Abstract – Tropical cyclones pose increasing risks to coastal communities, making seasonal prediction a critical scientific and societal challenge. This presentation shows the performance of four machine learning approaches: Lasso regression, K-nearest neighbors, artificial neural networks, and XGBoost, in forecasting tropical cyclone activity across the Atlantic basin, Caribbean Sea, and Gulf of Mexico. Using HURDAT records (1950–2024) and a set of physically motivated climate predictors from NOAA and NASA, models are tested under operational conditions with a sliding-window cross-validation framework. Results show that model performance varies significantly: Lasso provides robust and interpretable predictions, while XGBoost delivers the most consistent overall skill, especially for basin-wide storm counts. In contrast, ANN and KNN offer limited improvements due to data constraints. Forecast skill declines for regional and major hurricane targets, highlighting fundamental predictability limits. These findings emphasize that model choice and data limitations are critical in applying AI to hurricane prediction.