Discriminating Fradulent Financial Statements by Identifying Linguistic Hedging

Managerial financial fraud is estimated in the billions of dollars annually in the United States. Since fraud includes obfuscation, misdirection, and fabrication, this study proposes using deception theory as a means of detecting fraud in textual portions of financial statements (10K). A corpus of 101 fraudulent 10Ks was collected from the Securities and Exchange Commission along with 101 matching non-fraudulent 10Ks. Natural Language Processing techniques were applied to the corpus to generate raw counts and usage rates of hedging devices: hedging modal verbs, hedging adjectives, hedging adverbs, hedging conjunctions, hedging nouns, and hedging lexical verbs. A classification model, based on logistic regression, successfully discriminates with 69.3% accuracy and accounts for nearly 20% of the observed variance. Two machinelearning algorithms are investigated. Bayesian Network and JRip achieve accuracy results of 62.4% and 67.8% respectively. Both results are better than chance or of human deception detection suggesting the possibility of a diagnostic tool for auditors.