The national concept of consumption has changed to excessive consumption, and overdue debts have also increased. The surge of non-performing loans will not only lead to the liquidity difficulties of banks, but also lead to financial risks. Accurate prediction of personal credit overdue is one of the key issues to control financial risks. Traditional machine learning methods build classification models according to the characteristics of credit users, while ensemble learning can ensure high accuracy and prevent model overfitting, which is the mainstream of current application research. The Stacking method can fully combine the advantages of the base model and improve the model performance. The base model and hyperparameter selection have great influence on the prediction accuracy. Therefore, parameter selection according to the studied problem is the core of application. In this paper, the Stacking method is used to integrate multiple single models for credit user overdue prediction, and the parameters of the base model are optimized. The improved Bayesian optimization method is used to select appropriate parameter combinations to improve the model performance.