Research On Stock Price Prediction Methods Based On Deep Learning

The stock market occupies a crucial position in the national economy,and in recent years,research on stock prediction models based on deep learning has received widespread attention.Numerous scholars focus on the technical analysis of stock data,which is a technique that can predict the future stock price trends.The quality of feature representation for stock data directly affects the performance of prediction models.The existing stock prediction models are poor in the long sequence,lack of temporal dependency learning between adjacent time steps,and the issue of repetitive training with stock data.Therefore,this study focuses on researching stock data feature representation methods,stock price prediction methods,and model pre-training,the main research content of this thesis are summarized below:(1)Proposing a novel Multi-scale Time Series(MTS)stock data feature representation method,which treats the features of each trading day in the stock data as a token,and embeds the intraday data into the feature representation of the trading day token,consequently,it can enrich the feature expression of trading days.Based on this,a reverse Cross Attention(rCA)is proposed to improve the cross-attention mechanism in the Transformer model,which allows the model to focus more attention on data closer to the prediction date.Compared with other models,the rCA-Transformer stock prediction model,which combines rCA,exhibits better long sequence prediction and temporal dependency learning abilities.This thesis conducted experiments on multiple datasets,including regression prediction and stock rise/fall classification tasks,and the results show that the rCA-Transformer model combined with MTS significantly improves prediction accuracy compared to traditional methods.(2)Proposing a pre-training method of stock prediction model with Kalman filter(PTSK),which utilizes the period embedding of Kalman filter for pre-training on a pretraining dataset,capturing the common features between stock price trends and periods.Then,the pre-trained model is fine-tuned on small-scale or specific stock data to adapt to the training task.The PTSK pre-training method allows the model to learn more stock tokens and effectively apply the learned common features to the fine-tuning dataset or special cases with limited samples,such as stock price limits.Experimental results show that compared to stock prediction models without pre-training,PTSK exhibits significant performance improvement in both regression prediction and classification prediction.In addition,this study also conducts simulated quantitative trading experiments using two simple quantitative trading strategies,and the back testing results demonstrate that the fine-tuned PTSK stock prediction model achieves higher returns and Sharpe ratio.