Forecasting Time Series through Motif Structure Analysis within Visibility Graph

Time series forecasting is a fundamental and crucial problem in both academic research and industrial production. Various approaches have been developed to address this issue from different perspectives, each with its own limitations. In this work, we propose a motif-based approach for forecasting time series within the framework of visibility graph analysis. By representing time series interactions through a complex network, the most similar historical data point of the current one is identified by the motif distribution, which captures higher-order structural information beyond pairwise interactions. The fore-casted tendency of the time series is determined based on adjacent information and linear approximation, guided by proximity-based weighting. Through comparisons with typical network-based approaches using the real-world dataset, the effectiveness and applicability of our developed method are demonstrated, showcasing lower error indices.