Research On Extreme Value Prediction Of Stock Index Based On Complex Network And Visibility Graph

The financial market is the main place for ordinary investors to carry out investment activities.However,the emerging capital market represented by China's capital market is relatively more volatile due to its short establishment time and incomplete regulatory system.Effective prediction of the stock market has become a hot topic in risk management research in the field of financial investment.Predecessors modeled stocks or stock markets through various methods such as time series and machine learning to predict their volatility,avoid risks and increase returns.Based on previous work,this paper intends to map the stock index,which representing the overall movement of the stock market,into a series of weighted visibility graphs.We construct the networks based on the visible rule and the completely invisible rule,and the weight of edges are calculated by the depressed angle.We construct peak prediction indicators for the nodes in the weighted visibility graphs,and trough prediction indicators for the weighted invisibility graphs.Aiming at the problem that the properties of nodes in visibility/invisibility graphs algorithms are easily affected by their neighbor nodes,it takes neighbor nodes into predicating indicators.Finally,we use these indicators to predict the extreme values of peak and trough in the stock index.The results show that by using the methods proposed and improved in this paper for prediction,our methods' effect is better than previous methods.This article uses the 19-year Shanghai Stock Index daily closing price series from January 2001 to December 2018 to model,and to make effective predictions of extreme events during this period.It details the fluctuations of the stock market in 2015 and the changes in forecast indicators.Then the influence of parameters in the algorithm is analyzed.Finally,we extended the method to 12 major stock market indexes around the world to predict their extreme events.The results show that the method proposed in this paper has significantly improved over the previous methods and other algorithms.