A Tensor-based Approach For Social Media Aided Stock Market Prediction

Stock market prediction has been an active area of research for a long time.Previous works on stock market prediction mainly focus on traditional quantitative or machine learning approaches to analyze basic indicator information of a specific firm.The Efficient Market Hypothesis shows that stock price are driven by public information,however,previous words ignore two factors: financial news and investor sentiment.In this thesis,we aim to combine different types of information sources to improve the performance of stock market prediction.Although social networking data have the potential in boosting the prediction accuracy of quantitative methods,there are three problems: Traditional quantitative methods have less information which relative with stock and have less power to predict correctly;Social information is prone to be noisy,and the representation of a specific piece of news is sparse;Lack of effective prediction model for complicated data structure.To deal with the problem of information scarcity,we collected diverse information sources and explore their interaction effects to predict stock price,such as news representation and author sentiment of news,investor emotion,and features of firm,so that deal with the ineffective information problem.In above process,we use natural language process to understand the information and emotion of an article.To explore and capture the relation among different information sources,we first use tensors to model various information source features collected over time,and we propose a tensor-based information association(TIA)to learn more robust representation for stock price prediction.Specifically,TIA integrates the local neighborhood and discriminant information,derived from price movement,to map the original high-dimensional tensors to the low-dimensional ones with a set of basis matrices.Finally,we formulate the stock price prediction as a tensor regression problem,and we train a prediction model based on the higher rank tensor ridge regression algorithm to perform this.Experiments conducted on an entire year of data for stocks in China Securities Index demonstrate the effectiveness of our proposed model,and it shows higher prediction accuracy than state-of-the-arts.