Research And Implementation Of Stock Trend Prediction Model Based On Deep Learning

With the rapid development of the economy in China,the financial market goes into prosperity.As an important investment method in the financial field,the stock market has complex and changeable characteristics,its future trend is affected by many factors,such as industry development prospects,historical price information,national policies and regulations,and so on.Therefore,stock forecasting has become a valuable subject that is always concerned by many scholars and investors.By analyzing the impact factors of stock trend comprehensively,this dissertation uses capital,market and technical data besides trend data to builds a basic numerical data set.At the same time,as the emotional content,such as news reports,has a significant effect on the stock market,this dissertation crawls daily relevant research reports of target stocks,and use the Sentiment Analysis technology in the field of NLP to analyze these texts,by which we can get the input factor for our stock trend prediction model.In this dissertation,we discard the classic time series prediction model,but adopt GAN as our data model,a model framework that better matches the operating logic of the stock market,and proposes a text-assisted generative adversarial network stock prediction model STP-GAN.This model uses both numerical and textual datasets at the same time,and adds Text Pathway to the GAN model,which combines stock price prediction and sentiment analysis into a synchronous integrated network structure,thereby reducing the accumulation of errors in model training.At the same time,STP-GAN model introduces the Transformer structure to better improve the learning ability of stock-related data features.In addition,STP-GAN uses BERT to perform finetune in the financial field,and adds a financial dictionary to increase the adaptability of the model to specific financial fields.Based on STP-GAN,this dissertation further optimize model training.For the difficulty of convergence and easy overfitting of GAN network,a new optimizer RMSPropW,which is more suitable for this subject,is proposed.The optimizer uses the exponentially weighted moving average as the decay coefficient,and uses weight decay to improve the training stability and training speed of the prediction model.In terms of experimental demonstration,this dissertation designs multiple sets of control experiments in detail based on the core model constructed in the above chapters to prove the effectiveness of the STPGAN model and the RMSPropW optimizer.In the market back test verification,this dissertation constructs a quantitative back test platform based on zipline,and proves the good market application value of the STPGAN model by comparing the trading strategy with the market index.