How to predict corporate bankruptcy from open-source signals: A 7-category methodology

Corporate bankruptcy is rarely a sudden event. In analysis of 40+ post-2020 Chapter 11 filings, Semantic Visions identified clusters of decision-grade public signals appearing 6 to 34 months before formal filing, organized into seven categories: financial, governance, operational, regulatory, legal, supply-chain, and sentiment. Below is the framework, validated against named cases — Fisker (13 months of public distress before filing), Ebix (34 months), Energetický Holding Malina (14 weeks of compressed signals), Spirit Airlines.

Why financial filings are too late

Traditional credit risk models — Altman Z-score, Merton, KMV — operate on financial-statement data. These models are well-validated, widely understood, and structurally lagging.

The structural lag has three sources. Reporting cycles. Public companies file quarterly, often with 30–45 day lag. Private companies file annually with longer lag. Audit certification gaps. Going-concern warnings typically appear in 10-Q footnotes only after the auditor has formally questioned viability — which itself follows months of internal back-and-forth. Management discretion. Earnings restatements, covenant breach disclosures, and material adverse change announcements are subject to legal-counsel review and timing decisions.

By the time a going-concern warning hits a 10-Q footnote, average lead time before bankruptcy in our dataset is 90 days. By the time Bloomberg or Reuters covers the impending filing, average lead time is 18 days.

Open-source intelligence (OSINT) closes this gap by monitoring the same underlying business reality through different windows. Court filings, regulatory inquiries, supplier press, employee reviews, executive departures, and trade-press coverage all surface weeks or months before they consolidate into a single financial disclosure event.

The Business Decline Curve and where OSINT bridges the gap

A widely adopted framing of corporate financial decline tracks four sequential stages: underperformance (declining margins, missed targets), balance-sheet distress (covenant pressure, leverage stress), cash crisis (liquidity squeeze, missed payments), and insolvency (formal filing).

Most credit risk models become reliable only at stage 3 or 4. By that point, intervention options are narrow: forbearance, distressed sale, or filing.

OSINT is uniquely positioned to detect signals at stages 1 and 2 — precisely when intervention options are still wide. The signals at these stages don't appear in financial statements. They appear in court records, regulatory filings, supplier press, executive bios, employee review platforms, and customer complaint forums — exactly the public surface that Semantic Visions' platform monitors continuously across 12 languages.

The 7 categories of pre-bankruptcy distress signals

1. Financial signals

Credit-rating downgrades, covenant breaches, ATM offerings, convertible-debt structuring, missed debt payments, going-concern footnote changes, auditor resignations, earnings restatements. Sources: rating agency releases, SEC filings, lender disclosures, financial press.

2. Governance signals

CFO, CAO, or auditor departures within compressed timeframes; board-member resignations; chair turnover; executive succession announcements with no successor named. Sources: 8-K filings, exchange announcements, executive-search press, LinkedIn movement.

3. Operational signals

Production halts, factory closures, indefinite layoffs above 5%, capacity reductions, fulfillment delays surfacing in customer or supplier press. Sources: local press at facility locations, supplier announcements, trade publications, employee review platforms.

4. Regulatory signals

SEC inquiries, exchange compliance notices, regulator investigations (DOT, FDA, EPA, NHTSA, equivalents in other jurisdictions), license suspensions, sanctions designations. Sources: regulatory press releases, exchange notice systems, government enforcement databases.

5. Legal signals

Securities-class actions, vendor liens, customer lawsuits, contract-termination filings, fraud investigations, criminal complaints. Sources: court records (PACER, equivalents), legal-databases, law-firm press releases announcing investigations.

6. Supply-chain signals

Supplier insolvencies, contract cancellations announced supplier-side, force-majeure declarations, payment delays surfacing in supplier press, key-customer departures. Sources: supplier press releases, trade publications, customs/trade data, force-majeure registries.

7. Sentiment signals

Surge in negative coverage in non-Tier-1 outlets, employee review-score collapse on Glassdoor or local equivalents, customer-complaint spikes on review platforms, online creditor complaints. Sources: news aggregation across 12 languages, employee review platforms, consumer complaint forums.

A single signal in any one category is rarely conclusive. A cluster across three or more categories — particularly Financial + Governance + one Operational/Legal/Regulatory — is highly predictive in our dataset.

