The Improvement And Application Of LSTM In Multi-factor Quantitative Investment Model

In recent years,artificial intelligence has undoubtedly become the hottest technology research field in the world,especially the breakthrough progress of deep learning,which greatly promotes the popularization of artificial intelligence method.At present,artificial intelligence based on deep learning method has been successfully applied in the fields of speech recognition,handwriting recognition,computer vision,automatic driving and natural language processing,which achieved remarkable results.These research achievements have greatly promoted the progress of social science and technology.Currently,the basic conditions of artificial intelligence industry were available,with the depth of learning algorithms maturing and the accelerated growth of data resources,artificial intelligence technology is expected to be improved.At present,the artificial intelligence method based on deep learning has completely surpassed human in medical diagnosis and image recognition.Even in the uncertaintyfinancial market,the quantitative investment based on artificial intelligence technology has gradually replaced humansubjective investments.In fact,in addition to the increasing research on deep learning in financial investment by academics,financial institutions make extensive use of deep learning algorithms to invest in capital markets and begin to gradually replace traditional subjective investments.Many large-scale hedge fund companies are also beginning to deploy artificial intelligence,such as Bridgewater Fund,the Renaissance technology company,Two Sigma and Citadel,etc.they all have set up their own investment team of artificial intelligence research.According to statistics,at present,the quantitative investment in the United States has accounted for more than half of the market,while in China,this proportion is currently less than one-tenth.It is foreseeable that in the near future,the artificial intelligence based on deep learning will greatly change the existing financial quantitative investment theory and investment model in our country,which will inevitably have a great impact on future investment research in the financial field.The trend of quantitative investment gradually replacing traditional value investment is inevitable.However,unlike the tremendous achievements of deep learning in other fields,the research in financial field based on artificial intelligence technology has not make such a tremendous achievement by.Although many financial institutions devote a great deal of resources to the development of artificial intelligence technology,they can provide customers with intelligent investment consulting service by analyzing,forecasting transaction data and price change trends through machine learning,voice recognition and visual recognition.However,due to the great uncertainty of the financial market and the complicated and high-dimensional characteristics of the financial transaction data,the method of directly analyzing the financial data using the artificial intelligence method and designing the investment strategy from the obtained results has great challenge.Using the latest artificial intelligence algorithm technology to improve the accuracy of financial asset forecasting and reduce the investment risk is a key issue that deserves our attention.As the structure of deep learning model becomes more and more complex and the training data keep increasing,finding a suitable model to train the efficient modelefficiently is an urgent problem to be solved.The paper base on the time-series features of financial data,adopts cyclic neural network(LSTM)which has long-term memory effect in deep learning and conducts a multi-angle analysis of the financial market in China.In this paper,a factor selection method based on Elastic-net is proposed from a statistical point of view,and a multi-factor quantitative investment model with Elastic-net and LSTM is constructed.In addition,from the angle of risk management,a new RLF-LSTM model is proposed,and then the model is empirically analyzed in the options market and the stock market.Especially in the stock market,this paper also proposes a new elastic-net based factor selection method and constructs a multi-factor stock quantification strategy which is based on elastic-net and RLFLSTM,And the backtest of strategy is carried out.Specifically,the structure of this paper includes the following aspects.The first chapter of the thesis is the introduction.First introduced the rise of artificial intelligence in recent years and various industries are investing heavily in the application of artificial intelligence research.Second,it briefly introduced the great achievements of deep learning methods in many fields.This section analyzes the significance of introducing deep learning methods into the field of financial investment.Then listed the main research content and the logical framework of the study.Finally,this paper gives the innovative research work from many aspects.The second chapter is a literature review of deep learning research.The literature review is divided into four parts.The first part starts with the shallow neural network and introduces the development of artificial neural networks.From the first proposed neuron in the 1940 s to the introduction of the perceptron model in 1958,the BP algorithm was invented in 1986.This part of the literature shows several times of the rise and fall of neural networks.It also illustrates the hardships and twists of artificial intelligence research from another angle.The second part is the deep learning research literature.This section introduces several representative development process of deep learning algorithms,such as deep belief networks,deep automatic encoders,deep Boltzmann machines,convolutional neural networks and recurrent neural networks,and explored the latest research results of deep learning.The third part is an important part of the literature,which shows the latest research frontier of deep learning theory in the financial field.By reviewing the literature,we can find that the use of neural networks for asset price prediction has been very popular as early as the 1980 s.However,due to the “black box” problem of the computation of neural networks,many scholars of research on finance believe that artificial neural network models departed from the foundation of financial econometrics,it is also out of the scope of traditional financial econometrics.However,with the increasing ability of computers to process data and the natural advantages of deep learning in processing big data,the research results of applying deep learning methods to the prediction of financial asset prices have increased greatly in recent years.These research results have covered the traditional stock market,creditor's rights market and derivatives market,etc.,and these studies also show that the use of deep learning methods can have a better prediction effect than traditional financial statistics models.The review of these literature has also greatly enhanced our confidence in the use of deep learning methods to study China's financial markets.The fourth part summarizes and