A novel forecasting method for time series based on vector visibility graphs

Time series forecasting has consistently been recognized as a field of considerable importance. In this study, we propose a novel forecasting framework that integrates phase space reconstruction, vector visibility graphs, and linear regression to enhance the accuracy of time series predictions. The proposed model leverages the strengths of complex network theory and machine learning to improve the robustness of forecasting. Through comprehensive experiments conducted on nine real-world datasets, including stock indices, gold prices, power grid loads, and traffic flows, the model demonstrates superior performance over machine learning models and existing visibility graph-based forecasting methods. The results indicate significant improvements in prediction accuracy, as evidenced by lower values of MAPE, MASE, and NMSE across various datasets. The Model Confidence Set (MCS) test further corroborates the statistical significance of these improvements, affirming the model’s effectiveness in a variety of forecasting scenarios. Furthermore, this study evaluates the impact of node similarity measures based on the communicability matrix and local relative entropy on the prediction of the results, showing that the communicability matrix method improves the model performance.