Comparative Study Of Portfolio Strategies Based On MI-LSTM Model

In the context of a new round of industrial and technology revolution,with the vigorous development of artificial intelligence,cloud computing,big data,block chain and the internet,Fin Tech is applying new scientific and technological achievements in the financial sector.It also produces more profound impact on the organization system of traditional financial industry,the business process and the product design,etc.The Fin Tech development plan(2019-2021)released by the people's bank of China in September 2019 pointed out that accelerating the strategic deployment and security application of Fin Tech has become a necessary choice for deepening the structural reform of the financial supply side,enhancing the ability of financial services to the real economy,and making good efforts to prevent and defuse financial risks.The continuous development of Fin Tech is also promoting the development of quantitative investment.Automatic Financial transaction,scientific risk management,intelligent investment and financial management and other means enable quantitative investment to play huge advantages in the financial market.In view of the fact that quantitative investment in China is still in the early stage of development,the use of artificial intelligence,machine learning and other methods to characterize the non-linear characteristics of the financial market can help investors make investment decisions with broad prospects and great potential.The current academic research about the application of deep learning in the field of quantitative investment emphasized on using a reasonable deep learning network to predict the stock prices,choosing the input characteristics of the prediction and the processing methods of them,and constructing the portfolio strategy.Sezer et al.(2020)summarized the literature for financial time series prediction using deep learning from 2005 to 2019.There are about 31% of the literature researches concerning on stock price prediction,and models included multi-layer perceptron,recurrent neural network,long short-term memory,convolutional neural network,deep belief network and the autoencoder,etc.Zhai et al.(2007)used news information and technical indicators as input characteristics to improve the accuracy of stock price prediction.Tsai and Hsiao(2010)found that the methods of stepwise regression,principal component analysis and regularization were mostly used to extract effective information from input features.Arevalo et al.(2016)constructed a high-frequency "buy low and sell high" trading strategy based on the stock price prediction results from the fully connected feedforward network.According to the existing research results,most of the literature focus on using advanced network structure and appropriate feature selection methods to improve the prediction accuracy of stock price,but ignore the construction process of portfolio strategy and fail to effectively control portfolio risk.In the context,this paper will build the stock price prediction model based on the long short-term memory and attention mechanism,and introduce the prediction results into the improved portfolio optimization model to build the portfolio strategies.The attentional mechanism not only extracts the input characteristics,but also pays attention to the output information on different time steps.At the same time,this paper also carries out a comparative analysis on the changes of attention weight of input features during model training,the prediction effects of the same model on different stocks and different models on the same stocks.Finally,the performances of back testing and effective frontier changes of portfolio strategies outside the sample period are compared and analyzed,and the characteristics of each strategy under different market styles are discussed.Constructing portfolio strategies based on deep learning needs data collection and data preprocessing,training learning model and obtain predictions,asset allocation based on the prediction results and optimization model,and the evaluation of prediction models and strategies.It is also necessary to choose effective input characteristics,set up the appropriate model parameters,and consider the rationality of risk measures.Therefore,this paper adopts the research method of combining normal analysis with empirical analysis.In terms of normal analysis,this paper will summarize the current literature on the stock price forecasting methods based on the deep learning,feature selection methods in the task of forecasting stock prices,the application of quantitative strategy based on deep learning,and the mean variance portfolio theory.This paper also summarizes theoretical directions of current researches.In terms of empirical analysis,this paper uses the open source tool Tensor Flow to build a stock prediction model based on the long and short-term memory and attention mechanism to predict the closing price of constituent stocks of the SSE 50 index.Basic data processing and relevant numerical calculations are completed by Python,and optimization tool in the Scipy package is used to solve the optimization models.This paper has five chapters.The first chapter is the introduction,which describes the research background,research significance,research method,research content and innovation of this paper.The second chapter is literature review,which introduces in detail the research status of the stock price prediction methods based on deep learning,the feature selection methods in the stock prediction task,the application of deep learning in the construction of quantitative strategy,and the development of portfolio theory under the mean variance model.The third chapter is the research design.Firstly,the long short-term memory network is used to construct the stock price prediction model in this paper,and gives the embedding mode and principle of attention mechanism.Secondly,based