Open-source research software

Aetherius Risk Intelligence

A deterministic, auditable downside-risk screen over SEC EDGAR filings and GDELT financial news, with a reproducible crisis backtest, an honest market-timing analysis, and a pre-registered live forward test.

Code on GitHub Read the working paper Live forward-test log

The result, honestly

Replayed on 161 real archived headlines from three crises, the screen flagged every affected company before its formal failure date. But it never flagged a company before the press reported its stock already falling, and the backtest cannot say how often it raises false alarms.

9 / 9affected names flagged before resolution (95% CI 66–100%)

Untestedfalse alarms: the control companies never appear in the data

0 / 9first flags that came before the market reaction

So the system is fast, auditable triage of a market reaction already underway, not a predictor. The working paper (Section 4.5) shows the check and the tests that pin it.

Event	Mechanism	Flagged before resolution	Median lead to resolution	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

How it works

IngestSEC filings and whitelisted news, deduplicated so reruns never double-count.

MatchWord-boundary entity matching, so the ticker AI never matches "SAID".

ScoreFixed, published weights. The same input always gives the same score.

GateNo adverse language, no flag, however direct the name match.

ReviewA human approves every brief; delivery checks evidence and disclaimers.

Limitations

Small sample chosen with hindsight. Three famous failures, nine names. The events and the adverse-word list were chosen knowing the outcomes.
False alarms 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 this.
Known false positives. Tickers that are English words match ordinary text ("coin prices fall" matches Coinbase), and adverse terms match inside longer words.
Filings add provenance, not signal. SEC filings are scored on form type and date, not their text.

The live forward test

On 2026-10-16, before collecting any data, I committed a fixed list of 20 large US companies. Every day a GitHub Action scores live news and filings with the frozen model and commits the results. Because the record is append-only and timestamped by git, it can't be quietly edited afterwards. It will show how often the screen raises false alarms when nothing is happening, which the backtest cannot.

Run it yourself

git clone https://github.com/zariffromlatif/Aetherius-Risk-Intelligence.git
cd Aetherius-Risk-Intelligence
pip install -r requirements.txt
pytest aetherius/tests -q
python simulations/backtest/run_backtest.py --event svb-2023

About

Built by Zarif Latif as a study in building auditable data systems and evaluating them honestly. Python, FastAPI, PostgreSQL, Celery and GitHub Actions. Feedback, especially critical feedback, is welcome: zarif.latif.biz@gmail.com.