In this paper, four types of machine learning techniques, namely decision trees, random forests, extreme gradient boosting, and ridge regression, are employed to predict the excess return of ‘tushare’ decadal data at monthly frequencies. The importance of each machine learning parameter is ranked, and the training order is prioritized accordingly. For the tree model, this paper gives priority to the criteria for tree splitting strategy, followed by the number of trees, tree depth, and leaf nodes. In the case of ridge regression, the only parameter considered is the regularization parameter. The rolling training method is utilized to conduct time series cross-validation for tuning the models one by one. After the tuning process, the models are used for rolling predictions. Finally, the backtesting results are employed to assess the predictive performance of each model. It is concluded that the ridge regression model achieves the highest annualized return of 22.22%. The decision tree, random forest, and extreme gradient boosting models yield annualized returns of 8.63%, 16.6%, and 15.71% respectively.