Accurate seasonal drought prediction is crucial for mitigating socio-economic and ecological losses, yet dynamical models are often evaluated on drought indices rather than integrated events, and their skill variation linked to drought mechanisms remains unclear. This study assesses SEAS51, CFSv2, and CPS3 for five extreme droughts (2018-2022) over China using a 3D event-oriented framework based on the SPAI and DBSCAN clustering. The deterministic and probabilistic prediction skills are assessed using the threat score, and the models’ ability to capture critical precursors and circulation patterns is examined. Our results indicate that deterministic drought predictions generally skillful within lead times of 45 days. Probabilistic predictions extend the skillful lead time, in some cases beyond 120 days. Model performance varies substantially across events, closely linked to their capacity to simulate key drought-driving processes, such as the weakened Walker Circulation during the 2018 South China drought and Rossby wave dynamics associated with the 2022 Yangtze River Basin drought. Moreover, we identify the limitation of predicting persistent and large-scale precipitation deficits in dynamical models, which leads to an optimal probability threshold of 25% for ensemble-based drought prediction. These findings highlight the operational value of probabilistic ensemble predictions for drought early warning and provide a mechanistic basis for understanding model skill differences, supporting the development of drought prediction systems.

Seasonal drought prediction,3D drought event identification,Dynamical model evaluation,Ensemble probabilistic prediction