Forecasting the El Niño southern oscillation: physics, bias correction and combined models

Abstract

Because of the impact of the El Niño southern oscillation (ENSO) on climate and the economy, there has been extensive research on predicting its behavior. The literature on climatic forecasting falls into two broad categories, physics and time series models, the latter encompassing both statistical methods and artificial intelligence. This study compares nonlinear regressions, physics models and a combined model in which the physics forecasts are used as inputs in a neural net. The regressions are estimated in first differences, and use lags of the sea surface temperature in the equatorial Pacific. The physics forecasts are from the Seasonal-to-Multiyear Large Ensemble (SMYLE) database, which uses the Community Earth System Model version 2 (CESM2) run at the National Center for Atmospheric Research (NCAR). The physics model is tested with and without bias correction. The bias correction uses an adjustment factor calculated from earlier simulations. The combined model uses long lags of sea surface temperature and the physics forecasts. Forecasting experiments are run over 1–24-month horizons, starting at four inception points. The errors are then sorted by lead times, and ensemble averages are taken. Although the regressions capture more of the dependence between proximate values, their accuracy falls away rapidly as the horizon extends. The accuracy of the physics models is found to fluctuate substantially over the forecast horizon. Bias correction improves at some but not all horizons. The combined model achieves the most accurate forecasts at the majority of lead times, although there are cases where it is less accurate. Despite the ambiguity of the findings, the results suggest that the most promising approach is to combine physics models with artificial intelligence techniques.