Evaluating Machine Learning Algorithms in Predicting Financial Restatements

Abstract

The identification of financial statements which were willfully or accidentally misstated is important for all involved parties: Investors can expect improved returns, analysts preserve their reputation and auditors avoid costly litigation. In this paper, we chose six state-of-the-art machine learning methods which we analyze in their ability to detect misstatements. In addition to that we investigated the influence of a FeatureBoost algorithm, namely XG-Boost to all of the six machine learning methods. The underlying data is retrieved from Eikon, a financial database provided by Refinitiv (former provided by Thomson Reuters). In order to take out our experiments we chose about 9000 US-companies and 757 features per year over ten years. We offer six definitions of ground truth of which three can be calculated with the data extracted from the Eikon database. The other three definitions are created with the help of an external data source provided by Audit Analytics Europe. Our well structured results give an overview on the performance of current machine learning methods in order to identify misstatements.