Huijie Fan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Huijie Fan has a total of 3 co-authors. Here are the top co-authors by number of shared publications:
- Wenjie Ping
Georgetown University, Washington, DC, USA
1 shared publication - Xuguang Zhang
University of Gloucestershire, Cheltenham, U.K.
1 shared publication - Yanan Jiao
Long Island University, Brooklyn, NY, USA
1 shared publication
Latest Publications
- Open Access
- 27 August 2026
- Published by Elsevier in SSRN Electronic Journal
Artificial Intelligence-Enabled Detection of Financial Risk in U.S. Banks: Longitudinal Evidence from SEC 10-K Filings
Annual bank filings contain forward-looking descriptions of credit, liquidity, regulatory, operational, and market risks, but most computational approaches encode each filing independently. This paper studies whether changes in mandatory risk-factor disclosure provide incremental information about a bank's contemporaneous solvency condition beyond prior financial ratios and static text. We introduce LRDF (Longitudinal Residual Disclosure Fusion), a sparse, auditable model for SEC Form 10-K Item 1A. LRDF represents the current filing with TF-IDF, decomposes year-over-year textual change into two nonnegative channels for newly emphasized and de-emphasized terms, fuses the channels with eight lagged bank key performance indicators (KPIs), and estimates the residual change in normalized Z-score relative to the previous year. The design separates three effects that are confounded in conventional text regression: the current disclosure level, the direction of disclosure drift, and persistent financial state. Experiments use the public FETILDA U.S.-bank dataset: 2,479 Item 1A observations from 369 banks over 2006-2016. We preserve chronological ordering with 1,983 training, 248 validation, and 248 test observations and report MSE, RMSE, MAE, R 2 , Spearman, and Kendall statistics. On the held-out test period, LRDF obtains an MSE of 0.028470, compared with 0.030118 for an eight-KPI ridge model, 0.031800 for current-file TF-IDF, 0.029838 for TF-IDF plus lagged Z-score, and 0.028666 for current text plus the full lagged financial history. Relative to the historical-KPI and pure-text baselines, bank-cluster bootstrap tests yield positive improvement probabilities of 0.9969 and 0.9992, respectively. A separate bank-level holdout gives a 1.23% MSE reduction and improves Spearman correlation from 0.4732 to 0.4964 versus the matched direct-fusion baseline. Ablations show that current text, lagged KPIs, directional drift, and residual learning each contribute distinct predictive information. The resulting model is computationally light, directly reproducible on public data, and provides a technically explicit alternative to treating successive 10-Ks as independent documents.
- 27 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessArtificial Intelligence-Enabled Detection of Financial Risk in U.S. Banks: Longitudinal Evidence from SEC 10-K Filings
Annual bank filings contain forward-looking descriptions of credit, liquidity, regulatory, operational, and market risks, but most computational approaches encode each filing independently. This paper studies whether changes in mandatory risk-factor disclosure provide incremental information about a bank's contemporaneous solvency condition beyond prior financial ratios and static text. We introduce LRDF (Longitudinal Residual Disclosure Fusion), a sparse, auditable model for SEC Form 10-K Item 1A. LRDF represents the current filing with TF-IDF, decomposes year-over-year textual change into two nonnegative channels for newly emphasized and de-emphasized terms, fuses the channels with eight lagged bank key performance indicators (KPIs), and estimates the residual change in normalized Z-score relative to the previous year. The design separates three effects that are confounded in conventional text regression: the current disclosure level, the direction of disclosure drift, and persistent financial state. Experiments use the public FETILDA U.S.-bank dataset: 2,479 Item 1A observations from 369 banks over 2006-2016. We preserve chronological ordering with 1,983 training, 248 validation, and 248 test observations and report MSE, RMSE, MAE, R 2 , Spearman, and Kendall statistics. On the held-out test period, LRDF obtains an MSE of 0.028470, compared with 0.030118 for an eight-KPI ridge model, 0.031800 for current-file TF-IDF, 0.029838 for TF-IDF plus lagged Z-score, and 0.028666 for current text plus the full lagged financial history. Relative to the historical-KPI and pure-text baselines, bank-cluster bootstrap tests yield positive improvement probabilities of 0.9969 and 0.9992, respectively. A separate bank-level holdout gives a 1.23% MSE reduction and improves Spearman correlation from 0.4732 to 0.4964 versus the matched direct-fusion baseline. Ablations show that current text, lagged KPIs, directional drift, and residual learning each contribute distinct predictive information. The resulting model is computationally light, directly reproducible on public data, and provides a technically explicit alternative to treating successive 10-Ks as independent documents.
