Title:Entropic learning enables skilful forecasts of ENSO phase at up to two years lead time

Abstract:This paper builds on previous work in applying the entropy-optimal Sparse Probabilistic Approximation (eSPA) machine learning algorithm to the task of classifying the phase of ENSO, as determined by the Niño3.4 index (Groom et al., Artif. Intell. Earth Syst., 2024). In this study, only observations and reanalyses from the satellite era are employed when training and validating the entropic learning models, and a full set of hindcasts are performed over the period from 2012 to 2022 (with a maximum lead time of 24 months) in order to determine out-of-sample skill. The features used for prediction are the leading principal components from a delay-embedded EOF analysis of global sea surface temperature, the vertical derivative of temperature at the equator in the tropical Pacific (a proxy for thermocline variability) and the zonal and meridional wind stresses in the tropical Pacific. Despite the limited number of data instances available for training (ranging from 350 monthly averages for the earliest hindcast to 520 as of December 2024), eSPA is shown to avoid overfitting in this small data regime and produces probabilistic forecasts with comparable skill to the combined (model-based) probabilistic forecasts produced from the International Research Institute for Climate and Society (IRI) ENSO prediction plume. At lead times longer than those available from the IRI plume, eSPA maintains skill out to 23 months in terms of the ranked probability skill score and 24 months in terms of the area under the ROC curve, all at a small fraction of the computational cost of running a dynamical ensemble prediction system. Furthermore, eSPA is shown to successfully forecast the 2015/16 and 2018/19 El Niño events at 24 months lead time, the 2016/17, 2017/18 and 2020/21 La Niña events at 24 months lead time and the 2021/22 and 2022/23 La Niña events at 12 and 9 months lead time.