Research On Multi-Time Scale Stock Forecasting Method Based On Clustering

The stock market is a huge and dynamic system.The stock data has high noise and unstable characteristics.The stock price fluctuation is affected by many factors.Therefore,the stock price trend forecast is very difficult.In the stock trading,a huge amount of data information is generated.In order to realize the forecast of stock price trends,people use the technical,fundamental and machine learning methods to conduct empirical research on these data,hoping to find the fluctuation law of stock prices,so as to sell stocks at high prices and buy stocks at low prices to maximize revenue.In the past studies,generally only a single or a few stocks and a single time scale data were used for forecasting,and the intrinsic link between stocks was not fully exploited.The paper deals with the existence of links between stocks,groups similar volatility stocks together.According to different time scales covering different amounts of information,and combine two time scales for research.The main research contents of the paper are as follows:(1)By calculating the correlation between the various characteristics of the stock,we use Affinity Propagation Algorithm(AP)to analysis with two weaker closing prices and trading volume on the A shares,and use the similarity to measure the clustering result of the two features,and find that stocks in same cluster is basically in the same industry.(2)Using securities industry stocks to construct two inconsistent securities clusters and using deep neural network(DNN),long short term memory network(LSTM)and temporal convolutional network(TCN)to conduct securities stock and single stock in the cluster(such as Guojin Securities)on stock price trend forecasting.The experimental results show that the stocks in the same cluster have more industry-related information than the single stock,and the stocks in the clusters will interact,thus having better prediction results.(3)Using TCN to compare the securities cluster and single stock in the cluster on different time scales,the experimental results show that the larger the time samplingfrequency,the better the prediction effect.Considering that the daily data and the minute cover different amounts of information,construct a multi-scale fusion model and compare the prediction effects of different fusion scales on the securities cluster and single stock in the cluster.The experimental results show that the multi-time scale can more effectively reveal the fluctuation of stock price trends.The paper mainly explores the correlations between stocks and different time scales coveing different characteristics of information.Using AP algorithm to cluster the similarly volatility stocks,and comparing the forecast result of clusters and single stock in cluster under different neural networks.We find that the stock data in the same cluster will interact,thus improving the prediction effect of single stocks in the cluster;comparing the analysis results of clusters and single stocks in the cluster on multiple time sclaes and single time scale,and we find that the multi-time scale can improve the stock price trend prediction effect than the single time scale.and the smaller the difference in sampling frequency between the two fusion scales,the better the prediction effect.