Detecting Misreported Accounting: A Machine Learning Approach using Text Data

In this study, we aim to introduce a novel method that uses machine learning and text data to measure the probability of misreported accounting. Specifically, we apply XGBoost and other neural network classification models to the text in 10-K reports to predict the probability of a firm's misreporting financial information. We provide comprehensive validation tests for our new text-based measures of misreporting. First, firms with a higher probability of misreporting also have higher discretionary accruals. Second, firms with a higher probability of misreporting tend to have higher crash risks. Third, the probability of misreporting is positively related to the cost of equity and bank loan spreads. Last, it is positively related to stock volatility and negatively related to corporate good news. Therefore, our novel measure of text-based misreporting can be used for future research in accounting and finance.