Short-term Stock Price Prediction Based On High-frequency Data

Forecasts based on low-frequency financial data are long-term in time and depend on the overall economic environment,and it is often difficult to obtain satisfactory prediction results in the short term.In the context of today's big data era,benefiting from the rapid development of computer,communication and information storage technology,stock investors can more conveniently and timely access various transaction data of the stock market.The acquisition,operation,and analysis of minute-by-minute transaction data and transaction-by-transaction data become feasible.High-frequency financial data is rich in more information and is very significant for capturing micro-changes in the stock market.Therefore,using high-frequency financial time series to predict the short-term price trend of stocks,you can avoid risks more accurately and obtain better returns.BP(Back Propagation)neural network model was the most obviously that has achieved research results and most widely used neural networks.As a parallel and nonlinear complex dynamic system,BP neural network is based on the algorithm of error back propagation,so that it can also exhibit a very complex nonlinear system.However,because the performance of BP neural network is deeply affected by the input data,high-frequency financial data with non-stationary and strong volatility will cause some interference to the network,resulting in larger output errors.Therefore,the prediction result obtained by the traditional BP neural network algorithm is not ideal.Empirical mode decomposition is abbreviated algorithm,which was proposed by Huang et al.in 1998 when studying non-stationary and nonlinear problems.This method adaptively decomposes a non-stationary sequence into several intrinsic modal functions(IMF)with the original characteristics of the signal and a remainder.This feature fundamentally reduces the interference of complex data and plays a smoothing role and retain the maximum amount of information in the original data.In addition,EMD has an adaptive feature,which makes it more suitable for the analysis of some non-stationary and nonlinear complex sequences.In order to establish a more accurate short-term stock price prediction model,this paper proposes an EMD-BP neural network algorithm based on high-frequency data based on the idea of decomposition and reconstruction: First,the EMD algorithm is used to decompose the original price sequence into a finite number of IMF components and residual items.After correlation analysis of all components and the original price series,the BP neural network is used to perform rolling prediction on highly correlated components and new components combined with low correlation components,and finally,the equal weighted sum of the predicted values of the above components is reconstructed into the final prediction result.This paper takes the CSI 300 Index as an empirical research object,based on its 1-minute closing price high-frequency data and daily closing price low-frequency data,using the EMD-BP network algorithm proposed in this paper and BP neural network algorithm to make rolling predictions.These four sets of prediction results were evaluated using three performance indicators: mean square error(MSE),absolute mean error(MAE)and relative mean error absolute value(MAPE).The experimental results show that the prediction effect of the model based on high-frequency data is better than that of the model based on low-frequency data,and the prediction effect of the EMD-BP network model is better than that of the traditional BP network model.In summary,the EMD-BP neural network model based on high-frequency data can make a more accurate prediction of the fluctuation trend of the financial market,which is of great significance for the identification and control of financial risks.