Research Of Stock Price Prediction Based On Stacked Autoencoder And LSTM Prediction Model

Stocks,a financial product that emerged with the emergence of joint-stock companies,have a history of nearly 400 years so far.At the same time,joint-stock companies and Stock Exchanges were born with him,bringing great changes to the world economy.For investors,stock investment has a low threshold and high returns,which makes a large number of small and medium-sized investments to join in.However,Risks and returns coexist,there are always many factors affecting stock prices,coupled with the prosperity of international trade and the instability of the world economy.So seeking a practical stock selection strategy to reduce investment risks has become an overhang for many investors.In recent years,with the continuous development of machine learning disciplines and computer computing power in recent years,more and more people try to apply deep learning algorithms to various fields.Among them,integrated algorithms rely on their many substate models and generalization.,Especially popular in industry.In thepere major competitions in recent years,the LightGBM integrated model based on the gradient descent tree has emerged with its excellent calculation speed and accuracy.On the other hand,the recurrent neural network model has excellent performance in processing financial time series,and the longshort-term memory network model improved by the recurrent neural network is even more sought after.Used in stock price prediction.The stack self-encoding algorithm is an algorithm developed based on unsupervised machine learning.It has a unique performance in the extraction of advanced semantics and feature engineering.In this thesis,the stock price data of the pre-recovery stocks in Kweichow Moutai,from March 1,2009 to March 1,2019 for a total of 10 years(excluding holidays,double holidays,and suspensions)are used,a total of 2,420 data research.Six technical characteristics were selected: short-term 5-day moving average and long-term 10-day moving average;Bollinger upper line and Bollinger lower line;K line and D line.In addition,8 data characteristics: the highest price,the lowest price,the opening price,the closing price,the price change,the turnover rate,the trading volume,the Shanghai Composite Index.A total of 14 features were used as the original feature set of this thesis.The closing price of the next day is the marked value.Let's empirically prove the innovation of this thesis.In this thesis,the LightGBM model is transformed,and an improved LightGBM stock price prediction method based on stack self-coding is proposed.The stack self-coding algorithm is used to replace the mutually exclusive feature binding algorithm,and the stack self-coding algorithm is used for 14 original features.The set is used for feature extraction,and the extraction result is combined with the original feature set to form a composite feature set,which is then input into the LightGBM model for training,which increases the generalization of the decision tree set.Through the empirical data of Kweichow Moutai,comparing the improved LightGBM model based on the stack self-encoding algorithm with the traditional LightGBM prediction model,a more accurate fitting curve and a more stable yield curve are obtained,which has certain practical significance.On the other hand,in view of the non-sparse and non-high-dimensional characteristics of stock data,this thesis proposes a stock price prediction method based on stacked self-encoding and long-short-term memory networks.The network model is used in combination.First,the stack-type self-encoding algorithm is used to extract features from the 14 original feature sets.The extracted results are combined with the original feature sets to form a composite feature set,and then the long-short-term memory network model is input for training.According to the empirical data of the former reweighted stocks in Kweichow Moutai,based on the stack self-coding and the long-term and short-term memory network prediction model,compared with the traditional long-and short-term memory network model,the average error rate of the former model will be 5% under the time When the step length is 15,it has the most stable and accurate results.Compared with the LightGBM model based on the stack self-encoding algorithm,the prediction model based on the stack self-encoding and long-short-term memory network has increased in training time,but the gain curve has increased.