Wenjie Ping

Wenjie Ping

Auto-generated profile, based on publicly available information

Wenjie Ping

Georgetown University, Washington, DC, USA

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 January 2026 in IEEE Access
Verified
Journal Article
Open access
YLYiwen Liang
Yiwen Liang
YJYanan Jiao
Yanan Jiao
WPWenjie Ping
Wenjie Ping
HFHuijie Fan
Huijie Fan
XHXu Han
Xu Han

20 May 2026 in IEEE Access
Verified
Journal Article
Open access
WPWenjie Ping
Wenjie Ping
YJYanan Jiao
Yanan Jiao
HFHuijie Fan
Huijie Fan
XZXuguang Zhang
Xuguang Zhang
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.

11 June 2026 in IEEE Access
Verified
Journal Article
Open access
ZCZifan Chen
Zifan Chen
MWMengdie Wang
Mengdie Wang
ZZZhuoqi Zeng
Zhuoqi Zeng
WPWenjie Ping
Wenjie Ping
Financial forecasting presents a critical challenge in risk management, where traditional point estimation approaches fail to capture the inherent uncertainty and volatility of financial markets. While deep learning models have demonstrated remarkable predictive capabilities, they typically provide deterministic predictions without quantifying prediction confidence or risk boundaries. This paper introduces a practical uncertainty-aware forecasting framework that integrates Conformal Prediction (CP) with deep recurrent neural networks to generate statistically valid prediction intervals alongside point estimates. Our approach leverages Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures as base learners, enhanced with adaptive conformal inference mechanisms that provide approximate marginal coverage under realistic market dynamics. We propose three key design contributions: (1) a volatility-adaptive nonconformity measure that adjusts prediction intervals based on market regime detection, (2) a temporal split conformal framework specifically designed for financial time series with autocorrelation structures, and (3) an extreme event awareness mechanism that dynamically widens intervals during high-uncertainty periods. Comprehensive experiments on multiple stock market datasets spanning multiple years demonstrate that our framework achieves superior calibration (95.2% empirical coverage at 95% target level) while maintaining competitive predictive accuracy (RMSE reduction of 8.3% compared to baseline models). Notably, our method exhibits enhanced early warning capabilities for extreme market events, achieving 87.6% detection rate for significant price movements with controlled false positive rates. We additionally evaluate the framework on cryptocurrency data and with multiple modern backbone families across the same 10-stock subset, providing empirical evidence that VAC-FF is compatible with diverse base predictor architectures and asset classes. The proposed framework offers financial practitioners a transparent, statistically rigorous tool for risk-adjusted decision-making, bridging the gap between predictive accuracy and uncertainty quantification in high-stakes financial applications.

13 November 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
WPWenjie Ping
Wenjie Ping
YJYanan Jiao
Yanan Jiao
Financial misstatements rarely occur in isolation: subsidiaries, blockholders, shared insiders and related-party counterparties link reporting entities, and misreporting spreads along such ties. We examine whether inter-entity relationships carry misstatement-risk information beyond financial ratios and how robust it is to the frictions of extracting relationships from SEC disclosures. We build a typed, time-stamped Relational Risk Graph over 141,463 U.S. firm-years (1991-2014) with AAER labels, release dates and an SEC CIK-GVKEY registry, and derive leakage-safe structural, enforcement-exposure and anomaly-propagation features. On real registry and EDGAR colocation relations, a naive comparison raises out-of-sample AUC from 0.695 to 0.710 (+0.014 [0.003, 0.027]) across three learners, but the gain stems from crosswalk coverage, through which enforcement labels are assigned; relational content alone adds +0.005 (p = 0.37). A plasmode design that keeps real firms, financials, labels and enforcement timing while simulating subsidiary, ownership, common-control and related-party channels yields no gain under a placebo and +0.052 AUC with 3.6× NDCG@1% when relational mechanisms exist, carried mainly by relationship structure. Look-ahead use of enforcement outcomes inflates the AUC gain by 46% and NDCG@1% by 2.4×, static snapshots and matching errors erode it, and systematic concealment leaves a detectable footprint. Across 409 enforcement episodes, enforcement follows the last misstated year by a median of 3.5 years, so relational monitoring cannot wait for neighbours' enforcement outcomes.

