Stock Price Forecasting Based On Multi-scale Temporal Convolutional Neural Network

The Chinese economy is currently facing multiple tests such as a complex and severe international environment and domestic epidemic epidemics,but despite the difficulties,it has still delivered an economic answer sheet with notable achievements.China is the only country in the world to achieve positive GDP growth in 2020 and a near decade-high GDP growth rate of 8.1% in 2021,which is a hard-won achievement in today’s domestic and international context.Along with China’s rapid economic development and rising influence,China’s financial market is becoming increasingly important and domestic and foreign investors are competing more and more for the domestic capital market.Therefore,how to effectively build financial forecasting models for financial market analysis and portrayal has become a common topic of discussion between academia and industry.Through effective models,we can explore the intrinsic laws of the financial market,provide corresponding decisions and suggestions for investment,and thus achieve risk avoidance of related investment activities.Referring to the success of convolutional neural networks in the field of computer vision and recent examples of various variants of 1D convolution applied to tasks such as natural language processing,this paper applies a convolutional network model to the financial domain to handle stock price time series prediction tasks.In this paper,we construct a multi-scale temporal convolutional neural network(MTCNN)model by modifying the traditional 1D convolutional approach and the convolutional network architecture,which achieves feature extraction for different temporal ranges of the original data through convolutional kernels of various sizes,so as to achieve the effect of feature extraction from micro to macro and local to whole original data.The MTCNN model applies the concept of modularity to construct a time-domain convolutional module,and through the stacking of the module,deeper and more abstract feature extraction can be achieved,providing more detailed data and features for subsequent prediction work.In this paper,the model is analyzed empirically using all the data of the SSE 50 constituent stocks since their listing to date.From the cross-sectional comparison of the model output results,the MTCNN model performs better than the multilayer perceptron,traditional convolutional neural network and long and short-term memory models,and outperforms the other three classical models in predicting most stocks in the stock pool.Overall,this paper proposes a new network architecture MTCNN by improving and innovating the convolution method and network architecture of convolutional neural network,and migrates the convolutional neural network to the financial field for stock price prediction application,and achieves excellent experimental performance and detailed innovation.