Home

Welcome to the ML_AccountingFraud wiki!

In this paper, we examine the effectiveness of using machine learning (ML) techniques to develop fraud prediction models. We identify methodological issues in the recent papers which use ML methods to predict accounting irregularities; notably, we investigate the role of multicollinearity in the variables used as predictors by prior research. We then examine eight ML techniques to predict fraud. We find that a set of ML methods that are new to the accounting literature substantially outperform the effectiveness of the model used by prior studies. The results show that an adaptive boosting approach combined with logistic regression has the highest sensitivity and can detect AAERs up to 4 years earlier than SEC does.