Cross-Period Financial Statement Anomaly Detection via Contrastive Representation Learning

Abstract

To address the challenges of concealed anomaly patterns, easily drifting distributions, and scarce anomaly samples in cross-period financial statements, this paper proposes a cross-period anomaly pattern recognition framework based on contrastive learning. The method takes multi-period financial characteristics of an enterprise as input. First, it performs caliber alignment and standardization to improve cross-period comparability. Then, it constructs cross-period sample pairs and learns stable low-dimensional representations through an encoder and projector, ensuring that reasonable continuations between adjacent periods of the same entity remain close in the representation space and are differentiated from inconsistent patterns. Simultaneously, a cross-period smoothing constraint is introduced to suppress representation jitter caused by short-term noise, enhancing the continuity and robustness of the representation. In the detection phase, anomaly scores are constructed based on a joint characterization of adjacent period consistency and historical reference deviation, enabling the ranking and screening of risk levels for each reporting period of the enterprise. Comparative evaluation results show that the proposed framework has stronger comprehensive discrimination capabilities under a unified indicator system and achieves a more reasonable balance between detection capability and false alarm control, making it suitable for automated risk screening and auxiliary verification scenarios for large-scale cross-period financial statement data.