Bond transaction link prediction based on dynamic network embedding and time series analysis

Abstract

Trading behavior prediction is to estimate the possibility of the occurrence of links in a dynamic network of bond transactions. At present, most of the existing link prediction models are link predictions for static networks such as social networks that do not consider time dimension. Since the evolution of the network over time is not considered, it is difficult to meet the object of effective link prediction of bond transactions. In this paper, in order to meet the link forecasting demand of bond market risk warning, DNETSA's link prediction method is proposed to realize the link prediction task under dynamic network, which provides a basis for financial risk warning. The DNETSA method can effectively extract the advantage of the structural information of the network in each time period. Then combine it with the link number attribute information by means of the time series model, which realizes the prediction ability of the link in the dynamic network and overcomes the problem that the static network link prediction does not consider the shortcomings of the network evolution over time. The effective integration and utilization of the dynamic network structure information, time information and attribute information makes the DNETSA method increase the AUC value by 22% compared with the LMPF method, and the AUC value by 13% compared with the TS-sim method. Compared to the TS-occ method AUC, the value is increased by 12%, which is 9% higher than the AUC value of the SOTS method. In summary, the DNETSA method makes up for the shortcomings of other methods and can satisfy the prediction of bond trading behavior.