有效预警上市公司违规的递延所得税异动指标和人工智能模型

2000-2017年,3434家A股上市公司样本中的47.79%至少有一次违规记录,每年平均有17%的公司违规,而监管机构平均需要2.7年查证并通告违规行为。我们用当年数据构建递延所得税异动指标,可有效预判违规,并发现监管机构未能识别这一指标的警示作用,实际激励了违规公司通过操纵递延所得税提高财务指标以规避稽查。进一步构建决策树模型,对违规事件实现样本外精准判别。本文揭示了A股公司大面积违规而稽查过程冗长迟缓这一重要现象,并分析了违规机制,指出了所得税数据在稽查违规中可以发挥的预警作用,为监管者和投资者提供有效预警上市公司违规行为的新指标和方法。

Fraud Detection Using Deferred Income Tax: A Machine Learning Perspective

Information disclosure significantly affects listed companies' market value and investors' decisions. After years of practice, regulators have recognized that the public listing system should focus on information disclosure, rather than evaluating issuance quality, which should be left for the market to assess. Furthermore, the deferred income tax data associated with corporate profit and management flexibility has become an arena for gaming between regulators and listed companies.
This paper highlights the important role of deferred income tax in the early alerting of fraudulent activities. It reveals what motivates companies to abuse their deferred income tax data, both theoretically and empirically. A sample of A-share companies from 2000 to 2017 is analyzed, controlling for partial observability and endogeneity problems. The deferred income tax indicator is significantly positively correlated with the tendency of listed companies to commit fraud and significantly negatively correlated with the probability of fraud detection. This result implies that regulators focus on monitoring the financial indicators of A-share companies, but fail to recognize the relationship between deferred income tax data and corporate disclosure fraud. Such oversight may incentivize listed companies to manipulate deferred income tax data to boost their financial performance. According to Spence's signaling theory, listed companies transmit information about their corporate value to the market through signals, such as financial performance. Thus far, deferred income tax adjustment has served as a convenient tool for listed companies in managing earnings at low costs. It constitutes a “soft signal”, which bears a low threshold to mimic but conveys less reliable information on companies' fundamental value.
Three factors motivate financial fraud: financial distress, avoidance of negative performance reporting and excessive valuation upon corporate growth. The first two reflect competitive pressures and the third reflects the rewarding effect of the capital market. Separating accounting and tax standards yields differences between pre-tax accounting profits and taxable income. Factors such as the confirmation time and amount of deferred income tax assets fall within the scope of management discretion. This accounting-tax discrepancy provides an opportunity for companies to abuse their deferred income tax data for earnings management.
This paper makes three contributions to the literature. First, it reveals the anomaly of extensive occurrences of disclosure fraud among A-share companies and the lagging regulatory inspections upon fraud. Furthermore, it offers an effective solution based on a solid analysis of the mechanisms underlying fraud. The literature on disclosure fraud detection is limited. Such academic neglect may explain the regulatory insufficiency regarding corporate fraud and the lack of public monitoring of this regulatory insufficiency. Second, this paper demonstrates the informational value of income tax data in detecting disclosure fraud and its “hard signal” value in revealing the true financial status of a listed company. Third, it showcases the role of machine learning in improving capital market governance.
The Guotaian CSMAR database conveys information on public announcements of regulatory measures on corporate disclosure fraud from authorities. The financial data of A-share listed companies is from the Wind database. A binary probit model is constructed, controlling for endogeneity and partial observability, to analyze the relationship between the deferred income tax indicator and companies' tendency to commit fraud. In the interest of accurately identifying companies with instances of fraudulent activity, a decision tree model is built to implement out-of-sample prediction and form investment portfolios based on the prediction results. Among the out-of-sample firms, the decision tree model identifies 39% of the fraud-free firms with 95% accuracy and identifies 2% of the fraud-committing firms with 100% accuracy. The annual return on assets of the fraud-free firms is 2.80% higher than that of the fraud-committing firms, and the average annual stock return of the fraud-free firms is 3.69% higher than that of the fraud-committing firms.
This paper analyzes the mechanism underlying financial manipulation, shedding light on the predictive power of the deferred income tax indicator in detecting disclosure fraud. For regulators, the timely identification of disclosure fraud is of great significance for combating violations and maintaining healthy stock market development. For investors, risk management is a primary concern. Avoiding companies with major alleged fraud (e.g., LeTV) is a primitive demand for risk control. The decision tree model in this paper provides an analytic framework for assessing the likelihood of listed companies to commit fraud, with parameter settings that address the accuracy and coverage of such assessment.