Tropical cyclone activity prediction on subseasonal time-scales

We will first examine the skill of probabilistic tropical cyclone (TC) occurrence and intensity (ACE - accumulated cyclone energy) predictions in the Subseasonal to Seasonal (S2S) dataset. We show that some of the models in the S2S dataset have skill in predicting TC occurrence 4 weeks in advance. In contrast, only one of the models (ECMWF) has skill in predicting the anomaly of TC occurrence from the seasonal climatology beyond week 1. For models with significant mean biases, calibrating the forecast can significantly improve the models’ prediction skill. In contrast, for models with small mean biases, calibration does not guarantee an improvement in model skill as measured by the Brier Skill Score. We show that limiting factors for model skill are the errors in genesis prediction.

We then focus only on the ECMWF model and using cluster analysis examine the sensitivity of the North Atlantic TC tracks biases to various factors, such as model resolution, lead time, and tracking. We also explore how well the ECMWF North Atlantic TC model tracks in each cluster simulate the known response to climate modes, such as ENSO and MJO. By applying simple bias corrections to each cluster of Atlantic TC tracks, we examine if we can improve the model skill in landfall prediction in the US and Caribbean.