Can Regulation be Identified?Evidence from Machine Learning and Inquiry Letters
Scientific regulation and intelligent regulation are important means to improve the modern financial regulation system,but the focus of regulation is always a black box for enterprises,investors and the public.The research goal of this paper is to construct an out-of-sample prediction model based on the multi-dimensional characteristics of listed companies to identify whether the company will be subject to regulatory inquiry,and try to find out those who have strong ability to identify regulatory inquiry behavior and analyze its recognition model.The evidence shows that:(1)Among different machine learning methods,the ensemble learning method represented by stochastic gradient boosting tree and random forest has the best prediction effect on the company's receipt probability,and its prediction ability is significantly better than the logical regression method commonly used in previous studies.(2)The predictive ability of corporate financial characteristics to the regulatory inquiry behavior is better than that of corporate internal or external governance characteristics and the text characteristics of the annual report,which indicates that the exchange mainly examines the company's financial information when carrying out regulatory inquiry;(3)In the multi-dimensional characteristics of the company,the number of years a company has been listed,profitability,the going concern ability,the abnormal impairment loss,managerial ownership and the quality of internal control have the best recognition effect on the company's regulatory inquiry.Specifically,companies with short listing years,high management shareholding and poor internal control quality are more likely to receive inquiry letters;the company's probability of receiving inquiry letters shows a stepwise downward trend with the increase of profitability or sustainability indicators;the asset impairment loss ratio has a V-shaped relationship with the company's acceptance probability of receiving inquiry letters.Heterogeneity analysis shows that the focus of exchanges in different time periods is different,and the focus of different exchanges is also different;(4)Previous studies have found that some characteristics such as political connections)that have strong explanatory power for the regulatory behavior of inquiry letters have poor performance in predicting the probability of companies receiving letters,which reflects the difference between explanatory and predictive questions.The conclusion shows that machine learning method has advantages in identifying the regulatory behavior of exchange inquiry letters,and has important implications for the goal of'strengthening the use of regulatory technology'proposed in the 14th Five-Year Plan.And by evaluating and comparing the prediction effect of a series of machine learning methods on whether the company receives inquiry letter,this paper broadens the research scope of machine learning in the field of accounting and finance,and also provides ideas for the subsequent use of machine learning methods.
NETL
NSTL
万方数据
