171 / 2024-04-22 14:54:56

Accounting Fraud Detection through Decoding Corporate Event Sequence: A Novel Deep Learning Framework

Accounting fraud,Fraud detection,Corporate event sequence,Deep learning,Event embedding

The corporate event sequence, which describes a firm’s daily business events comprehensively and dynamically, can effectively help discover abnormal behaviors indicative of accounting fraud. However, it is rarely considered in fraud detection so far. This study explores whether corporate event sequences can provide valuable information to improve fraud detection accuracy. It is difficult for existing methods to handle these event sequences, in which event types are diverse and time intervals between successive events vary. Thus, we propose a novel deep learning model, which can process diverse event types and random occurrence time of events, and output a fraud probability for each sequence. Based on 43,541 firm-year observations of Chinese listed firms from 2010 to 2022, the empirical analysis shows that event sequence can significantly improve out-of-sample AUC performance of fraud detection from 72.84% to 76.18%, based on commonly-used financial variables. Moreover, event sequence can provide earlier fraud warnings than financial variables as prediction time horizon extends. Lastly, the abnormal events related to firms’ fraudulent activities are identified from the event sequences by our model’s attention mechanism. This study enlightens that investors and regulators should pay attention to the firm’s daily events in detecting accounting fraud.