Research On Stock Trading Model Based On ConvLSTM And Oriented To The Spatio-temporal Features Of Factors

For the last few years,the Chinese economy has an amazing increase.The stock market bacame an essential role in Chinese financial market.The stock market is not only a channel for companies to raise funds,but also an vital channel for ordinary investors to manage their financial investment.For investors,compared with other investment channels,stocks have lower investment thresholds and more profitable space,which attracts a large number of scattered investors to join it.However,risks and returns coexist,and there are too much factors affecting stock prices.With strong volatility,how to establish a feasible stock selection and trading strategy to reduce investment risks and obtain stable returns is a difficult problem for many investors and scholars.After entering the 21 st century,more advanced hardware was developed,almost everyboby use mobile phone which provide more strong computing power and big-data for deep learning study,making deep learning hot again.In the different areas of research,we can find the useage of deep learning.Meanwhile,there are more and more researches using it in financial markets and stock analysis,most of them are limited to the research of one-dimensional time series data.In this research,this paper innovatively converts one-dimensional sequence data of stocks into stock factor graph sequence data,and builds a new algorithmic trading model based on two-dimensional Conv LSTM and LSTM.Stocks have many indicator factors,and some indicators have certain spatial connections,which convey certain signals from the market and reflect the trend of stocks,such as golden chasing and dead chasing,which is the starting point of this article.Using the Conv LSTM fusion LSTM composite model to replace the traditional LSTM for stock forecasting research,and capturing the changing law of stock factor image sequence in market changes,find out its Spatio-temporal featurs.The main work and innovations are as flow:This article treats stock trading operations as an image sequence classification problem.This article uses trading behavior as labels,and the model predicts the best operations,such as buying,selling,and holding.This article uses several different technical indicators,each of which uses 6-27 days as the time calculation interval to calculate the factor value,and then selects 225 features from the calculated factors to construct a 15 × 15 size two,One-dimensional single-channel feature map.The feature maps calculated from different calculation targets(closing price,opening price)are used as different channels,and the daily stock data is converted into a two-dimensional three-channel image,and then images of consecutive t days are formed into an image sequence.This paper designs a composite model based on Conv LSTM and LSTM which used to train the images sequence.Through the experiment,it can be concluded that the model made by the this research performance best in every sizes which defeats the LSTM model.The quantitative trading tactics in this paper has achieved 16.7% annualized return on the CSI300 in 2017-2020.The total excess return rate is 26.34%.In stocks analysis,the model gets an average annualized return rate of 22.00% on 10 stocks,which far exceeded the traditional LSTM model 、 SMA model and the benchmark return.On the other hand,the investment risk was also at the lowest s level.The quantitative stock selection strategy developed at the same time also achieved an annualized return of 12.22% in the 2015-2020 backtest period,and the maximum retracement rate during the stock market crash was only 8.60%.The algorithmic trading model designed and developed in this paper can get excess benefit,and can help investors to reduce or avoid the risk of investment,the trading model has certain practical value.