Comparing Statistical and Machine-Learning Models for Seasonal Prediction of Atlantic Hurricane Activity

Tropical cyclones pose major risks to life and property, especially as coastal populations grow and climate change increases the likelihood of intense storms, making seasonal prediction of tropical cyclones an important scientific and societal goal. This study uses HURDAT best-track records from 1950–2024 to quantify annual tropical cyclone, hurricane, and major hurricane counts across the Atlantic basin, Caribbean Sea, and Gulf of Mexico. These nine targets are paired with 34 monthly climate predictors from NOAA and NASA GISS—including SST and ENSO indices, Main Development Region (MDR) wind and pressure fields, and latent heat flux empirical orthogonal functions—evaluated under nine predictor-set configurations. Four forecasting approaches are developed and tested under operationally realistic conditions: Lasso regression, K-nearest neighbors (KNN), an artificial neural network (ANN), and XGBoost, using a 30-year sliding-window cross-validation design and a Poisson log-likelihood skill score relative to climatology. Lasso performs reliably with concise, physically interpretable predictors, while XGBoost provides the most consistent overall skill, particularly for basin-wide total cyclone and hurricane counts. The skill of ANN is limited by small sample sizes, and KNN offers only marginal improvements. Forecast skill is the highest for basin-wide storm totals and decreases for regional and major-hurricane targets due to lower event frequencies and stronger predictability limits.