general circulation models (GCMs) are initialized every 5 years and they generate the decadal hindcasts with moderate skill. Here we test the hypothesis that computationally inexpensive data-driven models, such as multi-variate singular spectrum analysis (MSSA), which takes care of trends and oscillations, performs similar to GCMs. We pick up one of the most complex variables having low predictability, Indian summer monsoon rainfall (ISMR) and its possible causal sea surface temperatures (SST). We find that the MSSA approach performs similar to the GCMs in simulating SSTs beyond their nonlinear limits of predictability, which is ∼12 months. These SSTs are used for decadal predictions of ISMR and show improved skills compared to the GCMs. We conclude that data-driven models are possible alternatives to computationally expensive GCMs for decadal predictions.

Keywords

• decadal predictions
• limits of predictability
• computationally inexpensive data-driven models
• Indian summer monsoon rainfall