Stock Price Movement Prediction Based On Multi-sources And Heterogeneous Data

The stock market is an adaptive and complex system,existing many factors that affect stock price movement.In recent years,a large amount of valuable information in the form of text has emerged with the rapid development of the Internet.Especially with the breakthrough of deep learning algorithms and the improvement of computing power,multi-source and heterogeneous data can be utilized to predict the stock price movement.From the perspective of the intersection of computer and finance,this work tries to take into account the strong interpretability of financial theory and the advantages of deep learning's automatic feature learning,and builds a framework that integrates multi-source heterogeneous information to predict the stock price movement.The main contributions of this work are as follows:(1)A deep learning model that integrates industry rotation and co-movement information for stock price movement prediction is proposed.The model added a parameter matrix-the industry correlation coefficient matrix-in the data layer.It can be updated with model training and can automatically learn the influence factors between industries.Besides,the model used the deep neural network based on the factorization machine(DeepFM)to automatically extract the interactive features among the input data.The results showed that the model integrating multi-source industrial time series data had significantly improved prediction accuracy compared with that with only single-source time series data.Increasing the extraction of interactive features of input data enriched the degree of input data mining,and also significantly improved the prediction accuracy.In addition,after the model training,the learned industry correlation coefficient matrix is compared with that computed by dynamic time warping algorithm.It is found that theses two matrix were largely consistent,which indirectly reflected the excellent ability of deep learning model in dealing with complex prediction problems.(2)A deep learning model that integrates public sentiment data for stock price prediction is proposed.The model used an explicit kernel mapping layer constructed by random Fourier features at the last decision-making layer to replace the traditional multi-layer fully connected neural networks.Taking into account the influence of ownership concentration level on the prediction results,this work conducted model training and prediction under different ownership concentration levels.The results showed that the model using the explicit kernel mapping layer had significantly higher prediction accuracy than that with traditional multi-fully-connected layers.The explicit kernel mapping process did not bring additional trainable parameters,which could be effectively simplified the model structure and improved the generalization ability of the model.In addition,the level of ownership concentration was negatively correlated with the model prediction accuracy.(3)A deep learning model named Convolutional Neural Network Module(CCAM)that integrates news event data for stock price movement prediction is proposed.This model used the structured event triples instead of the average word embedings of keywords as text representation.The CCAM model is based on event category information and attention mechanism to extract semantic features from text data.The results showed that the prediction accuracy of the model with the structured event triples as input data was higher than that with the average word embedings of keywords.It meant that the structured text representation could retain more semantic information and reduce the feature confusion between event elements,which could help to improve the prediction accuracy.Besides,the prediction accuracy of the CCAM model was higher than the CBAM model as well as the original CNN model.It indicates that the attention mechanism can effectively distinguish the impact of different events on the stock prices movement,and the fusion of prior information-event category information-into the CBAM module can further improve the semantic features extraction ability of CNN.In addition,compared with the results in Chapter 5,the model with news headline data as input had higher prediction accuracy than that with public sentiment data as input for the stocks with high ownership concentration.It meant that news headline is more suitable as a source of input data than public sentiment data for stocks with high ownership concentration.