基于视频的欺骗检测与金融欺诈

摘要

Based on a comprehensive visual deception score constructed from China’s mandatorily disclosed IPO roadshow videos using computer vision and machine learning, we find that visual cues embedded in these videos predict IPO financial fraud. The predictive power is incremental to financial, textual, and audio indicators. Further analyses show that these cues are well grounded in psychology research and focusing on the psychology-based deception-cue-area features improves fraud predicting performance. Furthermore, decomposing each video into five topics, we find that only the deception score of the topic on financial performance is positively associated with financial fraud. We further show that insiders of firms with higher visual deception scores are more likely to sell shares shortly after lock-up periods expire, and these firms experience significantly lower subsequent stock returns. Our study demonstrates the feasibility and potential value of using video information for fraud detection in accounting and finance.