Case study: Fisker — 13 months of public distress before filing

Fisker Inc. filed for Chapter 11 on June 17, 2024. The first decision-grade distress signal cluster appeared in our system in May 2023.

‍

Six of the seven signal categories triggered before filing. Each event was public at the time. None, alone, would have moved a credit rating. Together — entity-linked, weighted, surfaced continuously, they painted a picture rating agencies caught up to roughly 90 days before bankruptcy.

The Fisker case illustrates the core methodology argument: lead time, not real-time, is the metric that matters. A monitoring system optimized for "real-time alerts on the day of bankruptcy" is documenting an event, not preventing exposure to it.

Case study: Energetický Holding Malina — 14 weeks of compressed signals (CEE)

The Fisker case demonstrates the methodology against a US-listed company with full SEC disclosure infrastructure. The Malina case demonstrates that the same framework operates in a Central European, primarily-Czech-language registry environment without comparable formal disclosure.

Energetický Holding Malina, a Czech residential solar installer, filed for insolvency on May 6, 2023. The first decision-grade signal appeared in late January 2023 — 14 weeks earlier.

‍

The compressed timeline (14 weeks vs Fisker's 13 months) reflects two structural differences. First, smaller private companies tend to deteriorate faster than large public ones, there's less institutional buffer. Second, the visible distress phase is shorter when a company doesn't have public-market disclosure infrastructure to slow the surfacing of signals.

What this case demonstrates is the architectural requirement underneath: the methodology only works if the underlying platform reads the languages in which the signals appear. Every Malina signal was in Czech-language sources. A platform monitoring English-only press would have detected this case roughly two days before filing a lead time of zero.

Case study: Ebix — 34 months from first major signal

Ebix Inc., a software supplier across insurance, finance, and healthcare, filed for Chapter 11 in December 2023. The first decision-grade governance signal — the abrupt resignation of independent auditor RSM — appeared in February 2021. That's a lead time of 34 months.

The Ebix case illustrates the slow-burn variant of the framework. Distress did not progress linearly. The auditor resignation in Feb 2021 triggered a wave of investor lawsuits and stock decline. The company stabilized for ~18 months while pursuing growth (Educomp acquisition, EbixCash IPO filing). The Hindenburg-related investigation in June 2022 reactivated scrutiny. Asset sales in July 2023 confirmed cash-flow stress. By October 2023, a credit-facility restructuring made the Chapter 11 filing in December effectively foregone.

For risk monitoring, the Ebix case argues for multi-year signal persistence: a flagged company that appears to "recover" should remain on elevated monitoring for 24+ months because the underlying distress patterns often re-emerge.

How to operationalize the framework

A risk-monitoring system that operationalizes this framework needs four capabilities, each of which is an architectural choice rather than a feature.

1. Multi-source ingestion across 12+ languages. Every signal category surfaces in different source types (court records, regulatory press, supplier press, employee reviews, customer complaints), and the source language depends on company geography. A platform monitoring English-only press will systematically miss CEE, Latin American, East Asian, and Middle Eastern cases.

2. Entity resolution that survives variation. Distress signals about "Fisker," "Fisker Group Inc.," and "Henrik Fisker EV" must be linked to a single entity. Signals about "AMD," "Advanced Micro Devices, Inc.," and 阿里巴巴集团 must each resolve to canonical entities. Without robust entity resolution, signal clustering fails.

3. Event-level taxonomy with category mapping. Each surfaced event must be classified to one of 720+ event types and aggregated to one of the 7 distress categories. This allows the cross-category clustering that drives prediction confidence.

4. Time-bounded signal weighting. A single CFO departure is unremarkable. A CFO departure within 60 days of a going-concern footnote and an SEC inquiry is a high-confidence signal cluster. The system must compute weighted, time-bounded combinations rather than treating signals as independent events.

Semantic Visions implements all four through its proprietary processing pipeline: 1.9 million news articles per day across 12 languages, fine-tuned LLM extraction with NER post-filtering, classification against 720+ event types, and continuous graph maintenance for entity-relationship persistence.

OSINT-based vs traditional credit-risk methodology

The 7-category OSINT framework is not a replacement for traditional financial risk modeling. It is a complement that closes specific gaps.

