On the seasonal predictability of the 2020 North Atlantic tropical cyclone season

Emma Lilly Levin, Mu-Ting Chien , Elizabeth Barnes, Haozhe He, Gabriel Vecchi, Wenchang Yang 

The 2020 Atlantic tropical cyclone (TC) season was unprecedented, producing a record storm count that exceeded the projected ranges of seasonal forecasts, despite their anticipation of above-normal activity. Here we assess the predictability of the extreme 2020 season from sea surface temperature (SST) forcing by examining a hierarchy of statistical, dynamical, and deep learning (DL) modeling frameworks, all constrained by observed SSTs. Although with observed SST forcing, physics-based models anticipated a relatively active season, the observed outcome lay beyond their ensemble ranges. To investigate this discrepancy, we first evaluate whether the large-scale environmental conditions of 2020 were conducive to the record-breaking activity. Although observed large-scale conditions in 2020 were generally favorable, they were not exceptionally so relative to other highly active seasons and did not indicate that a record-breaking season should have occurred. We next construct a 1,000-member ensemble using a DL model forced with observed SSTs to quantify the distribution of plausible outcomes. Using this large ensemble, the observed season emerges as a 0.5% event: highly unlikely, but not implausible. These results suggest that SST forcing provided only moderate predictive constraint, and that internal atmospheric variability could have played a role in the observed hyperactivity on top of the moderately favorable large-scale conditions. The inability of physics-based model ensembles to encompass the observed outcome does not indicate model failure or missing predictors, but reflects both limited predictability of the season from SST forcing and the inability of small ensembles to sample extreme tail risk.