A Stock Price Prediction Network Based On Multi-scale Convolution And Residual Multi-layer Perceptron

Since the establishment of China’s stock market,the rules and regulations have become more and more complete,and the degree of openness has become increasingly higher,which attracts a large number of investors.With the increasing attention of investors to the stock market,stock price prediction has become a crucial topic.Under the influence of the macroeconomy,policy,business conditions and investor sentiment,stock prices tend to be non-linear and highly volatile.Therefore,it is very challenging to accurately predict the future stock prices.Up to now,there are many stock price prediction methods,among which deep learning technology has made significant progress in the field of time series prediction,and has become a crucial tool for stock price prediction.Deep learning combined with decomposition algorithms is a common structure utilized in the current time series analysis works,and has been demonstrated to be effective in predicting stock prices.However,in the existing works,a single-scale convolution neural network is often used to identify local patterns in subseries,which cannot be adapted to stock price series with strong time-varying distribution for a long time,and has the disadvantage of information loss,thus weakening the generalization ability of prediction models.In addition,the prediction of subseries is usually independent,which ignores the interaction between interrelated subseries.This thesis proposes a stock price prediction network MCR-Net based on multi-scale convolution and Residual Multi-Layer Perceptron(ResMLP),which improves the prediction performance of the network by fully extracting the global and local features of the decomposed subseries and modeling their correlation.For the subseries decomposed by Singular Spectrum Analysis(SSA),MCR-Net captures long-term dependencies of each subseries through a Long Short-Term Memory(LSTM)network and identifies their local patterns at multiple time scales through a set of one-dimensional Convolutional Neural Networks(CNN)with different convolution kernels.For stock price series with time-varying data distribution,multi-scale convolution can extract the local feature of multiple scales and alleviate the loss of information caused by single-scale convolution.MCR-Net further uses ResMLP to model the correlation between global and local features of subseries,respectively.ResMLP’s special structure of internal data transposition and fully connected layer overlay captures the interaction between subseries at the feature level to improve prediction performance.Experimental results on stock indices in different countries around the world indicate that MCR-Net has improved prediction accuracy compared to recent advanced prediction methods.To overcome the drawbacks of stock price prediction methods based on decomposition and deep learning,MCR-Net provides a new solution:1)Using multi-scale convolution neural networks to identify multi-scale local patterns in subseries,which has wider applicability and less information loss than single-scale convolution.2)ResMLP model is used to model the correlation of subseries and to explore the potential interactions between them.