- Top Cited PapersOpen Access
- 20 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 80456-80468
Multimodal Fraud Detection in Financial Statements: A Trimodal Attention Network With Contrastive Evidence Chain Construction
Financial statement fraud inflicts enormous economic damage and continues to evade conventional detection methods that rely predominantly on structured accounting ratios. Such approaches discard the rich forensic signals present in narrative disclosures and document formatting—two channels that fraudsters routinely manipulate alongside the numbers themselves. This paper introduces TriAudit, a trimodal deep learning framework that jointly encodes three complementary evidence streams from corporate 10-K filings: (i) structured financial ratios and accrual indicators processed through a residual MLP, (ii) MDA textual semantics captured via a FinBERT-based encoder with cross-chunk attention, and (iii) visual document-layout features extracted using a LayoutLMv2 backbone augmented with a novel Layout Deviation Score that quantifies spatial anomalies against industry-normative templates. These modality-specific representations are integrated through a hierarchical cross-modal attention mechanism—mirroring the auditor’s evidential reasoning from numerical signals through narrative context to document structure—followed by a gated fusion module that adaptively weights each stream based on perfiling reliability. To support auditability, TriAudit incorporates a contrastive evidence-chain module that aligns anomaly scores across modalities into structured, human-interpretable audit trails linking specific financial irregularities, linguistic deviations, and layout inconsistencies. Evaluated on over 6,000 filings from the EDGAR Corpus with fraud labels derived from SEC Accounting and Auditing Enforcement Releases, TriAudit achieves an AUC-ROC of 0.923 and F1 of 0.847, surpassing the strongest bimodal baseline by 5.5 AUC points and the classical Beneish M-score by over 18 points. Ablation experiments confirm that each architectural component—including the layout deviation scoring, hierarchical attention ordering, contrastive alignment, and orthogonality regularization—contributes meaningfully to both predictive accuracy and interpretability.
- 20 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 80456-80468
Top Cited PapersOpen AccessMultimodal Fraud Detection in Financial Statements: A Trimodal Attention Network With Contrastive Evidence Chain Construction
Financial statement fraud inflicts enormous economic damage and continues to evade conventional detection methods that rely predominantly on structured accounting ratios. Such approaches discard the rich forensic signals present in narrative disclosures and document formatting—two channels that fraudsters routinely manipulate alongside the numbers themselves. This paper introduces TriAudit, a trimodal deep learning framework that jointly encodes three complementary evidence streams from corporate 10-K filings: (i) structured financial ratios and accrual indicators processed through a residual MLP, (ii) MDA textual semantics captured via a FinBERT-based encoder with cross-chunk attention, and (iii) visual document-layout features extracted using a LayoutLMv2 backbone augmented with a novel Layout Deviation Score that quantifies spatial anomalies against industry-normative templates. These modality-specific representations are integrated through a hierarchical cross-modal attention mechanism—mirroring the auditor’s evidential reasoning from numerical signals through narrative context to document structure—followed by a gated fusion module that adaptively weights each stream based on perfiling reliability. To support auditability, TriAudit incorporates a contrastive evidence-chain module that aligns anomaly scores across modalities into structured, human-interpretable audit trails linking specific financial irregularities, linguistic deviations, and layout inconsistencies. Evaluated on over 6,000 filings from the EDGAR Corpus with fraud labels derived from SEC Accounting and Auditing Enforcement Releases, TriAudit achieves an AUC-ROC of 0.923 and F1 of 0.847, surpassing the strongest bimodal baseline by 5.5 AUC points and the classical Beneish M-score by over 18 points. Ablation experiments confirm that each architectural component—including the layout deviation scoring, hierarchical attention ordering, contrastive alignment, and orthogonality regularization—contributes meaningfully to both predictive accuracy and interpretability.