A training set is obtained by combining diagnostic predictors calculated from the ECMWF finemesh reanalysis (ECthin) fields with the shortduration heavy rain cases from 2011 to 2014 between May and September. Based on the box difference indexes of all predictors, a thresholding method is proposed to rudimentarily decrease false alarms. A new classbalanced training set is reconstructed by using the Kmeans clustering, and meanwhile, predictors with greater average weights are selected by the Relief algorithm. A forecast model for shortduration heavy rainfall in the Chongqing region centered by the Xgboost algorithm is established. The results suggest three points: (1) this model provides probabilistic and deterministic binary forecasts generated by the customized threshold; (2) the verification of the independent validation set in 2015 shows that the model achieves better classification performance in general and outperforms the ECthin hourly total precipitation reanalysis when the probability threshold is set to be 0.1, with TS score reaching 0.3. Two case studies show that this Xgboostbased model can predict the probability and area of potential shortduration heavy rain events with higher precision than the hourly ECthin, scoring TS between 0.2 to 0.4. (3) The TS scores of the Xgboost model on cases from recent years are greater than 0.1, outperforming the ECthin and rivalling daily forecast operation, which means that its products are well worth referring to.