on mean variance model and tail risk measures,the portfolio optimization models under different risk measures are obtained.At last,this paper describes how to construct the portfolio strategy based on the prediction results of the models and the improved portfolio optimization models,and introduces the common risk evaluation indices in the back testing analysis.The fourth chapter is the empirical analysis.This paper uses the daily closing prices of SSE 50 index and some of its constituent stocks from 2008 to 2018 for empirical analysis.Firstly,based on the stock prediction models established in the previous chapter,the closing prices of component stocks are predicted by setting reasonable model parameters,and the changes in the attention weight of input features during model training,and the prediction effect of the same model on different stocks and different models on the same stocks were compared and analyzed.Secondly,the expected returns calculated based on the prediction results are used to replace the sample mean estimation used in the portfolio optimization models based on risk measures such as variance,Va R,CVa R and CDa R to construct the portfolio strategies,and the back testing performances of each strategy in the period outside the sample period was compared and analyzed.At last,it analyzes the back testing performance of each portfolio under the consideration of transaction fees and explores the prediction effect of the model on the stock index.The fifth chapter are the conclusions.Based on the above empirical analysis results,it summarizes the conclusions of this paper,puts forward relevant policy suggestions for financial market participants based on the research conclusions,and analyzes further research directions for insufficient details.The main conclusions of this paper mainly include:(1)The future price of the SSE 50 constituent stocks is not only influenced by their own historical price,but also influenced by the historical price of other stocks with positive/negative correlation and index.The historical price of stocks is most closely related to its future price.(2)The MI-LSTM model based on attention mechanism can capture the different importance of input features,adaptively extract useful information beneficial to the prediction task from multiple input features,and improve the prediction accuracy of the original LSTM model on the prediction of partial stock prices.(3)The prediction effects of the MI-LSTM model on the stock prediction tasks are closely related to the information contained in the input features,and the improper selection of the input features may lead to the poor prediction effects of the model.(4)The prediction errors of LSTM and MI-LSTM models on different stock prediction tasks are different,and the prediction effects on stock price series with more drastic price fluctuations are worse,which reflects the complexity and uncertainty of stock prices.(5)Using expected returns obtained from neural networks instead of the sample mean estimates can effectively improve the profit and risk resistance ability,and is able to outperform index and equally weighted portfolio.Using more precise input estimations can improve profitability,and reduce volatility and risk of drawdown.(6)From the perspective of market style,when the market is in a state of slow bull,the variance-based portfolio is more able to obtain excess returns.When the market is in a state of unilateral decline,the portfolio based on CDa R can resist the losses of asset value and avoid the investment risk effectively.This paper has several innovation points:(1)The stock prediction model based on attention mechanism is put forward,including attention mechanism embodied in the input features and time steps,considering the influence of positive and negative related to the stock price and index price.The new model can capture the different importance of input features,and extract useful information from the input features and the hidden states of time steps.(2)Existing literature have explored the application of portfolio optimization models in practice by using Bayesian approach,global minimum variance portfolio and robust optimization.Most of the methods used in these studies are based on the assumption of normality of security returns and ignore the information contained in the large amount of trading data.In this paper,the expected returns calculated based on the stock prediction model is used to replace the sample first-order moment estimation in the optimization model.(3)At present,there are few studies on the application of deep learning in the construction of quantitative strategies,which are often the "buy low and sell high" strategy,and the domestic research are mainly applied in the future and option market.In this paper,the advanced deep learning algorithm is combined with the traditional portfolio optimization model to construct an effective portfolio strategy for China's stock market,which enriches the application research of deep learning in the quantitative strategies and provides an effective reference for subsequent researchers.This paper also has the following shortcomings:(1)The input features in this model only contain price data,and the selection of features is subjective to some extent.It is necessary to explore whether the addition of some fundamental data such as profitability data,growth data,leverage data or public opinion data of major domestic financial and economic forums can have positive effects on the prediction of stock prices.(2)The asset allocation is conducted on the daily basis in this paper,and the adjustment frequency of fund manager is usually a month or a quarter,and it is really difficult to bear high fees for the daily adjustment.It is necessary to explore how to use high performance neural network to realize high quality prediction of low frequency data and make the research conform to real life situation as much as possible.