Positive-Unlabeled Learning for Financial Misstatement Detection under Realistic Constraints

Abstract

Detecting financial misstatements is critical for market integrity but remains challenging due to class imbalance, delayed discovery, and limited labeled data. We propose a novel Positive-Unlabeled (PU) learning framework that models the detection task under realistic constraints, where only a small subset of misstatements is known at training time. Our approach integrates unlabeled data into training, preserves temporal structure, and accounts for extreme imbalance. We construct and release a benchmark dataset reflecting these characteristics and evaluate several PU learning methods against recent baselines. Results show that PU-based models consistently outperform supervised approaches, highlighting their suitability for real-world misstatement detection.