A Hybrid Transformer and GNN Framework for Interpretable Fair Value Classification in Accounting

Abstract

Fair value classification in accounting requires accurate assessment of financial instruments across three hierarchical levels, where Level 1 represents quoted market prices, Level 2 involves observable market inputs, and Level 3 relies on unobservable inputs with significant estimation uncertainty. Traditional classification methods struggle with complex financial instruments that exhibit ambiguous characteristics across multiple levels, leading to inconsistent valuations and regulatory compliance issues. This study proposes a hybrid framework combining Transformer architecture with Graph Neural Networks (GNNs) for interpretable fair value classification. The Transformer component captures sequential patterns in financial instrument characteristics and market conditions, while the GNN component models relationships between related financial entities and market participants. An interpretability module provides transparent classification reasoning, enabling auditors and regulators to understand the decision-making process. Experiments on real-world financial datasets demonstrate that the proposed framework achieves superior classification accuracy compared to traditional methods while maintaining high interpretability. The integration of attention mechanisms and graph-based learning enables comprehensive analysis of financial instrument complexity, market liquidity, and valuation uncertainty, resulting in more reliable fair value classifications that align with accounting standards and regulatory requirements