Research On Ultra-Short-Term Forecasting Of Stocks And Stock Index Futures Based On Text And Price-Volume Data

The crash of China's stock market in 2015,the trade friction between China and the U.S.in recent years,and the COVID-19 have profoundly changed China's stock and stock index futures markets.Quantitative investment has strong capabilities for data analysis and is increasingly valued by institutional investors.Many public fund companies and securities companies have established quantitative investment departments in recent years.Quantitative investment has become the third important investment research method after fundamental analysis and technical analysis.Ultra-short-term trading refers to trading behaviors with high trading frequency and short holding time.Corresponding to the T+1 mechanism of stocks and the T+0mechanism of stock index futures,the holding time for ultra-short-term trading of stocks is one trading day,and the holding time for ultra-short-term trading of stock index futures is at the minute level.In the ultra-short-term environment,there are more trading opportunities and the profit accumulation effect is strong.Therefore,many institutional investors and researchers are attracted by short-term and ultra-short-term trading of stocks and stock index futures.The two most important elements in ultra-short-term forecasting are the forecasting models and the information on which they are based.In terms of models,the prices of stocks and stock index futures in the ultra-shortterm environment present strong nonlinear and complex characteristics,and all stocks and stock index futures form an overall complex system.Therefore,the traditional models that treat the price of a single stock or stock index futures as an isolated single time series will gradually be eliminated.All the models proposed in this paper emphasize the mining of cross-influence effects between multiple stocks and stock index futures.The deep learning model has a complex internal structure and a large-scale computation framework,which can fit the nonlinear characteristics of the data in an ultra-short-term environment.Traditional models can also be combined with text mining algorithms,obtain better performance by powerful information increment from the text information.In terms of information,the ultra-short-term forecasts of stocks and stock index futures are mainly relying on price-volume information and text information.Price and volume information is the internal information of the market,which reflects the comprehensive information in the market and is the result of the complex game of traders in the entire market.Text information is the external information of the market,which is an effective supplement to price-volume information,and plays an important role in the ultra-short-term forecast of stocks and stock index futures.Among them,the stock forum text is a direct reflection of investors' sentiment and contains important information.News text information contains huge price driving forces,and important news such as the COVID-19,relations between China and the U.S.and interest rate policies often lead to the huge volatility of the prices of stocks and stock index futures.Besides,the ultra-short-term trading of stocks is also closely related to the subject concept,which is an important path for emotional contagion among stocks.This paper integrates state-of-art deep learning models,text mining models and traditional financial models,improves forecasting models using stock forums text data,news text data,price-volume data and subject concept data.On this basis,this paper proposes new models to forecast the ultra-short-term price direction,arbitrage direction and yield of stocks and stock index futures,and constructs ultra-short-term trading and arbitrage strategies for stocks and stock index futures.The research content and contributions are as follows:(1)Ultra-short-term forecast of stock index futures based on price data and stochastic dominance.In terms of theory,this paper improves the state-of-art approximate stochastic dominance model based on the peak and biased characteristics of stock index futures and stock indexes in the ultra-short-term environment,proposes two new stochastic dominance models,and designs an ultra-short-term arbitrage strategy for stock index futures based on stochastic dominance under one-minute data.In terms of empirical research,this paper tests the proposed stochastic dominance model based on ultra-shortterm data of stock index futures in more than 8 years,and finds that it has advantages over the state-of-art models.The empirical research results based on the one-minute data of stock index futures in more than 5 years prove the advantages of our proposed arbitrage direction forecasting model based on stochastic dominance,and our arbitrage strategy has achieved good profits.The empirical research also analyzed the impact of the 2015 stock market crash and the strict supervision measures of the China Financial Futures Exchange on the market efficiency of stock index futures and stocks,and the results show that the damage caused by the strict supervision measures on the market efficiency is even more serious than the stock market crash.(2)Ultra-short-term forecasting of stocks and stock index futures based on pricevolume data,subject concept data and deep learning models.In terms of theory,this paper improves the state-of-art hierarchical neural network architecture based on the trading day segment embedding,solving the problem of discontinuous overnight price gaps,and further strengthens the ability to operate long-term sequences.This paper proposes a new embedding system for cross-influence between stocks which mines the cross-influence relationship among thousands of stocks at the minute level based on the subject concept information.A stock index futures forecasting model is proposed based on the stock index construction approaches,which predicts the price direction of stock index futures based on the predicted value of the stock price direction.In terms of empirical research,this paper conducts empirical analysis based on the one-minute price-volume data and subject concept data of all Chinese stocks in the past 9 years and stock index futures in the past 6 years.The empirical results verify the effectiveness of the proposed models,which have achieved better performance over the art-of-state models in both bull and bear markets.(3)Ultra-short-term forecasting of stocks and stock index futures based on the stock forums text and time series model.In terms of theory,this paper improves the state-ofart multivariate time series model by balancing the weights of the predicted asset and other assets,and then solves the parameter matrix based on the matrix partial derivative and maximum likelihood estimation.This paper also constructs sentiment indexes for all stocks and introduces them into our proposed multivariate time series model as an exogenous variable.In terms of empirical research,this paper conducts empirical analysis based on ultra-short-term stock data of stock index futures in the past 5 years and text data of Oriental Fortune stock forums in the past 1.5 years before and after the epidemic.The empirical results show that the model proposed in this paper has achieved very good performance.The empirical study also analyzed the impact of the COVID-19 on China's stock market,and the results show that the COVID-19 significantly reduced the textual sentiment index of Oriental Fortune stock forums,but increased the activity of stock forums,which enhanced the competitiveness of the ultra-short-term forecasting model based on text mining.(4)Ultra-short-term forecasting of stocks and stock index futures based on news text and deep learning.In terms of theory,this paper introduces an art-of-state text classification model in the computer science field by parameter migration and fine-tuning optimization with classifiers,which reduces the model's cost of computational resource,marks polarity of the news based on the abnormal fluctuations of stock index futures prices caused by news,and then constructs a new ultra-short-term forecasting model and proposes a corresponding trading strategy.This paper constructs a news stock mapping system based on subject concept data,mapping the news to individual stocks with the medium of subject concept,which has more advantages over traditional stock name-based mapping methods.In terms of empirical research,this paper conducts empirical analysis based on the news texts of the four major securities newspapers and Xinhua News Agency,one-minute price-volume data of stocks and stock index futures,and subject concept data in the past 5.5 years.The empirical results verify the effectiveness of text mining-based ultra-short-term forecasting model and news stocks mapping system.