Time Series Forecasting Based on Complex Network Analysis

Time series forecasting, especially from the perspective of the network, has been a hot research topic. In this paper, based on the analysis of complex network, a novel method is proposed for more accurate time series predictions. First, time series data are mapped into a network by visibility graph. Then, the link prediction method is adopted to calculate the similarity index. Considering that node distance is an important factor in the network, we take that into account to determine the weight coefficients and improve the predictive results. To fully verify the validity of the proposed method, it is applied to some representative time series data sets with different characteristics. The data values are recorded daily, monthly, and yearly. The error measurement and correlation analysis show that our method has a good prediction performance. It is believed that this paper will not only contribute to time series forecasting in theory but also take effect in practice.