Random analog prediction provides a benchmark for seasonal Arctic sea ice extent forecasting

Abstract

Abstract We develop a random analog predictor (RAP) algorithm for the forecasting of Arctic sea ice extent (SIE) on seasonal timescales. This is a stochastic variant for scalar time series of the method of the analogues that only uses the historical SIE record to produce ensemble forecasts. When comparing the observations with the most representative forecast of the ensemble (as identified through the band–depth, a centrality measure for functional data) the algorithm shows negligible bias and RMSE no larger than $$0.6\cdot 10^6$$ km $$^2$$ . Hindcasts of the September average SIE show a level of skill comparable to that of the models of the Sea Ice Prediction Network. We argue that simplicity, interpretability, independence on physical hypotheses, and the ability to attach an uncertainty estimate to its own forecasts, should make RAP a necessary baseline for physics–based and AI–based models alike: failure to outperform RAP should be considered as a serious shortcoming for models aiming at realistically describing sea ice dynamics.