Book a scoping call
Home / Work / AI Financial Trading Research Engine
Quantitative finance Colabs ventureIn production, research-only

AI Financial Trading Research Engine

A research terminal that tries to disprove its own findings before it believes them.

What we set out to build

A way to test trading ideas at scale without fooling yourself. The founder's problem was not a shortage of strategies, it was that every backtest looked good and almost none survived contact with reality.

What was built

  • 348 strategy implementations across grid, mean-reversion, momentum and market-making families
  • A regime engine written from scratch in pure numpy: Gaussian HMM and GMM, no ML dependencies
  • Multiple-testing correction across every study, so a strategy that looks significant among hundreds of candidates has to clear a much higher bar
  • A factor library where adding a factor automatically scores it, with the haircut applied
  • An edge scanner over 105 exchanges through a single interface, plus an opportunity ranker
  • A nightly self-refutation stage that re-tests confirmed edges and revives refuted ones

What it achieves

  • Seven measured edges survived the full multiple-testing haircut; six candidate edges were refuted with the measurement published rather than quietly dropped
  • The regime engine beats a naive volatility state on forward-volatility prediction, measured rather than asserted
  • Every claim on the terminal carries its sample size, its confidence interval and its limitations, so a weak result cannot be mistaken for a strong one
  • 857 tests covering the statistics, not just the plumbing
The hard part: The hard part was not finding edges, it was building something that would tell the founder an idea was worthless. Most research tools are optimism machines. This one is designed to argue with you, and it has refuted more of its own findings than it has confirmed.

The stack

Python, numpy, ccxt, React, FastAPI.

Common questions

How was this built?

This was built through Colabs, our sister company. A founder brought the idea and the domain knowledge, Colabs brought the team, the founder funded the build with a monthly subscription, and they hold equity in the company that resulted. We label every case study, because a studio that publishes its prices and argues the other side of its own comparisons has to be equally precise about its own portfolio.

What state is it in?

In production, research-only. We state that plainly rather than describing everything as production: several things here are deliberately gated, and the reason is usually evidence rather than a missing feature.

Can you build something like this for us?

That is the point of publishing it. The same patterns are in the catalogue at published prices, and you can scope your own version on the plain-English page without contacting us first.

Where to next

If something here is close to what you need, the same patterns are in the catalogue at published prices, or you can describe your version in plain English and get it scoped and priced without talking to anyone.

[ // / QUANTITATIVE FINANCE ]

Other quantitative finance work

Built with the same discipline, in the same domain.

Want something like this built?

The catalogue prices the patterns behind this work. Describe your version and the engine will scope it, including an allowance for anything nobody has costed.