A total of 184 hail weather cases in Chengde Mountains from April to September in 2000-2020 are analyzed by means of multi-source data including the hail observation, CINRAD/CB weather radar data, NCEP FNL reanalysis data, and NCEP-GFS forecasts. Firstly, the distribution characteristics and forecast thresholds of relevant ambient parameters such as water vapor, thermal instability, dynamic lift and characteristic height are analyzed in the form of box plots. Then, the initial optimal threshold values are set according to the results of box plots. The hail labels are determined according to the hail observation records or composite reflectivity greater than or equal to 60 dBz from April to September in 2014-2020. The hail labels are matched to the grids of reanalysis data according to the principle of near location and proximity time to construct the positive and negative sample dataset for feature parameter selection, interval segmentation and probability calculation. Next, five models for 3, 6, 9, 12 and 24 h hail potential forecast are established by the Bayesian method. The models are tested focusing on the weather processess from June to August during 2021-2022. The results suggest that the Bayesian-based hail potential prediction models have a certain feasibility in daily weather forecasting. The hit rates of all the models are above 90%, and the average critical success index is over 40.3%. Differing from the traditional probability and ingredient methods, the method can provide a better objective forecast of hail occurrence, which has a certain reference value for forecasting severe convective weather in mountainous areas. However, there are some false alarms as the spatio-temporal scale of the reanalysis data is much larger than that of severe convective weather, which needs to be improved in the future.