Research On Stock Recommendation Based On User Expertise Analysis Of Online Investment Community

In the era of Web2.0 and big data,the online investment community(OIC)has become the main network platform for investors to share and exchange investment opinions and suggestions.The massive amount of user-generated content(UGC)provides rich data and a novel research perspective for analyzing the investors' behavior and mining the value of the online investment community.Given the tremendous volume of user investment opinions,a significant research problem with practical application value is to provide stock decision support by integrating the investment opinions based on the “wisdom of crowds” theory.However,the effectiveness of opinion aggregation is likely to be affected by a large number of low-quality or misleading opinions published by non-experts or malicious users,so the value mining of the online investment community is still facing great challenges.In this context,the thesis takes the users' expertise as the breakthrough point to alleviate this problem.Based on the users' investment opinions of the Stocktwits platform,the thesis is organized as follows: First,how to assess user expertise more accurately,and then how to integrate users' opinions of different expertise levels more effectively to provide investment recommendations,and finally how to make investment recommendations more helpful for users.Overall,this thesis studies the stock recommendation problem based on the users' opinions of the online investment community.The main research contents and contributions of this thesis include:1.To resolve the defects of the existing user expertise assessment methods,this research proposes a novel method called correlation-based robust dynamic expertise(CBRDE)to model the expertise of users more accurately by considering the correlations among stocks and verify the effectiveness of the method based on the performance of the crowd wisdom and stock recommendation.CBRDE models the correlations among stocks based on the past expertise distribution of all investors' past opinions over stocks.In this way,the correlations can reflect how helpful a user's performance about a stock is for inferring his/her performance about other stocks.Different from the existing static expertise assessment method which is based on the overall performance of a user,CBRDE gives different weights to a user's historical opinions about different stocks.Compared with the method of only using the performance of investors' historical opinions on a particular stock to evaluate their expertise level of the stock,CBRDE makes full use of all the available historical opinions of users.The experimental results show that,compared with the traditional methods,CBRDE has better performance with different parameters(number of stocks recommended,length of holding periods,etc.)and evaluation methods(portfolio evaluation metrics,group performance),which indicates that CBRDQ is a more reliable and accurate dynamic expertise assessment method.The findings demonstrate that for the stock recommendation utilizing the user-generated content of the online investment community,it is beneficial to notice the variability of investors' expertise in different stocks.This study provides important implications for trading practice and judgment of the quality of users' opinions in the online investment community.2.The thesis proposes a more flexible and intelligent user opinions aggregation method(Follow All Kinds of Investors,FAKOI)based on the machine learning technology,and applies it to the stock recommendation.In the existing literature,opinion aggregation methods with equal weight or expertise weight are based on experience and artificially defined rules,which can not effectively reflect the complex relationships between opinions posted by different levels of users and group performance.To solve this problem,for the first time this thesis applies machine learning technology to the task of opinion aggregation.Firstly,users are classified into different groups based on their expertise,and the derivative features of their opinions are constructed.Then,a function of opinion aggregation quality about the derivative features of different groups is learned by using machine learning technology,and the user opinions of different expertise levels are integrated into stock investment decisions more flexibly and intelligently.The experimental results show that the performance of stock recommendation based on machine learning integration method is better than that of equal weight and expertise weighted integration method,which verifies that machine learning technology is a more effective method of group opinion aggregation than the method of artificially defined rules.3.Based on the users' behavior data of the online investment community,this thesis proposes a novel stock movement aware personalized preference modeling method and explores the possibility of improving the fit degree between the recommended stock list and investors' personalized preference on the basis of high-quality stock recommendation.A large number of researches are devoted to providing high return potential stock recommendations while ignoring the personalized investment preference of investors.A few researchers use traditional recommendation algorithms to realize personalized recommendation,but they can not effectively use the specific factors of the stock market that affect investors' preference and ignore the return potential of recommended stocks.Compared with traditional methods,the personalized preference modeling method proposed in this thesis can capture the stock movement patterns preferred by users and find users with similar trend preferences or investment philosophies.On this basis,this thesis innovatively proposes a hybrid recommendation algorithm that takes into account both investors' personalized preferences and stock return potential and explores the influence of high-quality and personalized factors on the recommendation performance.The experimental results show that the personalized preference modeling method outperforms the traditional collaborative filtering algorithm based on users or items,and preliminarily prove that the hybrid recommendation algorithm can improve the personalized recommendation effect with little loss of revenue potential.