Time Series Forecasting Based on Temporal Networks Evolution and Dynamic Constraints

Time series forecasting holds significant application value in financial market analysis, biological prediction, and other domains. Analyzing time series from a network perspective offers novel insights for forecasting. This article proposes an innovative time series prediction method based on the network characteristics of time series. First, the time series is mapped to a complex network with node weights assigned according to association of nodes. Subsequently, leveraging the unique evolutionary patterns of temporal networks, we employ matrix evolution to predict the topological structure of the network at the next time step. A loss function is constructed by integrating the ability of weighted temporal networks to reconstruct the original time series data and visibility-based constraint features, enabling the computation of predicted values for next time step. The proposed method achieves prediction considering global topological characteristics of temporal networks during the forecasting process. To validate its performance across diverse scenarios, five time series datasets from distinct domains are selected. These datasets encompass varied temporal granularities and exhibiting periodic, stochastic, trend, and seasonal patterns. Prediction errors are calculated to evaluate method efficacy. Statistical tests and error metric results demonstrate that the proposed method achieves superior prediction accuracy compared to existing forecasting approaches under consideration, which confirms that this work not only contributes theoretically but also provides a practical solution for time series forecasting applications.