Bound on forecasting skill for models of North Atlantic tropical cyclone counts

TL;DR

This paper establishes a fundamental lower bound on the forecasting error for models predicting North Atlantic tropical cyclone counts, showing that simple models already achieve this bound and complexity offers no improvement.

Contribution

It introduces a lower bound on forecast accuracy for Poisson-based TC count models and demonstrates that simple linear models reach this bound, limiting the potential for more complex models.

Findings

A simple linear model explains ~50% of variance in TC counts.

More complex models do not improve forecast skill.

Observed TC counts are consistent with a Poisson process.

Abstract

Annual North Atlantic tropical cyclone (TC) counts are frequently modeled as a Poisson process with a state-dependent rate. We provide a lower bound on the forecasting error of this class of models. Remarkably we find that this bound is already saturated by a simple linear model that explains roughly 50 percent of the annual variance using three climate indices: El Ni\~no Southern Oscillation (ENSO), average sea surface temperature (SST) in the main development region (MDR) of the North Atlantic and the North Atlantic oscillation (NAO) atmospheric circulation index (Kozar et al 2012). As expected under the bound, increased model complexity does not help: we demonstrate that allowing for quadratic and interaction terms, or using an Elastic Net to forecast TC counts using global SST maps, produces no detectable increase in skill. We provide evidence that observed TC counts are consistent with a Poisson process, limiting possible improvements in TC modeling by relaxing the Poisson assumption.