ARIMA Model Based on Weighted Bagging Algorithm and Its Application
Jinghua Huang, He Zhang · 2024
ARIMA model is a statistical method for shortterm prediction of stationary time series, and in the specific modeling process, AIC and BIC criteria are generally used to select the number of hyperparameters of the model, but the AIC and BIC criteria based on the goal of likelihood maximization may not be consistent with the purpose of predicting optimality. In this paper, an improved ARIMA model based on the weighted Bagging algorithm was proposed, in which several ARIMA models were trained by setting all feasible parameter combinations and forming a set of candidate models, and then the weighted Bagging algorithm was used to average the models of several sub-models to obtain an ensemble learning model based on ARIMA. On the one hand, the proposed method can theoretically improve the prediction performance of the original model from the perspective of prediction accuracy, and on the other hand, the method weighs the bias and variance of the prediction model through the model averaging technique. The actual data analysis results show that the proposed method has a significant improvement in prediction accuracy compared with the ARIMA model based on AIC and BIC criteria.