MGR Trade — Quantitative research
Evidencebeforeexposure.
We wanted a platform that would protect us from ourselves. We built a lab where every idea must prove it holds before a single dollar is committed — and an interlock nobody can work around.
In-house platform — access reserved to our team
- Year
- 2026
- Type
- In-house lab · partners wanted
- Stack
- Python 3.12 · FastAPI · PostgreSQL
- Status
- Private access
On every screen
One codebase. Every screen.
These aren't screenshots. It's MGR Trade's homepage, rebuilt in code and rendered twice: at 1440 px in the laptop, at 390 px in the phone. The layout really reflows.
The lab
Drag the edge. Watch it reflow.
Drag the handle, or use the ruler with your keyboard.
Gate 01Why we built it
Not one more bot. A method that refuses to lie to itself.
We were not looking for a miracle strategy: we wanted to know, with evidence, which ones were worthless. Hence a system where the hypothesis is written before the test, out-of-sample data stays sealed, and the exchange's real costs count from the start.
Gate 02The method
Observe. Test. Reject. Keep only what survives.
A hypothesis passes through successive gates: pre-registration, training, validation, real costs, return concentration. One closed gate is enough to reject it, and the verdict is final for the version tested.
- 01Observe
- 02Hypothesise
- 03Test
- 04Reject
- 05Rehearse on paper
- 06Earn eligibility
- 07Shadow
A rejected idea goes back to the start, under a new name and a new registration — never quietly.
Gate 03The evidence
Most ideas end here.
Twelve hypotheses were pre-registered, tested and refuted before a single dollar was committed. The ledger below says which gate closed — the strategy behind each idea stays private.
| Hypothesis | Gate that closed | Verdict |
|---|---|---|
| H-01 | Returns concentrated in a few trades, sign flipped between training and validation | Rejected |
| H-02 | Failed the placebo: a fake lag predicted as well as the real one | Rejected |
| H-03 | Too little cross-market consistency, effect carried by two assets of nine | Rejected |
| H-04 | A real effect, but the opposite way to the hypothesis | Rejected |
| H-05 | Added value not shown in validation | Rejected |
| H-06 | Validation correlation under the set threshold (0.096 for 0.15) | Rejected |
| H-07 | Sign opposite to the hypothesis in validation | Rejected |
| H-08 | Failed in training | Rejected |
| H-09 | Diversification weakest exactly when it is needed | Rejected |
| H-10 | The exchange's real costs five times the expected premium | Rejected |
| H-11 | Cost reduction too small to make the idea viable | Rejected |
| H-12 | Returns too concentrated (78% where 40% is allowed) | Rejected |
Gate 04Architecture
From raw market to decision, one verifiable chain.
A recorder captures the microstructure of seven public markets and turns it into a research dataset. The lab tests, the risk centre bounds, paper and then shadow trading rehearse the strategy without real money, and every step is audited, replayable and reported.
- 01Market recorderThe microstructure of seven public markets, captured segment by segment, checksummed and archived.
- 02Research labBacktests, walk-forward validation, bootstrap and Monte Carlo, economic-viability analysis.
- 03Pre-registered hypothesesCriteria written before the test; out-of-sample data sealed until an idea has passed training and validation.
- 04Risk centreLimits, exposure and utilisation tracked live, with risk sessions dated in UTC.
- 05Paper and shadow tradingThe strategy rehearses on real data without real money: first on paper, then shadowing the market.
- 06AI analyst copilotA read-only research assistant: it analyses and explains, it never executes.
- 07Audit, replay and reportsEvery decision is logged, replayable and documented in reports.
- 08Two-factor sign-inPasswords hashed with Argon2id and a mandatory TOTP code: no password-only path exists.
- 09Real-trading interlockIf the environment tries to enable real trading without authorisation, the application refuses to start.
Gate 05Live trading
The last gate stays shut until something deserves it.
No strategy is approved, so no real order is possible. This is not a setting: it is written in the code, tested, and checked at start-up.
APPROVED_STRATEGY_VERSION = NONELIVE_TRADING_ENABLED = FALSELIVE_GATE_APPROVED = FALSELIVE_AUTOPILOT_STATUS = UNAVAILABLE
Partners wanted
A complete lab, looking for its partners.
Built in August 2026, MGR Trade brings together research, risk, simulation, observability and an AI analyst copilot — with a safety interlock that stops the application from starting if anyone tries to enable real trading without authorisation. Research itself is expensive: market data, compute and time. We are opening the lab to partners and sponsors.
- hypotheses refuted
- 12
- Rejected on criteria written in advance, before committing a single dollar.
- markets recorded
- 7
- Public microstructure captured continuously and turned into a dataset.
- automated tests
- 1,517
- Deterministic, no network, run before every delivery.
- real orders possible
- 0
- Real trading is locked in code until a strategy deserves it.
- Python 3.12
- FastAPI
- PostgreSQL
- SQLAlchemy 2
- React 18
- TypeScript
- Vite
Measured
Measured, not promised.
Numbers you can check: the site's quality today, the history of its code, and what was done for people who read differently.
AQuality, today
An in-house platform, for our team only: its scores are not published.
BThe build
- lines of code
- 0
- automated tests, no network
- 0
Aug 9, 2026 → Aug 19, 2026 · 11 days
Built in eleven days in August 2026: the real-trading interlock is tested right up to the application's start.
Read from git on Oct 21, 2026. Application code only: no tests, config or build output.
CAccessibility
- 01Reduced motion respectedThe landing site's animations hold still when the system asks for less motion.
- 02Charts that say what they areEvery illustrative visual is labelled "illustrative structure — not live prices".
- 03Named controlsThe market selector is a labelled tab list, usable by keyboard and screen readers.
DBuilt with our tools
This product uses the open-source tools we publish on npm under the MIT licence. The same code, open to anyone.
See our open-source tools- 01maisongr-skillsThe discipline of our AI coding agents: understand before building, change only what was asked, prove the result.
The budget
What the market charges for this scope.
$80,000 – $150,000
- Closest tier
- Scaling SaaS product
- Continuous market data, a research lab, risk, simulation, AI and hardened security: the market's “scaling SaaS” tier.
- Typical timeline
- 6 to 9 months
A range published by third parties, not a Maison GR quote.
Estimate my project→Maison GR
A research partner or sponsor?
Market data, compute, or research funding: we are looking for partners who share our standard of evidence. MGR Trade is not an investment product, and no returns are promised.