Financial fraud detection using vocal, linguistic and financial cues

摘要

Corporate financial fraud has a severe negative impact on investors and the capital market in general. The current resources committed to financial fraud detection (FFD), however, are insufficient to identify all occurrences in a timely fashion. Methods for automating FFD have mainly relied on financial statistics, although some recent research has suggested that linguistic or vocal cues may also be useful indicators of deception. Tools based on financial numbers, linguistic behavior, and non-verbal vocal cues have each demonstrated the potential for detecting financial fraud. However, the performance of these tools continues to be poorer than desired, limiting their use on a stand-alone basis to help identify companies for further investigation. The hypothesis investigated in this study is that an improved tool could be developed if specific attributes from these feature categories were analyzed concurrently. Combining features across categories provided better fraud detection than was achieved by any of the feature categories alone. However, performance improvements were only observed if feature selection was used suggesting that it is important to discard non-informative features.