A Temporal Convolutional Neural Network For Stock Trend Prediction

With the continuous development of global economic financialization and financial globalization,the stability of the financial market plays an important role in the operation of the real economy.The economic crisis transmitted by the financial crisis has a wider scope of influence and greater actual harm.Therefore,a certain degree of financial forecasting is a very important and necessary economic means to control and stabilize the development of the financial market.Over the past 40 years,China’s economy has made significant achievements and achieved considerable development.While the real economy has made rapid progress,the corresponding financial market has become increasingly important.How to give full play to the resource allocation and value discovery functions of the financial market,especially the stock market,is also particularly important.Therefore,the study of stock market and stock market volatility has significant practical significance.The stock market is a dynamic system,in which nonlinear,time-varying and complex correlations are widely distributed,and its trend is considered almost unpredictable.The early stock market prediction is usually a linear model dominated by measurement models.The measurement method has a strong explanatory power due to the construction of rigorous assumptions and the use of statistical methods,but at the same time,the linear structure of the model is constrained,It is often unable to predict the real price fluctuation.With the rapid development of artificial intelligence,more and more deep learning models are applied to the prediction of the stock market.Deep learning can receive massive structured data,support nonlinear high-dimensional models,and have greater data flexibility.Its model structure is closer to the data generation process of the financial system and even the economic system.By adding a trainable weight to each possible influencing factor,the model parameters are automatically adjusted through backpropagation,Get all the influencing factors and the interaction between these influencing factors to the greatest extent,extract the important characteristics of the hypothesis space of the stock market,so as to realize the prediction beyond the measurement method.Based on the above background and the characteristics of China’s stock market,this paper selects the Shanghai and Shenzhen 300 Index as the research object,and uses the method of deep learning to try to predict the volatility of the index by classification.The empirical analysis is conducted from the perspective of traditional econometric models,LSTM,GRU,TCN and improved TCN respectively.The next day’s price rise,fall and flat of time series data are classified as categories,and the accuracy of these models for this classification forecast is compared.Through empirical research,it is found that TCN model has better prediction effect and higher robustness than traditional econometric model and traditional deep learning model.Considering the particularity of China’s stock market,there is no long-term equilibrium relationship between financial markets in China.This paper attempts to improve the data set of the general TCN.By introducing core data from other financial markets,such as bond market,exchange rate market,money market and gold market,the potential of the TCN model is further explored,and better performance and prediction accuracy are achieved.In addition,because TCN model is usually applied to machine translation,speech recognition,image recognition and other fields,it is necessary to optimize the structure of the model when dealing with financial time series.This paper starts from the characteristics of financial time series and builds an adaptive TCN model from scratch,including the design of model structure modules and super parameter selection.Later,attention mechanism is added.As a new model in the field of deep learning,attention mechanism can model serial data well,and has a strong ability to capture remote dependent information,so as to better grasp the nonlinear characteristics of long time span input samples,and improve the model performance.The improved TCN model shows better performance and forecasting ability in the Chinese stock market,a special market.