FinGuard

FinGuard | Audit Intelligence for Financial Misstatement Detection
Proof of Concept for audit-risk intelligence

Catch financial misstatements before they cost billions.

FinGuard converts scarce, delayed misstatement labels into an explainable risk-scoring engine for SEC filings — helping regulators, auditors, and financial partners find hidden risk earlier and allocate oversight resources with precision.

3.4x detection lift 3.6x ROI improvement 45K validated cases
Early-risk signal
Recall @ 10% coverage
34%
ROI multiplier
3.6x
Ranking efficiency
Baseline
10%
Sup. ML
22%
PU Pipeline
34%

Prioritize the highest-risk filings

White-box components provide per-instance explainability for targeted review and partner-ready trust.

The oversight gap

Financial fraud is a needle-in-a-haystack problem with expensive consequences.

Audit teams and regulators monitor a vast universe of companies with capped budgets, manual workflows, and labels that arrive years after the damage has already spread.

$2.3M

Average audit cost per firm

Manual review consumes high-value expert time, making full-market scrutiny economically unrealistic.

10K+

SEC-listed companies to monitor

A massive search space creates a structural mismatch between available personnel and the volume of risk.

3–5 yrs

Average detection lag

Misstatements often surface late, when reputational, regulatory, and investor losses have already compounded.

The scientific edge

Most fraud is hidden, not labeled.

Standard supervised machine learning assumes a clean split between “fraud” and “clean” firms. In financial misstatement detection, the so-called clean set is often contaminated by undetected cases. FinGuard’s Positive-Unlabeled approach treats unlabeled firms as unknowns — not negatives.

Positive-Unlabeled learning

Models confirmed positives while learning from an unlabeled universe that may contain hidden misstatements.

Built for realistic constraints

Robust to detection lag, label noise, and severe class imbalance — the actual data conditions of audit oversight.

Explainable by design

Per-instance interpretability helps reviewers understand why a company is prioritized for further inspection.

confirmed positives unlabeled unknowns likely clean
Validated performance

A step-change in detection efficiency and capital allocation.

The proposed PU pipeline was validated on a 12-year longitudinal study of SEC filings covering 45,000 cases and 7,000 companies from 2003 to 2014.

3.4x

More misstatements detected at fixed coverage

At just 10% market coverage, FinGuard’s PU pipeline can identify roughly 34% of misstatements — outperforming the default baseline and existing supervised ML approaches.

0.69 vs 0.50 AUC Recall improvement over sequential selection, improving the quality of the ranked review queue.
$26.1M Estimated cost per material misstatement detection, down from a ~$94.3M baseline.
Feature / Metric Default Baseline Existing Supervised ML FinGuard PU Pipeline
Handles unlabeled data
Recall @ 10% coverage ~10% ~22% ~34%
Cost per detection $94M+ $40M+ $26M+
ROI multiplier 1.0x 2.1x 3.6x
Stakeholder impact

One risk layer. Three high-value entry points.

FinGuard is designed as a force multiplier for institutions that need to prioritize oversight, diligence, and review under real resource constraints.

Regulators

Catch roughly one-third of misstatements at just 10% market coverage — turning capped budgets into a sharper, data-driven enforcement funnel.

Auditors

Direct expert review to high-risk targets, reducing low-value manual screening and converting sunk audit cost into recovered value.

Investors

Create an early disclosure-risk signal for portfolio screening, lending decisions, underwriting workflows, and governance diligence.

Market potential

From audit research to deployable financial-risk infrastructure.

The PoC focuses on proving a practical, explainable risk-scoring layer for the 10,000+ company oversight universe — with expansion routes across regulatory technology, audit planning, investor diligence, lending, and insurance underwriting.

RegTech Audit analytics Portfolio risk Credit underwriting Disclosure intelligence
10K+

SEC-listed companies create a large monitoring universe with constrained review capacity.

coverage
10%

Targeted review coverage can become a high-yield queue instead of a broad manual sweep.

focus
34%

Misstatement recall at fixed coverage creates a compelling wedge for pilot deployment.

signal
PoC roadmap

A focused path from research foundation to a TRL-6 prototype.

The PoC is structured to move from integrated financial data and model calibration to live scoring, backtesting, and pilot-ready deployment.

Data integration & feature engineering

Build the ingestion layer and feature foundation for SEC filing-based risk scoring.

PU-model calibration & backtesting

Refine the learning pipeline against historical misstatement outcomes and realistic label constraints.

Live risk scoring & pilot deployment

Deliver a review-ready scoring interface for partner evaluation and workflow testing.

TRL 4 → 6 market entry

Advance from validated lab prototype toward relevant-environment demonstration and commercialization readiness.

The team

DICE team of NCSRD

Ilias Zavitsanos Konstantinos Bougiatiotis Andreas Sideras Dimitris Kelesis Nikos Reskos
Track record

Deep ML research, built for applied financial oversight.

Published foundation Research track record across ICAIF ’21 & ’25, ACM Computing Surveys ’25, and Applied Intelligence ’25.
Multidisciplinary expertise NLP, graph ML, end-to-end ML pipelines, industrial experience, and Horizon Europe project delivery.
Realistic constraints Approach designed around scarce labels, delayed confirmation, imbalanced outcomes, and institutional review needs.
Partner-ready PoC Seeking scientific and investor partners to move FinGuard from research validation to pilot deployment.
Call to action

Next steps with FinGuard

We are seeking Proof-of-Concept funding and strategic partners to validate live risk scoring, pilot deployment, and the commercialization path for explainable financial misstatement intelligence.