Financial Time Series Prediction Based On Plate Effect And Multivariate Features

Stocks are an important form of investment in people’s daily lives,and it has always been a concern for investors and financiers to reasonably analyze stock price movements and make accurate predictions.However,there are characteristics of stock data such as nonlinearity and multiple features,and it is difficult for traditional forecasting models constructed based on statistics to make accurate forecasts of stock prices.In recent years,artificial neural networks have made great progress relying on advances in hardware technology and can efficiently process large amounts of multidimensional data,which provides new research directions in the field of stock prediction.Stock price movements are related to a variety of factors,such as plate effects,macroeconomic policies,technical indicators,and other factors that affect the prediction accuracy of the model,while current stock forecasting methods often consider only one aspect of the characteristics.In addition,the stock prediction model composed of a single deep learning neural network will appear over-fitting phenomenon in the prediction process,and it is difficult to make accurate predictions of stock price.To address the above two problems,the following two financial time series forecasting methods are proposed in this paper.1.A deep learning stock price trend prediction method based on industry effect is proposed.The research finds that certain factors（such as policy changes and social emergencies,etc.）will have the same influence on the prices of multiple stocks in the same industry,leading to a similar trend of multiple stocks in the same industry in a certain period of time,namely,plate effect.To address this phenomenon,the method first analyzes the closing prices of multiple stocks under the same sector using Pearson correlation coefficient and XGBoost algorithm to select stocks with high correlation and uses principal component analysis to downscale the screened stock data to extract the price trends of the stocks;Secondly,a CNN-LSTM prediction model is constructed to solve the problem of poor prediction accuracy of a single neural network.The model uses stocks of banks,pharmaceuticals,alcohol and entertainment media as the experimental dataset.Through comparative experimental analysis,the deep learning stock price trend prediction method based on the plate effect achieved better indicators in all four plate datasets.2.A generative adversarial network model based on multivariate features is proposed.To address the two issues of the impact of fundamental and technical analysis of stocks on stock prices and the still poor prediction accuracy of CNN-LSTM models.The model first uses wavelet transform to denoise the exchange rate features and uses multiple financial technical indicators generated by Ta-lib financial database and principal component analysis method to reduce the dimensionality of the generated technical indicators.Secondly,the model uses generative adversarial network as the prediction framework.In the generator,the CNN-LSTM model based on attention mechanism is used to predict stock data.In the discriminator model,convolutional neural network and multi-layer fully connected network are used.The method is experimentally proven to outperform the comparison model used in terms of fitting and prediction accuracy.In response to the stock prediction model,stock prices are affected by the plate effect,stock fundamental analysis and technical analysis,and the low prediction accuracy of current prediction models.Two financial forecasting methods are proposed and verified in several financial data sets.The results show that the two financial forecasting methods proposed in this paper are better than the comparison model,and have a better fitting effect and prediction accuracy.