A deterministic, auditable downside-risk screen over SEC EDGAR filings and GDELT financial news, with a reproducible crisis backtest, an honest timing analysis, and a pre-registered live forward test.
Aetherius watches a list of companies. It reads new headlines and filings, decides conservatively whether each one really mentions a watched company, scores it with fixed, explainable weights, and puts only evidence-backed, adverse items in front of a human reviewer.
Headline result. Replayed on 161 real archived headlines from three crises (SVB 2023, Wirecard 2020, FTX 2022), it flagged 9 of 9 affected names before their formal resolution. The backtest says nothing about false alarms: its control names never appear in the data (§4.6).
The honest qualifier. In 0 of 9 cases did the first flag come before the first headline reporting the stock already falling. The system is fast, auditable triage of a market reaction, not a predictor (working paper §4.5).
What's next. A pre-registered live forward test started on 2026-10-16. A daily GitHub Action scores live data with the frozen model and commits timestamped results.
| Area | Where to look |
|---|---|
| Data ingestion from two public APIs (SEC EDGAR, GDELT DOC 2.0), with retries, rate limits, dedupe keys and idempotent reruns | ingestion/ |
| Precision-first entity resolution: word-boundary regex, acronym stoplist, aliases, counterparty links | entity_mapping/ |
| Deterministic scoring: one shared adverse-language list and formula used by the API, backtest and forward test | scoring/ |
| Backend engineering: FastAPI, PostgreSQL + Alembic, JWT auth with per-client operator scoping, Celery workers and schedules | aetherius/ |
| Quality gates before anything reaches a reader: banned hype language, required disclaimer, evidence for every high-severity item, retry with backoff, dead-lettering | delivery/ |
| Reproducible research: frozen fixtures, bit-identical metrics, tests that pin the published numbers | simulations/backtest/, tests/ |
| Research honesty: confidence intervals, a market-timing check that weakens the headline claim, documented false-positive classes | working paper |
| Engineering hygiene: 116 tests, ruff lint and CI on every PR | .github/workflows/ |
| Event | Mechanism | Affected names flagged before resolution | Median lead to resolution | First flag before market reaction |
|---|---|---|---|---|
| SVB-2023 | Regional-bank contagion | 5 / 5 | 2.34 days | 0 / 5 |
| Wirecard-2020 | Accounting fraud | 1 / 1 | 6.60 days | 0 / 1 |
| FTX-2022 | Crypto counterparty contagion | 3 / 3 | 7.12 days | 0 / 3 |
| All | 9 / 9 (95% CI 66–100%) | 2.48 days | 0 / 9 |
"Resolution" means the formal event: receivership, insolvency filing or Chapter 11. The "market reaction" is the first headline in the corpus reporting the stock already falling (market_reaction_check.py).
- Small sample, chosen with hindsight. There are three famous failures and nine affected names, and both the events and the adverse-term list were chosen knowing the outcomes. The confidence intervals above are wide.
- Triage, not anticipation. Flags arrive together with the press reporting a sell-off.
- False alarms are untested. Each event has a control company, but none appears in its headlines, so "no control was flagged" proves nothing. The live forward test measures false alarms instead.
- Known false positives. Tickers that are English words match ordinary text ("coin prices fall" matches
COIN). Adverse terms match as substrings ("fall" fires inside "shortfall"). - EDGAR adds provenance, not signal. Filings are scored on metadata (form and date), never on the filing text.
- Whitelisted sources only. Coverage is a fixed list of financial-news domains, which is thin for non-U.S. events.
All of these are written up in §5 of the working paper. The forward test is the start of answering them.
EDGAR filings ─┐
├─▶ ingest + dedupe ─▶ entity mapping ─▶ deterministic score ─▶ adverse gate ─▶ signal
GDELT news ────┘ (idempotent) (word-boundary) (fixed weights) (≥1 downside │
term required) ▼
operator review ◀─ daily brief / urgent alert drafts ◀────┘
│
▼
quality gates ─▶ email + PDF delivery (retry, dead-letter)
git clone https://github.com/zariffromlatif/Aetherius-Risk-Intelligence.git
cd Aetherius-Risk-Intelligence
pip install -r requirements.txt
pytest aetherius/tests -q # 116 tests
python simulations/backtest/run_backtest.py --event svb-2023 # or wirecard-2020, ftx-2022
PYTHONPATH=aetherius python simulations/backtest/market_reaction_check.py --markdownGenerate a per-company stress-test deck (HTML, print to PDF) from the frozen SVB data:
python scripts/build_deck.py \
--ticker SIVB --name "SVB Financial Group" --sector "Regional Banks" \
--aliases "Silicon Valley Bank,SVB" \
--thesis "Concentrated regional-bank position with rate-sensitive HTM book." \
--counterparty "FRC:First Republic Bank:peer:0.6:First Republic" \
--window 2023-03-06:2023-03-12 \
--fixture-jsonl simulations/backtest/events/svb-2023/observations.jsonl \
--out sivb-deck.htmlTo run the full backend (API, worker, scheduler, Postgres, Redis, Mailpit), see aetherius/README.md.
aetherius/
app/
api/ # FastAPI routers, JWT auth, per-client scoping
services/
ingestion/ # SEC EDGAR and GDELT adapters
entity_mapping/ # Word-boundary matching, acronym stoplist
scoring/ # Shared adverse-language list and deterministic formulas
signals/ # Evidence -> scored, typed risk signal
pipeline/ # Observation -> evidence -> signal flow (API and workers)
drafting/ review/ # Brief drafting and operator review workflow
delivery/ # Quality gates, email/PDF, retries, dead-letter
reporting/ # Weekly summaries, stress-test deck assembly
workers/ # Celery beat: hourly ingestion, briefs, retries, retention
alembic/ migrations/ # Schema
tests/ # 116 tests
simulations/
backtest/ # Frozen crisis fixtures, replay harness, market-timing check
forward_test/ # Pre-registered live forward test + daily summarizer
docs/
working_paper/ # Methodology, results and limitations (Revision 2)
site/ # Static project page (Cloudflare Pages build output directory)
See CONTRIBUTING.md. PRs must keep scoring deterministic, include tests, and keep the frozen fixtures bit-identical.
Apache License 2.0. See LICENSE.
Research software. It does not provide investment advice, fiduciary services or return forecasts.