Large climate model ensembles reveal underdispersion in seasonal Atlantic tropical cyclone counts

Emma Lilly Levin, Gabriel Vecchi, Gabriele Villarini

Seasonal Atlantic tropical cyclone (TC) counts are commonly modeled as a conditional Poisson process, implying that the distribution of possible seasonal outcomes—the range of TC counts that could plausibly occur in a given year—exhibits equidispersion for a given climate state, with its variance equal to its mean. This assumption underlies many statistical frameworks used for seasonal TC prediction and risk assessment, yet to the best of our knowledge has not been extensively tested directly. Using large ensembles from physics-based climate models (AM2.5-C360 and HiRAM) and a deep learning-based climate emulator (ACE2; Ai2 Climate Emulator version 2), we examine the distributional properties of seasonal TC counts in a controlled modeling framework. Across all models and years, we find that the ensemble distribution of within-season TC counts is systematically underdispersed, with ensemble variances smaller than their corresponding means. This behavior violates the equidispersion implied by a Poisson process and is consistent with a finite opportunity or binomial framework. This interpretation of TC genesis suggests that storms arise from a limited number of precursor disturbances whose likelihood of development depends on large-scale environmental conditions. We fit seasonal TC count distributions from a 1,000-member ACE2 ensemble to the Poisson and binomial distributions and find statistical evidence that the binomial formulation provides a better fit to TC count distributions. The Poisson assumption is not supported by current climate model simulations, given that simulated seasonal TC counts are more constrained than implied by traditional Poisson-based frameworks.