AI-Assisted Market Analysis: measure first, believe later
A research platform that collects official regulatory disclosures from listed companies, sorts and reads them with AI, and puts every idea for a market signal through statistical tests before any number is trusted. Built and operated by Agentas.
- Status
- Research phase complete
- Data
- Over 99,000 disclosures since 2023
- Stack
- Python, SQLite, FastAPI, Claude
- Source
- Private — walkthrough on request
What it does
Official disclosures, timed to the millisecond
Every regulated disclosure is stored with the exchange's own timestamp and category. The collector splits its queries so no message is silently dropped, and the exact timestamp lets every test use only what was public at that moment.
AI where reading is needed
The regulator's own category sorts each disclosure at no cost. Claude reads the text where judgement is needed, with a prompt that holds only facts known at that point in time.
The null test comes first
Random events run through the same harness before any signal is measured. If pure noise shows a gain, the run fails and no signal number from it is believed.
Permutation gating
Each signal is compared with the same events on shuffled dates. If the real dates do no better than the shuffled ones, the event date carries no information and the signal stops there.
Costs that never flatter
The cost model takes the larger of two spread estimates and puts stocks of unknown liquidity in a costly tier, because understating costs is the one error that manufactures a fake edge. Gross and net are always reported side by side.
AI tested blind
To test whether AI could read disclosures well enough, it got batches with no outcomes, a prompt fixed in advance, one run with no tuning, and only events after the model's knowledge cutoff, so it could not answer from memory.
Relevant if you're looking for…
This is a private research platform that gives no investment advice; it is relevant if you want to build or understand analysis of this kind.
- Analysing company announcements with AI: the platform has collected over 99,000 official disclosures since 2023, each stored with its official timestamp and category, and AI reads the text where judgement is needed.
- Sentiment analysis for the stock market: any signal read from text has to beat a null test with random events and a permutation test on shuffled dates before it is believed.
- Avoiding look-ahead bias: every test uses only what was public at that moment, with prices and news cut off before the decision.
- A custom analysis tool for your business: we bring the same discipline to any analysis where a convincing chart could still be wrong, and test the numbers against chance before anyone builds on them.
We build this for clients: Data and reporting platform · AI and LLM integration into your systems
How it works
- 01
Collect
Official disclosures and daily prices in one database.
- 02
Classify
The exchange's category first, AI reading where it adds something.
- 03
Test
Null test, permutation gate and cost model, in that order.
- 04
Verdict
Each hypothesis gets a written verdict in the research log.
How it's built
The platform is a Python service with one SQLite database and a single-file dashboard, deployed as a hardened container. The dashboard shows measured facts: its alert feed deliberately carries no predicted direction, and anything that spends AI tokens is marked in its own colour.
The working rule is measure before build. The platform has no order execution, and a signal would have to clear the gate before any were added. Today it runs as a disclosure reader with measured base rates per category.
Engineering highlights
- 60+ automated tests; the suite gates every deploy
- Look-ahead guards on every input: entry at the next tradable open, prices and news cut off before the decision
- A hard daily cap on AI spending, checked before every paid step; automation is off by default
- Claude runs headless in a locked-down sandbox with every tool denied
- Each deploy must prove the running container serves the new commit
- Read-only container, non-root, with all capabilities dropped
Questions
Can AI be used for stock analysis?
AI can read and classify company disclosures, but whether its reading adds anything has to be measured. We tested it blind: a prompt fixed in advance, one run with no tuning, and only events after the model's knowledge cutoff, so it could not answer from memory.
Does the platform give buy or sell recommendations?
No. It is a research platform with no order execution, and its alert feed deliberately shows no predicted direction. Nothing on this page is investment advice.
How do you keep AI costs under control in a system like this?
The regulator's own category sorts each disclosure at no cost, and AI only reads where judgement is needed. A hard daily cap on AI spending is checked before every paid step, and automation is off by default.
Want something like this?
Before you build on a number, test it against chance and against realistic costs. We bring the same discipline to any analysis where a convincing chart could still be wrong.
Get in touch