A Vector Autoregressive ENSO Prediction Model

Abstract

Abstract The authors investigate a sea surface temperature anomaly (SSTA)-only vector autoregressive (VAR) model for prediction of El Niño–Southern Oscillation (ENSO). VAR generalizes the linear inverse method (LIM) framework to incorporate an extended state vector including many months of recent prior SSTA in addition to the present state. An SSTA-only VAR model implicitly captures subsurface forcing observable in the LIM residual as red noise. Optimal skill is achieved using a state vector of order 14–17 months in an exhaustive 120-yr cross-validated hindcast assessment. It is found that VAR outperforms LIM, increasing forecast skill by 3 months, in a 30-yr retrospective forecast experiment.