Vol. I · Downside-Risk Intelligence · Est. 2026
A per-target Stress-Test Deck built from SEC filings, counterparty mapping, and a scoring pipeline whose detection track record is disclosed on the cover page — not asserted in marketing.
Every claim in every deck links to a primary source. The full pipeline is open source, Apache-2.0. Audit the repository.
SVB 2023 · Wirecard 2020 · FTX 2022
| Event | Recall | Median lead | FP |
|---|---|---|---|
| SVB 2023 | 5 / 5 | 2.34 d | 0 |
| Wirecard 2020 | 1 / 1 | 6.60 d | 0 |
| FTX 2022 | 3 / 3 | 7.12 d | 0 |
| Aggregate | 9 / 9 | 2.48 d | 0 / 3 |
n = 161 observations across three frozen historical windows (SVB 2023, Wirecard 2020, FTX 2022). Real primary-source news from the GDELT archive, measured on information available at the time only. Early-stage validation; coverage expanding.
§ I · Methodology
SEC EDGAR filings (8-K, 10-Q, 10-K) and primary-source news from a curated whitelist of financial wires are pulled for the target and its declared counterparties — suppliers, customers, funding sources, regulators.
edgar_adapter.py · gdelt_adapter.py
A conservative entity-mapping pipeline links each observation to the right name with word-boundary matching and an acronym stoplist. A deterministic 7-factor risk-scoring formula grades every signal. No LLM decides severity.
entity_mapping/service.py · scoring/service.py
You receive a structured 8–12 page PDF: target & counterparty dependency map, 90-day evidence timeline with primary-source URLs, invalidation markers, and the historical detection track record on structurally comparable names.
target_stress_test.py · templates/pdf/
§ II · The working paper
We replay a systematic entity-mapping and severity-scoring pipeline against real primary-source news headlines from three widely-studied financial crises and report per-name detection timing, mapping precision, and false-positive control behavior.
“Across the three frozen historical windows (161 real news observations sourced from GDELT DOC 2.0), the pipeline flagged 9 of 9 affected names at ‘elevated’ or ‘high’ severity before the realized event date, with a median lead time of 2.5 days and zero false positives on the three unrelated control names.”
§ III · Track record
Detection is validated on frozen historical windows — the pipeline sees only the information that existed at the time. Three crises, each replayed against real primary-source news from the GDELT archive:
| Event | Failure mode | Obs. | Recall | Median lead | FP on control |
|---|---|---|---|---|---|
| SVB 2023 Regional-bank contagion, Mar 6–12 |
Deposit run | 98 | 5 / 5 | 2.34 days | 0 |
| Wirecard 2020 Accounting fraud + insolvency, Jun 15–30 |
Idiosyncratic fraud | 14 | 1 / 1 | 6.60 days | 0 |
| FTX 2022 Crypto counterparty contagion, Nov 2–14 |
Counterparty risk | 49 | 3 / 3 | 7.12 days | 0 |
| Aggregate | — | 161 | 9 / 9 | 2.48 days | 0 / 3 |
Early-detection and prioritization evidence on a historical dataset. Not a return forecast. Not a guarantee of future results. Full methodology, watchlists, ground-truth files, and pinned tests are published in the open repository.
§ IV · The deliverable
Each Stress-Test Deck is a self-contained PDF built by the same pipeline the working paper documents. Human-reviewed before delivery.
§ V · Engagement structure
Tier 01
$500–$1,500
Tier 02 · Most requested
$2,000–$3,000
Tier 03 · Ongoing coverage
$2,500/mo
All engagements: pay first, deliver on the committed timeline. Stripe Invoicing or bank transfer.
§ VI · About
Aetherius is built and run by Zarif Latif, a computer-science undergraduate at BRAC University and the engineer behind every component this site documents. The SEC EDGAR ingestion, the conservative entity-mapping layer, the deterministic scoring formula, and the backtest harness — 75 pinned tests running against frozen GDELT corpora — are all his work, and all of it is open for audit under Apache-2.0. His focus is systematic, evidence-first risk research: shipping the part of the problem that can be measured and reproduced today, rather than the part that can only be asserted. He also builds Railo, an automated platform that ships deterministic security fixes in every pull request. Based in Dhaka; engagements conducted remotely.