Global SST forecast using empirical model: causality based non-linear modeling

In the framework of a linear Vector Autoregressive (VAR) model for the global SST anomalies (Chapman et al., 2015), this study compares the forecast skill of VAR with a model using time-varying coefficients for each time level. The global SST forecast using VAR is developed using the principal components of global SST EOFs with multiple time levels as the predictors and it has been proven successful especially for 9-12 month lead forecasts for ENSO and Tropical Atlantic SST. The time-varying coefficients are obtained according to the causality amongst the principal components based on the Convergent Cross Mapping method by Sugihara et al. (2012). In this study, we take a closer look at the causality in all the possible pairs amongst the involved principal components of SST modes, and assess the potential skill improvement by introducing time-varying coefficients.