13 November 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
WPWenjie Ping
Wenjie Ping
YLYiwen Liang
Yiwen Liang
Material misstatements are typically revealed years after the fact, so practical screening must rank firm-years from current filings under a strict alert budget. Existing detectors treat financial statements and MD&A text as separate feature blocks and ignore whether the narrative is consistent with the numbers. We propose narrative-numeric inconsistency (NNI) features that extract quantitative claims from MD&A text, anchor them to statement figures through accounting identities and measure three gaps: revenue growth versus narrated activity, reported margins versus narrated unit costs, and tone versus fundamentals. We evaluate NNI on 119,155 U.S. firm-years (1991-2010) with Compustat-based statements and SEC AAER labels, paired with a mechanism-calibrated MD&A layer in which misstatements affect text only through economic channels documented in prior research. Under a rolling out-of-time protocol (training on years up to t-2, testing on 2003-2010) with serial-fraud recoding, we compare numbers-only, text-only and numbers+text detectors. NNI fusion attains an AUC of 0.709, on par with the best numbersonly model (0.711), while delivering the best early ranking: NDCG@1% of 0.031 versus at most 0.028 for all baselines, and recall of 15.3% at a 5% alert budget versus 11.7% for a boosted numbers-only model, with the lowest false-positive rate. Naive early fusion of text representations lowers AUC to 0.688, unanchored claims add nothing, and without retraining NNI loses only 0.012 AUC versus 0.036 for numbers-only detection. A nullsignal control confirms that the gains arise from inconsistency signals rather than leakage.

27 August 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
WPWenjie Ping
Wenjie Ping
This paper develops and tests a new disclosure-based measure of valuation uncertainty for Level 3 fair values. The accounting hierarchy identifies whether a valuation relies on unobservable inputs, but it does not distinguish a tightly parameterized Level 3 estimate from one supported by a wide range of plausible assumptions. I exploit the quantitative ASC 820 disclosures of significant unobservable inputs made by U.S. business development companies (BDCs) and construct the Level-3 Valuation Uncertainty Index (L3-VUI). For each valuation row, L3-VUI transforms the disclosed high-low range of each input into a scale-invariant interval dispersion, aggregates multiple inputs by a root-mean-square operator, weights the resulting uncertainty once by the associated fair value, and scales by total investments. Algebraically, the index decomposes into Level 3 intensity, disclosure coverage, and within-Level-3 input dispersion, which permits direct ablation tests. The empirical sample contains 13 publicly traded BDCs, 26 firm-years in 2023-2024, 234 disclosed input rows, and 170 measurable valuation rows. The pooled year-controlled association between L3-VUI and log price-to-NAV is negative (coefficient-0.504, HC3 p = 0.016). The result remains negative and significant when conventional Level 3 intensity is included (coefficient-0.383, p = 0.021), while Level 3 intensity itself is positively associated with price-to-NAV. Within-Level-3 dispersion, rather than disclosure coverage alone, accounts for the negative relation. The direction is invariant to leave-one-firm-out estimation, alternative aggregation, a non-log valuation outcome, robust regression, and exclusion of a merger-affected observation. A 50,000-draw within-year permutation test yields p = 0.0109. Small-cluster procedures are directionally consistent but produce reference-distribution-dependent significance. Firm fixed effects and first differences do not show a short-horizon within-firm effect, locating the contribution in cross-sectional valuation stratification rather than a causal repricing claim. The evidence shows that the marketrelevant content of Level 3 reporting is not exhausted by the amount classified in Level 3; the disclosed geometry of the underlying valuation inputs contains distinct information.