Research On The Prediction Of Text Writing Style And Fund Performance Based On Machine Learning Method

We found that when investors focus on information disclosure in the current field of fund research,they often focus on the explicit information of the fund,such as the financial condition of the fund,company information,identity characteristics of the fund manager,etc.,but the large amount of text data in the information disclosure actually also contains incremental information about the future performance of the fund,which can also be considered as some implicit information.With the development of computer technology,information mining and analysis of massive text is now achievable.Many domestic and foreign scholars have verified the usefulness of text information in decision-making from an empirical perspective,such as annual reports,prospectuses,and performance briefings,but there are still few related research on text analysis of performance forecasts.Some people analyze the future development trend of the stock market or company from financial news,forum posts,video bullet screens,comments and so on.However,considering that the information are affected by false information and group effects to a certain extent,social news and panic mentality will cause changes in public opinion of the financial market,and the current relevant supervision departments cannot monitor the public opinion dynamics of the financial market in time and guide its development correctly,which will cause distortion of information in the process of information transmission in the financial market.Investors will be disturbed by these signals,affecting their accurate judgment,resulting in unnecessary losses.It can be said that robust,reliable and effective text information has become a scarce resource for current fund market research.Compared with financial news,the fund company’s annual report "brief outlook of macroeconomic,securities market and industry trend" and the detailed plan of the fund company’s entire business situation and future development strategy in the whole annual report information disclosure,not only requires professional auditors to investigate,but also has the work personnel of the supervision department to review the content of the company’s annual report multiple times,thus ensuring the accuracy and effectiveness of the content of the annual report.In addition,the fund annual report has a good fixed format and mandatory release,which is very suitable for academic research.Therefore,based on the text content of the "Outlook" part of the fund annual report from 2010 to 2020,this paper constructs the multi-semantic layered sentiment tone index of the fund manager through the dictionary method,and tries to explore the predictive ability of the fund manager’s writing style on the fund performance.The main research contents are as follows: this paper attempts to mine the subtle emotional changes and the content conveyed in the writing style of the author from the annual report text regularly released by the fund.On the basis of the fund annual report as the research corpus,we preprocessed the data with PYTHON related libraries,and used unsupervised machine learning algorithms and manual selection to generate sentiment dictionaries.Our dictionary is not only divided into positive and negative sentiment words,but also further divided into three layers to express the confidence level: determinacy,expression intensity and emotional exaggeration.We can use these three standards to measure the confidence level expressed in the open letter written by the fund manager to the investors.Finally,we quantified the confidence level through word embedding technology and tone measurement method TF-IDF,and used it as a feature to be integrated with the financial feature vector and added to the training and prediction process of the model in the form of vector to study whether the fund manager’s sentiment tone in the text information can significantly improve the prediction ability of the fund performance.Finally,this paper draws many important conclusions related to the confidence level of the fund manager: the prediction effect of the fund performance based on the semantic layered sentiment score is significantly improved compared with the traditional positive and negative dictionaries.Our research shows that measuring the confidence level of fund text is fundamentally different from the polarity intensity of confidence perceived by people in communication,and the writing style contains important forward-looking information.On this basis,we applied the score to the selection of investment portfolios,and the results showed that the investment portfolios selected by the semantic layered sentiment score obtained excess returns.