‍

Used together, OSINT signals provide early warning that triggers deeper financial analysis. Traditional models confirm and quantify the financial trajectory once signals cluster.

Limitations and false-positive analysis

Honest framing of the methodology requires acknowledging where it produces false positives.

Recovery cases. Approximately 18% of companies that exhibit a 3+ category signal cluster in our dataset do not file for bankruptcy within 24 months — they recover, restructure outside bankruptcy court, or are acquired. Examples include several airlines that exhibited Fisker-pattern signals in 2020–2021 but stabilized post-COVID. The framework predicts severe distress, not bankruptcy specifically.

Industry-specific calibration. Signal categories carry different weights by industry. CFO departures matter more in financial services. NHTSA investigations matter more in automotive. Production halts matter more in manufacturing than in software. Threshold calibration by industry is required.

Sentiment noise. Sentiment signals are the noisiest of the seven categories. Used alone, sentiment is unreliable. Used as a confirmation layer when 2+ structural signals (financial, governance, operational, regulatory, legal) have already triggered, sentiment adds meaningful confidence.

Signal exhaustion. Companies under chronic distress can generate years of signals without filing. Constellation of signals, not signal count, drives prediction. The Ebix case is illustrative: distress signals appeared in 2021, persisted through 2022, and finally clustered into bankruptcy in late 2023.

Frequently asked questions

Can AI predict bankruptcy?
AI alone does not predict bankruptcy. AI applied to a structured signal taxonomy across multilingual sources can predict severe corporate distress with 6–34 month lead time, validated against named historical cases. The methodology requires ingestion architecture, not just a model.

How accurate are OSINT-based bankruptcy predictions?
In Semantic Visions' analysis of 40+ post-2020 cases, 3+ category signal clusters preceded eventual bankruptcy or insolvency in 82% of cases within a 24-month window. The remaining 18% recovered or were acquired before filing. The framework is most useful as severe-distress detection, not as a deterministic bankruptcy predictor.

What is a going-concern warning?
A going-concern warning is auditor language indicating substantial doubt about a company's ability to continue operating for the next 12 months. It typically appears as a footnote in audited financial statements. By the time it appears, average lead time before bankruptcy in our dataset is 90 days.

How does this differ from credit-rating monitoring?
Credit ratings synthesize lagging financial-statement data with limited qualitative input. The OSINT framework operates on real-time public signals across non-financial dimensions (governance, operations, legal, sentiment) that credit ratings only incorporate after they consolidate into financial impact.

Can this methodology work for private companies?
Yes, with adjustment. Private companies don't file SEC reports, but they do appear in court records, regulatory filings, supplier press, employee reviews, and customer complaints. Six of the seven signal categories are accessible without public-market disclosure. The Malina case is a private-company illustration.

What languages does the methodology require coverage of?
Coverage requirements depend on company geography. Minimum useful coverage for a global portfolio includes English, Mandarin, Spanish, Portuguese, German, French, Russian, Japanese, Korean, Arabic, Czech, and Polish. Semantic Visions covers all 12 natively.

How is this different from adverse-media screening?
Adverse-media screening monitors a single category (Sentiment) against a vendor list. The 7-category framework monitors all seven categories with cross-category clustering and operates against a continuously maintained relationship graph rather than a fixed list. Adverse media is one input; the framework is the analytical layer.

Methodology note

The methodology described in this article is operationalized through Semantic Visions' Early Warning Signals platform, drawing on a multi-source intake of 1.9 million articles per day across 12 languages, classified against a 720+ event type taxonomy, and aggregated to the 7-category distress framework described above. Case-study data is drawn from Semantic Visions' published whitepaper Early Warning Signals: Leveraging OSINT for Predicting Business Distress and Bankruptcies (2025) and the use-case studies on Fisker, Ebix, and Spirit Airlines. The Malina case is from Semantic Visions' CEE-language coverage research.

Full sources, taxonomy, and per-case signal trails are available to enterprise customers through the svEye platform.

Further reading

• White paper: Early Warning Signals — Leveraging OSINT for Predicting Business Distress and Bankruptcies

Talk to us

Want to know whether the 7-category framework would have flagged a counterparty in your portfolio? We can run a free, named-account historical replay using the same methodology applied above. Book a 30-minute review.