Predictability of temperature extremes in multi-annual forecasts

Abstract Temperature extreme events, associated with major impacts on various socioeconomic sectors, exhibit trends related to global warming and undergo important variability across different timescales. While attention has been given to seasonal and decadal predictions, there is growing interest in exploring the potential for skillful predictions at interannual timescales. In the current study, we assess the deterministic skill of the CESM2-SMYLE in predicting temperature extremes, globally and in all calendar seasons, up to two years. This 20-member ensemble of 24-month hindcasts enables a comprehensive assessment of interannual predictability since it is initialized quarterly per year. The study evaluates the capability of the prediction system to forecast the number of days belonging to episodes of extreme temperature anomalies, considering such anomalies in all calendar seasons (DJF, MAM, JJA, SON) and studying each forecast season independently up to 2 years ahead. In general significant predictive skill is found over many regions (depending on the calendar season). As expected, the skill is higher in the first forecast season and generally decreases over time. However, notable skill persists in some regions even up to forecast-season seven. Importantly, in certain regions, significant skill remains up to approximately forecast-season four, even after removing the externally forced signal as estimated from the corresponding uninitialized historical simulations. This suggests that in certain areas, internal variability contributes to the predictability of temperature extremes even beyond the seasonal timescale. Since these areas are affected by ENSO teleconnections, we also assess the role of ENSO as a source of predictability and we find significant contributions especially for the boreal winter (DJF) and spring (MAM) up to forecast-season four. However, there is evidence that additional sources of predictability may contribute.