An Evidence-Driven Hybrid Architecture for Financial Statement Fraud Detection With Long-Context Transformers

Abstract

Detecting financial statement fraud from narrative disclosures is difficult because evidence is sparse and distributed across long, structured public-company annual reports filed on Form 10-K. We present a time-aware benchmark focused on Management’s Discussion and Analysis (MD&A) for financial statement fraud detection (FSFD) and evaluate representative long-document architectures under time-aware forward-chaining splits. We compare hierarchical encoders (hierarchical attention transformer and hierarchical document transformer), a finance-domain long encoder (LongFinBERT), a sparse-attention model (Longformer/LED), and a short-context head+tail BERT baseline. Performance is assessed using the area under the receiver operating characteristic curve (AUROC), average precision, audit-budget ranking quality based on cumulative gain, and deployability metrics such as peak GPU memory and per-filing latency. We further propose an evidence-driven Hybrid pipeline (Selector → Reranker → Calibrated Stacker) that selects paragraph evidence with high recall, applies cross-evidence reranking, and fuses calibrated scores with a monotone stacker. On the held-out 2014–2019 horizon, the Hybrid achieves AUROC = 0.871 and average precision = 0.392 and improves audit-budget ranking over the strongest single model. In our benchmark, hierarchical encoders offer the strongest single-model accuracy–efficiency trade-off, outperforming the sparse-attention baseline at lower memory and latency, while head+tail BERT remains a competitive low-cost anchor. Year- and issuer-stratified block bootstrap analysis indicates that gains in average precision and cumulative-gain ranking quality are robust under our time-aware evaluation protocol.