Your backtest worked. Your trading didn't.
Edge isn't designed.
It's what survives.
Candidates come from named mechanisms: who is forced to trade, and what it costs them. Then we try to kill every expression of the idea - under pass/fail criteria registered before the data is touched. The algorithm will always find structure. The question is whether it means anything.
Our process
Edge is selected,
not designed.
An idea is never one strategy. It's a family of expressions - transforms, lags, combinations, parameterizations - and we test the family, not the favorite.
The tooling: evolutionary search across candidate expressions, walk-forward validation with purge and embargo, pre-registered pass/fail criteria.
Surviving the search is necessary. It's never sufficient. Structure only counts when the mechanism behind it holds.
Your thesis. The right question.
You've built intuition from watching markets for years, and we translate it into something measurable. But before we optimize anything, we figure out what's worth optimizing. Most traders go straight to tuning the dials - sizing, entries, exits - without asking whether they're turning the right ones. The same data answers completely different questions depending on what you're trying to predict. We start there.
Production
Your edge survived. Now it needs to run.
Data pipelines, historical-to-live matching, real-time execution. From data ingestion to deployed system - not a PDF.
If this sounds familiar - here's what the engagement looks like: a research engagement that starts with your thesis and ends with a verdict - and if it survives, a production system that trades it. See services
Why the results hold
No future information
Signals are stamped to what you'd actually know at decision time. Fills are modeled at realistic execution windows - not the close that generated the signal.
Execution realism
When your signal is strongest, the market is usually thinnest. We model that. Costs scale with conviction, not just volume.
Holdout validation
The search never touches out-of-sample data - not for signal selection, not for parameter tuning, not for stopping rules. One pass. No reruns.
In the past year, four candidate strategies entered our validation protocol under pre-registered criteria. Most failed. Every verdict is provable.
Most strategies don't survive the process.
The ones that do are worth trading.
Tell us what you're working on. We'll let you know if there's something testable.
Start a conversation