In order to select the indicators of bond default, to determine an effective default warning time window, and to establish a bond default warning model which is practically and has high prediction accuracy in different default states, SMOTE and XGBoost are applied to process imbalanced samples and determine the default warning model and the optimal warning time window according to the default warning accuracy and the number of indicators in the indicator system respectively. The results show that the prediction effect is better when the default warning time window is $t-1$, which means that using indicator data one year in advance can better predict whether bonds will default. Using the embedding feature selection of XGBoost algorithm to establish the default early warning model can complete the model training and indicator system dimension reduction simultaneously, which makes the compute easier. Comparing with the other 7 common default prediction methods, the proposed model has higher default prediction accuracy, more effective dimension reduction, less computing time, and stronger interpretability.