Multilevel vector autoregressive prediction of sea surface temperature in the North Tropical Atlantic Ocean and the Caribbean Sea

摘要

We use a multilevel vector autoregressive model (VAR-L), to forecast sea surface temperature anomalies (SSTAs) in the Atlantic hurricane Main Development Region (MDR). VAR-L is a linear regression model using global SSTA data from L prior months as predictors. In hindcasts for the recent 30 years, the multilevel VAR-L outperforms a state-of-the-art dynamic forecast model, as well as the commonly used linear inverse model (LIM). The multilevel VAR-L model shows skill in 6-12 month forecasts, with its greatest skill in the months of the active hurricane season. The optimized model for the best long-range skill score in the MDR, chosen by a cross-validation procedure, has 12 time levels and 12 empirical orthogonal function modes. We investigate the optimal initial conditions for MDR SSTA prediction using a generalized singular vector decomposition of the propagation matrix. We find that the added temporal degrees of freedom for the predictands in VAR12 as compared with a LIM model, which allow the model to capture both the local wind-evaporation-SST feedback in the Tropical Atlantic and the impact on the Atlantic of an improved medium-range ENSO forecast, elevate the long-range forecast skill in the MDR.