Detecting Earnings Management

This dissertation discusses the challenges of earnings management research and aims to improve the understanding through the following essays.Chapter 2 examines the earnings discontinuity literature, the methods used, and identifies the research questions for Chapters 3 and 4. For the earnings changes and earnings surprises thresholds, we highlight the different design choices in the studies.Chapter 3 is a joint paper with Martin Wallmeier and Peter Fiechter published in Finance Research Letters.The dissonance in the U.S. zero earnings literature between Dechow et al. (2003), who find a decreasing discontinuity in the late 1990s, and Gilliam et al. (2015), who suggest that the Sarbanes-Oxley Act (SOX) is causally responsible for the disappearance of the kink, asks for further investigation.We identify the dotcom boom at the turn of the millennium as a possible confounding effect that may have led to mechanically more firms being in the small loss interval.After excluding firms with almost no sales, it is unclear whether the zero earnings kink has completely disappeared.More importantly, the decline in the zero earnings kink appears to be gradual rather than a sharp decline caused by SOX.Chapter 4 is a joint paper with Martin Wallmeier that is under review at the time of writing.It examines the evolution of the zero earnings discontinuity in Europe and finds that, in contrast to the U.S., the kink has remained remarkably stable.At the country level, we find that firms in environments with higher uncertainty avoidance are associated with a higher zero earnings discontinuity.The cultural dimension seems to be more important than the adoption of the International Financial Reporting Standards (IFRS).v ABSTRACT Chapter 5 evaluates three different firm-level earnings management models using the Accounting and Auditing Enforcement Releases (AAER) sample, which includes firms known to have manipulated earnings.The first model is the commonly used modified Jones model.The second model extends the first to include earnings information, and the third model is calculated using the earnings histogram.The results show that the discretionary accruals of firms committing fraud are significantly higher in the first and second models.These two models also lead to a slightly better than a random prediction of overstated firm-years.The third model does not show any superiority over random selection.Chapter 6 predicts next year's accruals.We use a large dataset including 447 explanatory variables with the standard ordinary least squares (OLS) and three supervised machine learning models, the least absolute shrinkage and selection operator (LASSO), random forest, and support vector machines (SVMs).We find that the LASSO model has the highest out-of-sample prediction accuracy.In general, the machine learning models outperform the OLS model.However, even the most accurate models still contain a relatively high prediction error, showing that accruals remain challenging to predict.vi