Compared with departures, predicting the weather impact on arrival delays is more challenging because of possible non-linear, cascading effects, and higher uncertainty. Existing weather impact studies are location-dependent and often neglect the impacts of dangerous phenomena. We propose a data-driven model for severe weather impact quantification on airport arrival on-time performance based on the Bayesian approach to address these issues. Our model considers the impact of the dangerous phenomenon by evaluating the mean shift and is flexible enough to be applied to different airports. Using two years’ worth of data (2017-2018) from the Hong Kong International Airport, we studied over 55,000 local meteorological reports and analyzed over 430,000 arrival flights. Across all three key performance metrics considered, a non-linear relationship with the weather score, akin to a phase transition, could be observed. This framework allows a comparison between the sensitivity of each airport’s arrival performance metric towards severe weather. Delay rate is the most sensitive metric, while cancellation rate is the least. For the impacts of dangerous phenomena, cumulonimbus has the most significant impact on the delay rate. Shower rainfall/cumulonimbus has a similar and vital impact on the mean arrival delay per hour. Because of its potential applications in different airports, this framework can provide a deeper insight into weather impact on air traffic networks.