Building Financial Misstatement Detection Models using Multi-class Cost-sensitive Learning and Feature Generation from CFO survey

Abstract

Material misstatements are such omissions and misstatements of financial information included in the financial statements that can affect the economic decisions of the users of financial statements. When building material misstatement detection or prediction models, researchers should consider two important issues: first, financial misstatements can be classified as involving either errors (i.e., unintentional misapplications of accounting rules) or irregularities (i.e., intentional misreporting). Second, there is selection bias in target variables. Many firms that manipulate earnings are likely to go unidentified and there could be selection biases in cases pursued by financial authorities. I develop financial misstatement detection models using machine learning and statistical models addressing the two issues present in target variables. First, to address the issue of fraud intention in the target variable, I develop multi-class financial misstatement detection models to detect misstatements by fraud intention. Hennes, Leone and Miller (2008) performed post-event analysis on financial restatements and classified restatements by fraud intention (i.e. intentional vs. unintentional misstatements). Using their result (along with non-misstated firms) as a three-class target variable, I develop three multi-class classifiers, multinomial logistic regression, support vector machine, and Bayesian networks as predictive tools to detect and classify misstatements by fraud intention. To deal with class imbalance and asymmetric misclassification costs, I perform cost-sensitive learning using MetaCost. Second, one way to reduce the effect of selection bias in the target variable is to perform domain expert guided feature selection. I propose to utilize the earnings management survey to public company CFOs by Dichev et al. (2016) for feature selection and feature generation. To detect material misstatements, I create features using the survey result and build binary detection models. I compare the performance of the new models with the existing scoring models from accounting and finance literature.