This dataset consists of tropical cyclone activity forecasts for the period 15 May 2016 to 31 December 2022 from four models trained and evaluated within CLINT workpackage 3. Tropical cyclone activity is defined as the occurrence of at least one tropical cyclone within a 300-km radius and a 48-h time window, evaluated on a 2.5°-2.5° grid. While the core AI model was trained on a set of predictors taken from ERA5 reanalysis (references below), the four forecast models differ in terms of how they are applied for inference. Like the core model, the 'pureAI' model takes ERA5 predictors as input. In the hybrid versions, however, the core model was applied to operational IFS forecasts, by using 1) the raw control forecast (called 'hybridCF'), 2) a mean-bias-corrected version of the control forecast, and 3) the raw perturbed forecasts. Forecasts are provided for seven regions: Northern Indian Ocean (30°E to 100°E, 0°N to 30°N), Northwest Pacific (100°E to 170°E, 5°N to 35°N), East Pacific (170°W to 100°W, 0°N to 30°N), North Atlantic (90°W to 20°W, 10°N to 40°N), Southern Indian Ocean (20°E to 90°E, 30°S to 0°N), Australia (90°E to 160°E, 30°S to 0°N), South Pacific (160° to 230°E, 30S° to 0°N).
Forecast data is provided in csv format, with columns for index, latitude, longitude, raw predictor values, and prediction value for each lag. Lags range from 0 to 13 for the pureAI mode, and from 0 to 9 for the other. Predictor variables comprise 850-hPa relative vorticity, 700-hPa relative humidity, u and v wind-components at 200 and 850 hPa, sea surface temperature, total column vertically-integrated water vapour, total column cloud liquid water, and total column cloud ice water.