reviews domestic and foreign literature.Through the review of domestic and foreign literatures,we have found that in terms of how domestic and foreign companies use deep learning to predict the price of financial assets,they mainly focus on high-frequency data.Through the analysis of massive amounts of transaction data,we look for the law of asset price changes.However,there are few literatures on how to use low-frequency data for deep learning.The third chapter is the theoretical basis of deep learning algorithm.This section first introduced the basic principles of traditional machine learning algorithms and common loss functions.Then introduced the theory of neural network and deep learning algorithm,including the basic principle and development history of artificial neural network,and compared the properties of common neural network activation function.The core of this chapter is the introduction of deep learning algorithms,starting with the principles of deep learning algorithms,focusing on the network structure,working principle,and network variants of long and short memory networks in cyclic neural networks.The fourth chapter of the thesis constructs a multi-factor quantitative investment modeltheory based on the regularization method Elastic-net and LSTM.The test data is from January 1,2005 to December 30,2016.The daily stock price and return of all Shanghai and Shenzhen stocks include a total of 3,166 stocks over the past 2,915 trading days.In terms of factor selection,this article selects from valuation factors,market value factors,leverage factors,financial factors,momentum reversal factors,and a total of 39 stock factors.The first part of this chapter begins with the capital asset pricing model(CAPM),introduces the basic idea of the stock factor model,and sorts out some research results on the current stock factor screening and factor score evaluation.The second part of this chapter first introduced the regularization method of stock factor screening,compared the advantages and disadvantages of lasso,Ridge,and Elastic Net,and compared the factor regression result of lasso and Elastic net.Through experimental comparison,it was found that the Lasso method compressed the factor coefficients,directly removed the 14 factors with small regression coefficients,and retained only 25 valid stock factors,while the Elastic Net method compressed 10 factors,leaving 29 valid Stock factor.From the subsequent OLS,lasso,and Elastic Net building a quantitative investment model,Using the Elastic net method to select 29 factors to construct the backtest result of quantitative investment model is better than the OLS and Lasso.The third part of this chapter is based on the 29 stock factors selected by the second part of Elastic Net.This part of the experiment uses the Elastic Net method to select 29 stock factors as the experimental objects.The results compare the multi-factor quantitative investment model constructed using the LSTM with 39 stock factors and the 29 stock factors selected by using Elastic Net with the LSTM model.The experimental results show that when the number of iterations suitable,the effectiveness of the strategy constructed based on the Elastic Net-LSTM model is better than that only using the LSTM model.The fifth chapter of the thesis considers the characteristics that investors pay more attention to in the actual operation,and then revise the loss function,and proposes three kinds of linear stretching function,exponential stretching function and logarithmic stretching function.Then we analyzed the properties of the three kinds of stretching functions,and the linear stretching function was finally selected as the correction term of the mean square error,then a new loss function RLF was constructed,and the optimization algorithm logic for the new loss function RLF was given.In addition,due to the selection of the parameters of the stretching function,this chapter gives the idea of using the trial algorithm to optimize,that is,under the conditions of fixing other parameters,by iterative experiment according to a certain step in the scope of the experimental data,choose the parameter value with the smallest prediction error is used as the parameter to modify the stretching function,and then the algorithm logic for parameter optimization of the stretching function is given.The fourth section of this chapter is the empirical analysis.Based on the stretching function theory and the 29 stock factors selected by Elastic Net,the RLF-LSTM quantitative investment model is constructed,and several experiments are performed at different iteration times.Backtesting results show that the RLF-LSTM model can effectively reduce risk and increase investment return.The sixth chapter of this thesis begins with the structure of the model,considering that the key to the LSTM model is the introduction of a set of characteristics of the memory unit.When the LSTM is applied to quantitative investment optimization research,it is necessary to consider the self-regulation of the price change of financial assets.And according to the characteristics of different financial market to design research methods.In order to make the deep learning algorithm closer to the real situation of China's stock market,this chapter improves the LSTM network structure so that the oblivion gate function only memorizes the information of the stock price in the range of [-1,1] within 5 trading days,at he same time to forget the other nosiy data,this improves the quality of the information carried by the data entering the model.In addition,the transaction cost factor of Chinese stock market is added to the model of oblivion function and a new RFG-LSTM model is proposed.Then a deep learning portfolio model based on RFG-LSTM is constructed.The model was tested several times with different iteration times.The backtesting results of the model showes the great application value of the RFG-LSTM model.The seventh chapter of this thesis is the conclusion and the outlook.This section summarizes the main research results of this paper and believes that the research in this paper expands the application boundary of deep learning in the financial field.However,it also points out the inadequacies of the research,looks into the direction of future research,and provides guidance for further research.The innovation of this paper is to combine deep learning with the stock multi-factor model,and put forward a variety of improvement programs to master the frontiers of research in this field.The perspectives and ideas of the selected topics have certain novelty.Among the literature that the authors are concerned with,this article is the first to use Elastic-Net to conduct stock factor screening and construct a quantitative investment model based on multi-factor and deep learning.The model application is innovative.In addition,this paper is also the first time which propose to use the stretching function to reduce the risk of quantitative investment model.It is also the first time to propose a modified oblivion gate,and to construct a study of the RLFLSTM model and the RFG-LSTM model which provides a new idea to promote the application of artificial intelligence in Chinese stock market.