Research On Stock Index Prediction Model Based On Complex Network And Reinforcement Learning

In recent years,stock index prediction has attracted widespread attention in academia and industry,the stock correlation network has become a hot topic in the research of complex networks.The stock market is a highly volatilr,nonstationary,and nonlinear dynamic.Stock index forecast results provide the best time to buy and sell stock index futures.It has important implications for the research of stock index trading strategies.Most traditional stock index prediction research methods are based on the technical analysis,but no single technical indicator can continuously and accurately predict the trend of the stock market.Machine learning solves the problem of the failure of traditional technical analysis methods by transforming the stock index prediction problem into a binary classification problem,but these algorithms' effectiveness highly depends on the selection of input variables.The input variables that have been selected can improve the performance of the machine learning classifier model.Selecting these features is still a problem worth solving in the field of data mining,this is a worthy problem in the field of data mining.In addition,different machine learning classifiers have significant performance differences in different stages of different stock markets.How to integrate the prediction results of different machine learning classifiers to learn credible trading strategy is also an urgent problem to be studied.To solve this question,we propose stock index prediction model based on complex network and reinforcement learning.First,we build stock distance matrices based on the visibility graph method,generate stock correlation networks using planar maximally filtered graph model,extract topological mesoscale indicators based on complex network theory,and combine technical indicators as predictive variables.Then we calculate the creterion functions derived from quantitative investing analysis to train machine learning classifier models and learn the most credible trading strategies for each machine learning classifier using the consistency-based decision-making mechanism.Finally,we employ the reinforcement learning model to characterize the interaction betweeen an investor and the stock market,use the prospect theory to select the best investment strategy.In order to verify and evaluate the effectiveness of the proposed method,this paper selects the daily data of four representative stock markets at home and abroad(Shanghai Shenzhen CSI 300,American S&P 500,British FTSE 100 and Japanese Nikkei 225).Firstly,different criterion functions are selected for experiments on each classifier and rank sum ratio values of all criterion pairs are compared.We learn the best criterion pair for each classifier on the corresponding training set.Then,we set up different classifiers according to different predictive variables and analyze the effectiveness of topological mesoscale indicators through investment performance.Finally,the reinforcement learning model is trained based on the real historical data of stock index,and the simulated investment experiment is carried out in the real stock market.The experimental results show that the combination of TIs and TMIs can result in better trading strategies.Using criterion function to learn candidate prediction signals brings better investment benefits.At the same time,compared with the baseline prediction models,our proposed reinforcement learning method is more robust,and obtains relatively better investment performance.