Detecting financial misstatements with fraud intention using multi-class cost-sensitive learning

• Multi-class financial misstatement detection models are developed.
• The models classify financial misstatements according to fraud intention.
• MetaCost is employed to perform cost-sensitive learning in a multi-class setting.
• Features are evaluated to detect fraud intention and material misstatements.

We develop multi-class financial misstatement detection models to detect misstatements with fraud intention. Hennes, Leone and Miller (2008) conducted a post-event analysis of financial restatements and classified restatements as intentional or unintentional. Using their results (along with non-misstated firms) in the form of a three-class target variable, we develop three multi-class classifiers, multinomial logistic regression, support vector machine, and Bayesian networks, as predictive tools to detect and classify misstatements according to the presence of fraud intention. To deal with class imbalance and asymmetric misclassification costs, we undertake cost-sensitive learning using MetaCost. We evaluate features from previous studies of detecting fraudulent intention and material misstatements. Features such as the short interest ratio and the firm-efficiency measure show discriminatory potential. The yearly and quarterly context-based feature set created further improves the performance of the classifiers.