A Machine Learning Analysis of Earnings Management Practices in Four East Asian Economies

Due to earnings management, firms’ true performance is distorted that misleads stakeholders’ decisions based on accounting numbers (Healy and Wahlen, 1999). Extant earnings management research mainly focuses on explaining firms’ earnings management behaviour using study variables in parametric models, but these are constrained in their predictive function. This study applies a machine learning model to predict earnings management across firms in four East Asian economies. The machine learning model performs well in prediction in terms of the model’s goodness of fit and low prediction errors. The East Asian market is chosen as a research context because of its phenomenal economic growth coupled with inadequate earnings management discussion in the literature. In particular, it is characterized by family-controlled ownership structure in firms. Using aggregate discretionary accruals as a proxy for earnings management, this study employs two commonly used earnings management detection models: the Performance Adjusted Model and the Modified Jones Model, to obtain discretionary accruals. Applying a high-dimensional machine learning model, which is implemented in TreeNet®, this research examines 54 earnings management determinants in one model to explore what are the factors that most drive earnings management behaviour in four East Asian economies: Taiwan, Indonesia, Thailand, and South Korea. The 54 earnings management determinants are grouped into five dimensions: firm-specific characteristics factors, market price variables, macroeconomic indicators, corporate governance metrics, and legal environment measures. This study finds that eight variables from the legal environment dimension, together with the categorical variable “economy”, make no contribution to earnings management prediction in both TreeNet® Models. The result is counter-intuitive in that legal environment determinants do not influence firms’ earnings management behaviour in the four East Asian economies, and there are no earnings management practices differences in firms across the four East Asian economies. The corporate governance dimension has approximately the same average relevant variables importance (RVI) score as the macroeconomic dimension in the Performance Adjusted TreeNet® Model, that is, they are the two lowest average RVI score dimensions. It has the lowest average RVI score among all the dimensions in the Modified Jones TreeNet® Model. This result provides evidence for the argument that corporate governance is inefficient when firms are family-controlled. Seven variables ‒ ROA, firm size, debt-to-equity ratio, Z score, total enterprise value, institutional ownership, and market competition ‒ rank in the top 10 RVI scores variables in both TreeNet® Models that address the drivers of earnings management among firms in the four East Asian economies. High predictive contribution variables come from different dimensions, supporting the argument that managers’ discretion on financial reporting is a collective effect of multiple factors. The partial dependence plots display that the relationship between the selected input variable and discretionary accruals is non-linear, which is consistent with the real-world context (Jones,2017). The result implies that blindly assuming a linear relationship in earnings management study may lead to an arbitrary conclusion