A Novel Intelligent Prediction Model for the Containerized Freight Index: A New Perspective of Adaptive Model Selection for Subseries

Abstract

The prediction of the containerized freight index has important economic and social significance. Previous research has mostly applied sub-predictors directly for integration, which cannot be optimized for different datasets. To fill this research gap and improve prediction accuracy, this study innovatively proposes a new prediction model based on adaptive model selection and multi-objective ensemble to predict the containerized freight index. The proposed model comprises the following four modules: adaptive data preprocessing, model library, adaptive model selection, and multi-objective ensemble. Specifically, an adaptive data preprocessing module is established based on a novel modal decomposition technology that can effectively reduce the impact of perturbations in historical data on the prediction model. Second, a new model library is constructed to predict the subseries, consisting of four basic predictors. Then, the adaptive model selection module is established based on Lasso feature selection to choose valid predictors for subseries. For the subseries, different predictors can produce different effects; thus, to obtain better prediction results, the weights of each predictor must be reconsidered. Therefore, a multi-objective artificial vulture optimization algorithm is introduced into the multi-objective ensemble module, which can effectively improve the accuracy and stability of the prediction model. In addition, an important discovery is that the proposed model can acquire different models, adaptively varying with different extracted data features in various datasets, and it is common for multiple models or no model to be selected for the subseries.The proposed model demonstrates superior forecasting performance in the real freight market, achieving average MAE, RMSE, MAPE, IA, and TIC values of 9.55567, 11.29675, 0.44222%, 0.99787, and 0.00268, respectively, across four datasets. These results indicate that the proposed model has excellent predictive ability and robustness.