The Enigma of Unlabeled Positives - Using Semi-Supervised Learning to Detect and Predict Unlabeled Restatements

This paper introduces a semi-supervised machine learning technique, Positive-Unlabeled (PU) Learning, to address the issue of “unlabeled positives” in accounting research. It uses financial statement restatement detection/prediction as a research setting. In this setting, unlabeled positives are audited financial statements that were not identified to be restated when data was collected but were subsequently restated. Extant research in restatement detection/prediction uses supervised learning, which treats unlabeled positives as clean financial reports. This improper handling of unlabeled positives in machine learning studies may affect model performance in restatements detection/prediction. Our analysis shows that PU Learning is significantly superior to supervised learning in detecting unlabeled restatements from past financial statements and predicting restatements from newly issued financial statements based on different metrics. In particular, the best-performing PU Learning method can outperform the benchmark supervised learning model by 113.2% (110.6%) in terms of the recall ratio for detection (prediction) and 34% (51%) in F2 Score. The methodology introduced in this study can also be applied to detecting or predicting other accounting outcomes that have unlabeled positives, such as fraud and accounting misconduct