A study on stock market forecasting based on time series with integrated multi-feature LSTM modeling

This paper constructs an effective stock prediction model, which can help investors better understand the market trend and accurately grasp the fluctuation pattern of stock prices, thus guiding them to formulate reasonable investment strategies. Taking Hengrui Medicine as the research object, we collected the stock data, established the feature engineering, set up the index evaluation system for the stock features, and constructed the six features of turnover, average price, relative increase, turnover change volume, increase/decrease amplitude, and amplitude from the aspects of enterprise value, enterprise development potential, and capital stability, respectively. In addition, principal component analysis is used to extract effective information and remove redundant information from the constructed features and original features. The decomposition model, exponential smoothing model, ARIMA and SARIMA are applied to the stock price prediction task of Hengrui Medicine. For the decomposition model, the STL and SEATS decomposition methods are used to decompose the closing rate into seasonal terms, trend-period terms, and residual terms. After that, we tried to integrate machine learning models, such as XGBoost, Adaboost and other Boosting models, and Random Forest and other Bagging models, and formed the meta-model into a Stacking model to evaluate the effect of different methods on stock feature extraction. Finally, we use LSTM model in deep learning field and apply its single-feature and multi-feature input dimensions respectively, and combine the previous models to form an integrated machine learning-multi-feature LSTM stock prediction model, and horizontally compare the traditional and modern methods, and get the conclusion that it has a good fitting effect.