关键词: 信息异质性, 财务危机, 文本分析, 机器学习

Abstract: The financial crisis of a listed company will cause investors to suffer huge economic losses and even have a serious negative impact on the entire society. Scientific and accurate financial crisis prediction of listed companies can effectively reduce investment risks and related losses. The composition of the existing enterprise financial crisis early warning index system focuses on the use of quantitative financial statements, annual reports and news reports and other data itself carried by the information, but due to the different entities of information release, release purposes, information content and release forms, the response to the operation of listed companies may have different tendencies. This paper defines the difference in tendencies as the “information heterogeneity” between different channel data, and in order to improve the accuracy of enterprise financial crisis early warning, it is necessary to study its impact on an early warning of financial crisis of listed companies. It expands the economic value of information heterogeneity, improves the accuracy of financial crisis prediction, further supports investors for making investment decisions, and helps maintain financial system and social stability. It provides new ideas for the early warning of corporate financial crisis from a new perspective, which is of certain significance for enriching the research in this field in China.
Taking the characteristics and differences of financial crisis early warning data of listed companies from different sources as the main starting point, this paper sorts out the root causes of information heterogeneity, and deeply analyzes the relationship between information heterogeneity and enterprise crisis state. Based on this, the information heterogeneity index is defined, the measurement method of information heterogeneity is proposed, and the impact of information heterogeneity on the early warning of financial crisis of listed companies is verified. Firstly, the paper analyzes different types of data content and characteristics. It focuses on the causes and means of falsification of financial data, artificial manipulation of news texts and excessive embellishment of annual report texts, and explains the root causes of information heterogeneity. Secondly, a method to measure information heterogeneity is proposed. The performance of quantitative financial statements is scored by the power coefficient method, the sentiment value of annual reports and news texts is calculated by combining dictionary method and deep learning, and the calculation formula of information heterogeneity index is proposed to measure the information difference contained in the three data. Finally, based on the tree relationship between information heterogeneity and enterprise financial crisis status, XGBoost is selected to establish a financial crisis prediction model of listed companies, and the prediction effect of the total sample model before and after the information heterogeneity index is added and compared, and the contribution of information heterogeneity to improving the accuracy of model early warning is verified. Furthermore, the total samples are divided into four types of specific sub-samples based on the performance of financial statements and the level of information heterogeneity, the relationship between information heterogeneity and financial crisis status of specific subsamples is analyzed, the model prediction effect of information heterogeneity indicators before and after the addition of subsamples is compared, the information heterogeneity indicators that whether they could effectively distinguish crisis companies and healthy ones are verified, and the relationship between information heterogeneity indicators and financial crises is further analyzed from the perspective of feature importance.
The empirical results show that the accuracy, Recall, AUC, F1 and other indicators of the model built based on the total samples and four types of subsamples are significantly improved by using the information heterogeneity index for financial crisis prediction in the machine learning model. And information heterogeneity is the first important indicator of XGBoost model. It can be seen that information heterogeneity is a major feature to judge whether an enterprise has encountered a financial crisis, and plays an important role in distinguishing crisis companies from healthy ones.
This paper shifts the research perspective from the information contained in the data of each channel to the information heterogeneity reflected by the data from different sources, and finds that it has good value in the prediction of corporate financial crisis. In future work, it can also be applied to other research fields, such as credit risk assessment and financial fraud identification. At the same time, more other data, such as investor comments in social media, can be introduced to better optimize the financial crisis early warning model of listed companies and improve the prediction effect of the model by analyzing the heterogeneity of information.

Key words: information heterogeneity, financial crisis, text analytics, machine learning