| Stock market has always been one of the hot topics in the financial field,and the fluctuation of stock price can affect the psychology of investors.Meanwhile,the stock price is in turn influenced by market participants,resulting in fluctuations in capital asset pricing.When stock prices change,investors tend to use the Internet to analyze the stock tendency and make comments or even blow off steam on relevant discussion forums according to the current economic situation.And since there are large overlap between the investors and the online forum users,the interaction effect between the stock tendency and the posts on the forum is relatively strong.Therefore,set in specific marketing environment,this disseration,after analyzing related online forum statistics,concentrates on the emotion and acting rules of those posting groups.In terms of empirical data,this disseration has made feasibility analysis on whether or not it is allowed to access the site contents both politically and technically.The forum data acquired by using the crawler technology consist two parts: First,the posting time(accurate to one minute)of total 1,814,500 posts on one certain online forum starting from November 1st,2017 to March 31 st,2019.Second,the contents of 91,328 posts released during the stock rising and falling period in 2015.According to those two parts of data,and also connecting the monthly line trend of Shanghai Composite Index released on East Money Net during the opening,we acquired the daily closing prices from the database of CSMAR Financial Database and studied relatively the sentiments and actions of the posting group during specific periods of the stock market.For the study of emotions,this disseration proposes a CNN(Convolutional Neural Network)sentiment classification model that incorporates LDA(Theme Model).It extracts the information of the poster to determine the positive and negative tendencies of the sentiment of the post.At the beginning of the experiment,taking the mentality of shareholders into account,we labeled the data in rising time as positive and those in falling time as negative.Then we added the emotion field vocabulary to the data,combining the TF-IDF(word frequency-inverse document frequency)algorithm to carries out preprocessing work.The feature words of sentiment in the post are filtered by the LDA topic model,and passed to the Le Net-5 convolutional neural network,the convolution and pooling operations are added on the basis of Le Net-5 to make the model learn features better,with an accuracy rate of 70%.Therefore,we brought up three optimized ways and a new labeling method,making comparisons to the former baseline model.With the abundant vocabulary in the sentiment word dictionaries,we firstly tried to merge both positive and negative words in three dictionaries together,and added some collected field words to the training set,and the results turned out to be better.Moreover,because there are some posts with too many words,where key words,which can represent those articles,need to be extracted,we added TF-IDF algorithm to do that,and better results could be seen.Especially after adding the LDA topic model,acquiring the words that best represent the training,such as stock market and emotion,with an increase of 20% in the accuracy.Apart from that,considering the label method can be improved,we used Snow NLP technology to score the text data emotion value to make polarity judgment,and set up value range of positive and negative sides and label the testing data with the new method of labeling.By adding the above model after changing the annotation method,the accuracy of the final model reached up to 76.45%,and acquired the optimized model combined with LDA and CNN was acquired after training.Lastly,we made an assessment of the LDA-CNN-combined sentiment analyzing model,and also made a comparison to the traditional machine learning classification algorithm.After inputting the training dataset of the forum,we can come to the conclusion that the model adapted in our experiment is better than all the other classification algorithm.We committed researches in three aspects on people's behavior.Firstly,we deeply discussed the group behavior regularity by analyzing the relevance between the closing pricing and the post count.Comparing the one decade data from 2010 to 2019 as a whole,we can see that the post count fluctuates with the stock market,with relevance rate reaching 0.6,suggesting the closing price is moderately correlated with the volume of posts.The stock market went through a cycle of gains and losses in 2015,making the year of 2015 a best time to study the spreading effect,where we can see that with the rise of the closing price,the group post count also goes up,and vice versa.To further study the relation between theclosing pricing and the post count,we divided the time into three stages: Turbulence,rise and fall,and calculate the relevance rate.Among them we found that the closing price from June12,2015 to July 10,2015 and the post count from July 10,2015 to August 7,2015 had the closest relevance,with a rate of 0.841,suggesting a high relevance between posting behavior and the closing price after a shout delay.Therefore,we can conclude that the sharp rise and fall of the stock have obvious effects on the posting of investors,and there is also a delay.Secondly,we analyzed the distribution of these interval times of the investors' posting to study the group behavior regularity.We found that the amount of the posts would affect the length of the intervals.We also found that in the ordinary coordinate system and the log-log coordinate system,most posting intervals were short.Comparing their behavior in the rise-and-fall period and the turbulence period,we could see that in the rise-and-fall period,the post count was denser,the number of posters bigger,the fat tail of the distribution presenting fatter than usual.We hold the opinion that the online group behavior is quite related to the fluctuation of the stock,and the changes in the stock market make it possible for more investors to share their emotions.Thirdly,we studied the group behavior by analyzing the post counts in different periods.On the basis of dividing three stages,we analyzed the changes in the volume of the posts in a day and in a week.After comparing that in the three different period,we found that in the falling period,the count was the largest,followed by rising period and the turbulence period.The distribution of posts per hour of the day accorded with people's daily activity pattern.To be more concrete,regardless of which stage they were in,there were always three peaks in daily post count in the opening time;while there was a little increase in the count in the evening during the cessation,compared with that in the day time.Under the circumstance of not distinguishing the market day,the post count characteristics of non-trading day would be covered,presenting generally three peaks.And during the week,the post count increased and then decreased.The post amount in 2015 was relatively high due to the influence of stock market fluctuations,while in other years,it was steadier. |
PDF Full Text Request