Methodology
How Cascade maps a global event to public-company exposure, what data feeds the answer, and what the POC deliberately doesn't try to do.
§ 01Pipeline
v1 is hand-curated. v2 replaces the hand-curation with Claude-driven NLP extraction over the full S&P 500 corpus.
Scenario definition → 3 maritime chokepoints
↓
Corpus selection → EDGAR full-text search + curated S&P-100 names
↓
Filing retrieval → SEC EDGAR REST API (most recent 10-K per company)
↓
Verbatim extraction → Sentence-window extraction around the keyword
↓
Tri-axis scoring → Geographic / commodity / concentration (0–3 each)
↓
Composite + confidence → 0–9 sum; explicit / implicit / inferred tier
↓
Static render → cascade.georg.miami
§ 02Scoring rubric
Geographic exposure
Does the 10-K name the affected region?
0
No regional mention
1
Region implied via broader geography
2
Region named in segment data or general risk language
3
Region named explicitly in Item 1A or Item 7
Commodity exposure
Does the company depend on a commodity that transits the chokepoint?
0
No commodity dependency
1
Generic commodity exposure
2
Specific commodity transits the chokepoint
3
Critical / single-commodity dependency on chokepoint flow
Concentration / single-source
Single-source supplier or supplier-region risk disclosed?
0
No concentration disclosed
1
General supplier-concentration language
2
Specific supplier or supplier region named
3
Disclosed single-source dependency or material concentration risk
Confidence tiers
- Explicit: the 10-K names the specific region, chokepoint, or supplier by name.
- Implicit: the filing describes a dependency that includes the chokepoint but does not call it out by name.
- Inferred: score relies on industry context plus a thin disclosure. Flagged in the UI.
A worked example
Par Pacific Holdings (PARR), 2025 10-K:
The Russia-Ukraine war, the Israel-Palestine conflict, the political activity in Venezuela, Houthi-related disruptions in the Red Sea, and tensions involving Iran and the Strait of Hormuz have all continued to disrupt global trade patterns, increase crude oil price volatility, and, at times, increase freight costs and delivery times.
- Geographic = 3 (names Iran + Strait of Hormuz explicitly)
- Commodity = 3 (crude oil, exactly what transits the chokepoint)
- Concentration = 2 (refining is single-source dependent on crude feedstock supply)
- Composite = 8, Confidence = Explicit
By contrast, ExxonMobil's 2025 10-K does not use the phrase "Strait of Hormuz" anywhere. Its Middle East presence is described through Saudi-joint-venture segment data plus general "geopolitical" risk-factor language. That earns implicit confidence at the same geographic level: a different evidence quality despite higher absolute exposure. This contrast is the most important insight from the v1 corpus.
§ 03Data sources used in v1
| Source | Access | Used For |
|---|---|---|
| SEC EDGAR REST API | Free · User-Agent required · 10 req/sec | Filing index + 10-K retrieval |
| EDGAR Full-Text Search | Free | Identify candidate corpus per scenario |
| UN Comtrade API | Free with registration | Commodity trade-flow context in scenarios |
| US Census USA Trade Online | Free with key | HS-code import flow references |
| NASA Worldview Snapshots | Free · no key | Satellite imagery for chokepoint hero panels (MODIS, Blue Marble) |
| Copernicus Sentinel-2 / Sentinel-1 | Free with registration | Cited for illustrative imagery; full tiles deferred to v2 |
§ 04Data sources NOT used in v1
| Source | Why deferred | Reserved for |
|---|---|---|
| GDELT 2.0 | Real-time triggering not needed for static demo | v2: live event auto-fire |
| ImportYeti | Scraping fragility; rate-limited free tier | v2/v3: tier-1 supplier validation |
| AISStream.io | Real-time AIS unnecessary for cached scenarios | v3: chokepoint vessel-density overlay |
| ACLED | Registration barrier; weekly cadence | v2: geopolitical violence overlay |
| GLEIF LEI | Entity resolution unnecessary at v1 scale | v2: disambiguate 500+ filers |
| Wikidata SPARQL | Sparse for non-US subsidiaries | v2: foreign-supplier graph |
| Bloomberg SPLC · FactSet SCR · Sayari · Panjiva | Paid · the moat for commercial tools | Deliberately not matched. Cascade stops at the free-data ceiling |
§ 05The Reg S-K caveat
In 2020, SEC Regulation S-K Item 101(c) was modernized away from a prescriptive "name your principal suppliers" rule toward a principles-based materiality standard.
The practical effect for Cascade:
- Some public companies that would have disclosed supplier identity under the old rule are now legally permitted to stay silent.
- Risk-factor disclosures (Item 1A) and MD&A (Item 7) remain the practical homes for chokepoint language, but it is unstructured prose, not tagged data.
- No XBRL concept tags supplier identity. Supplier extraction is fundamentally an NLP problem.
This shift is exactly why Cascade names the dimensions of what's disclosed vs. what's hidden as a feature rather than a footnote.
§ 06Known limitations
- Coverage is shallow. v1 has 30 companies. Full S&P 500 coverage requires automated extraction.
- No tier-2 or tier-3 visibility. If a public company's tier-1 supplier is a private foreign manufacturer dependent on a chokepoint, that chain is invisible in EDGAR.
- Historical bias. Filings reflect what management chose to disclose at the time of filing.
- EDGAR full-text search has recall gaps for image-based PDFs and older filings.
- Maritime chokepoints are easier to model than non-maritime risks. An undersea-cable cut affecting fintech infrastructure is harder to capture.
- Survivorship bias. Companies that failed under prior disruption stop filing.
§ 07What changes in v2
| Aspect | v1 (now) | v2 (proposed) |
|---|---|---|
| Corpus | 30 hand-picked | Full S&P 500 (~500 filings) |
| Extraction | Hand-curated quotes | Claude NLP over EDGAR full-text |
| Scoring | Hand-assigned 0–3 | Claude scoring + human review checkpoints |
| Trigger | Pre-built scenarios | Live GDELT event detection → auto-scenario |
| Visualization | NASA Worldview stills + drawn SVG vessel routes | AISStream.io vessel-density overlay + Sentinel-1 SAR |
| Output | Static HTML | Static HTML + email/Slack alerts |
§ 08Reproducibility
Every excerpt in data/companies.json includes cik, filing_url, filing_date, the keyword that triggered extraction, and the verbatim excerpt. Every claim in Cascade is traceable back to a public SEC filing.
§ 09Disclaimers
- Not investment advice. Cascade is a research/journalism tool, not a portfolio-construction signal.
- Data current as of 2026-05-13. Filings will update; this static snapshot will not.
- No warranty. Best-effort accuracy; users should always verify against the linked source.