Unearthing Financial Statement Fraud: Insights from News Coverage Analysis

We propose a financial statement (FS) fraud detection framework, called PeerMeta, that makes improvements in all three components of the detection procedure: label measurement, feature set, and detection model. For the label measurement, prior studies mainly adopt FS fraud events that have already been disclosed and confirmed. We construct a new measure based on news coverage that can reflect unrevealed FS fraud behaviors as well. For the feature set, we innovatively add peer factors learned through the business description texts in financial reports. For the detection model, two meta-learning algorithms are applied to aggregate the 19 popular classifiers. The results indicate that the proposed method has amazingly high recall of real fraud cases announced by regulatory authorities, reaching a staggering value of 0.982. We document that all components in PeerMeta contribute to the improvements of FS fraud detection and also showcase the significant economic value of the detection framework and find that recall is more crucial for the economic value than precision.