A Study On Quantitative Investment Strategy Based On Improved Recurrent Neural Network

The stock market is a dynamic,non-linear,high-noise system.Due to many complicated factors such as political factors,economic conditions and investor expectations,stock price changes are often nonlinear and unsustainable.At the same time,with the advancement of science and technology and the escalation of information dissemination systems,the time required for incidents to be reflected in stock prices has become shorter,which has led to a more sensitive stock price response to related events and a higher frequency of volatility.However,due to the possibility of obtaining high returns,the stock market has always been a market focus in the fields of finance,engineering and mathematics.With the development of science and technology,the stock market theory that people have and the maturity of investment skills,more and more stock analysis and stock price forecasting systems have been developed into stock investment practices.For the time being,the main stock price forecasting techniques in the financial market are mainly divided into basic analysis method,technical analysis method and quantitative analysis method.The further development of mathematical models and financial theory has made the role of quantitative analysis in stock forecasting more and more important.The traditional quantitative analysis model is suitable for processing linear,low-noise time series data,but most of the financial time series data is characterized by nonlinearity and high noise.The neural network model has been increasingly valued by scholars and investors due to its excellent ability to process nonlinear and high-noise data.Previous studies by researchers have shown that different neural network models can achieve good results when selecting appropriate parameters,but the structure of the neural network model is also very important because it directly affects the performance of the entire neural network model.Based on the previous studies,this paper improves the traditional artificial neural network(ANN)model structure,and uses the LSTM which is good at learning and preserving information from long steps,to construct an improved recursive neural network(ANN-LSTM)model.This paper is divided into four chapters.The first chapter briefly introduces the research background,research significance and the research progress of the neural network model for stock forecasting technology.The second chapter introduces the theory of stock forecasting theory,analyzes the structure and algorithm of neural network.The third chapter briefly describes and preprocesses the data used in this paper to eliminate the influence of missing values and outliers,and uses the dimensionality reduction technique to construct the most relevant and most influential feature variables.The fourth chapter is to study and train the ANN-LSTM model in this paper,and use the ANN model as a comparison.The two trained models will be used to forecast the direction of the closing price of the Harvest CSI 300 stock index fund,accuracy and loss function of test set are the main metric.The fifth chapter develops a quantitative investment strategy based on the prediction results of the two models,two benchmarks are used to measure how well the models can perform.Through empirical comparative analysis,it is found that the ANNLSTM model designed in this paper obtains higher prediction accuracy than the ANN model when it is used to predict the price of the Harvest CSI 300 stock index fund.In the quantitative investment strategy research,the ANN-LSTM model we designed also showed better practical application ability,which showed higher cumulative yield and risk-adjusted return.Then the sample data is divided into two parts based on the time node when the 2015 stock market crash occurs.The results show that the ANN-LSTM model can achieve better performance when the stock market is repairing or steady than when the stock market is in a high period.