Independent quantitative research · since March 2026
The model, the size of the edge, the cost of trading it — written down first, then graded by the tape a day later. The rules that decide what may touch capital were frozen in writing before any result existed, and they do not move afterwards.
A strategy passes every one or it never trades. Most do not — and the ones that fail are written down with the measurement that killed them.
every candidate entersfew reach the aperture
Per-asset gradient-boosted classifiers — eighty-three engineered features, trained across three hundred ninety thousand walk-forward samples — call the next twenty-four hours, every hour, priced at real venue costs. A day later the market settles each call against a fixed barrier. Nothing is backfilled; nothing is quietly withdrawn.
A direction-neutral engine watches what the derivatives market pays to hold a hedged position — hourly, across two venues. It deploys only above a rate fixed in advance, and holds cash, deliberately, until the market pays it.
Two hundred numbered entries since March, every verdict with the measurement behind it, hourly calls with their settlements, and the paper documenting the stack — federated learning with differential privacy and ZK-verified inference.