Sentiment Mining Algorithm Of Annual Report Of Listed Companies Based On Multi-kernel Learning

Annual report of listed company is a comprehensive report on the company's major activities and operations in the previous fiscal year,providing a detailed disclosure of the company's business,risk,operations status,financial conditions,and profit drivers.The listed companies' annual report is the major sources for investors to analysis the company's performance.However,80% of the annual report is in the form of text,and the sentiment in these text reflects the prospects of the company's management in companies future development.This emotional change is highly related to the company's performance and stock price movement.How to use the machine learning approach to analysis annual reports' emotional characteristics is very crucial for stock price forecasting.The existed works mainly studied a different quantitative financial data to predict stock price movement.However,in recent years,more and more studies have found that sentiment analysis on financial text can be effectively used to predict stock price movement.Although lots of work having proved the effectiveness of the measurement method for the emotional characteristics of the annual report,it is still necessary to explore an innovative measurement method,which can integrate with other heterogeneous features in the listed company's annual report to more accurately predict the stock price movement..The paper used the Hong Kong listed companies' annual report as the analysis target.It verifies the effectiveness of the emotional tone change in the listed company's annual report for the stock trend prediction,and analyzes Hong Kong financial market's sensitive event window on the sentiment information of the listed company's annual report,as well as the performance of the emotional characteristics of the annual report during the financial crisis.Then,the paper proposes a new emotional tone change vector,which is designed to measure the emotional change in the financial annual report.Moreover,the concept of emotional word weight in linguistics is introduced,and combined with the traditional vocabulary weight concept.An innovative emotional smoothing kernel is designed.Finally,a multi-kernel algorithm framework based on emotional smoothing kernel and diverse heterogeneous features is proposed.The financial text emotional characteristics are introduced into the financial multi-kernel learning prediction framework,and the financial text emotional characteristics are combined with the financial quantitative data characteristics to comprehensively analyze the market conditions.The experimental results show that the research conclusions of North American financial market also works for Hong Kong financial market.The emotional smoothing kernel proposed in this paper is especially suitable for the study of emotional characteristics contained in emotional tone change vector.The multi-kernel algorithm framework is used in stock trend forecasting,whose performance is significantly better than the traditional multi-kernel algorithm framework.