Research On Stock Forecasting Based On Deep Learning And Text Mining

The financial market is an important part of the national capital market.It plays an irreplaceable role in maintaining the sound and rapid development of the national economy.As a typical representative,stocks are deeply concerned by investors.Through the research and forecast of stock trends,investors hope to explore successful profit-making models,and managers can regulate in time.Due to the nonlinearity and volatility of stocks,research on stocks has not made a big breakthrough,but its enormous challenges and temptations inspire people to constantly explore.With the deepening of concepts such as smart finance and quantitative investment,people want to find more powerful intelligent algorithms for financial markets.Machine learning and deep learning have achieved outstanding results in many fields due to their powerful representation learning ability.Under this trend,they are gradually being applied in the financial field.With the need for modeling stock market integrity and the increase of data volume,text information has gradually become a mining object for scholars to study stocks.Since stock forecasts are still in the research stage,establishing an effective portfolio has become a common method of stock investment,neural network models can also help achieve this goal.The main contributions of this paper are as follows:1)For the problem of stock trading timing prediction,a complete forecasting method was established.This method uses adaptive threshold piecewise linear representation algorithm to generate transaction labels for each trading day,and proposed to construct technical index factors into multi-channel pictures as input data.This method trains a convolutional neural network to predict the trading operation of the next trading day.Experiments show that the predictive classification performance of the model has practical significance.The experimental results of simulated trading and portfolio investment show that the model is suitable for short-term trading and provides more profit opportunities.2)This article explores the use of fundamental information to forecast stocks.The research reports issued by the securities company and the financial indicators disclosed in the company's financial statements are collected for analysis.For the research report,the chi-square test and the word bag model are used for text representation.The experimental comparison shows that the random forest has obvious advantages.For the financial indicator data,feature engineering construction and the idea of local normalization are proposed to remove the influence of enterprise differences on financial factors.Experiments show that the accuracy of classification prediction is improved.In the task of predicting the long-term trend of stocks,the best accuracy of both experiments reached more than 60%.3)To establish an effective index tracking stock portfolio,the decoding coding neural network model is proposed.Decoding network restores the stock information from the stock index,those stocks with good restore effects will be selected as the combined components by ranking the correlation coefficient.Coding network simulates combined reconstruction process,determining combination weight through training.The validity of the method in this paper is proved in the combination construction of the Shanghai Stock Index.