An anonymised look at a quantitative system LambdaOrbit designed, built and operates: a genetic search engine that discovers candidate trading strategies, an adversarial statistical pipeline that filters them, and a control app over a cloud server running 24/7.
Volatility and options markets are noisy enough that a careless search will always “find” profitable-looking strategies that are pure overfitting. The brief was to build a system that could explore a vast space of strategy ideas automatically, yet hold each one to a standard of evidence high enough that survivors are worth real capital.
Four moving parts, engineered to work together unattended: discovery, validation, operation and control.
Each strategy is encoded as a compact specification — an entry condition on a market indicator, a target instrument, a holding horizon and a direction. A genetic algorithm treats these specs as a population: it mutates thresholds and horizons, crosses promising parents, and occasionally injects entirely new ideas.
To stop the search from collapsing onto one crowded idea, candidates are organised across a grid of more than 1,700 behaviour cells. The engine deliberately keeps the best performer in each cell — a quality-diversity approach that explores broadly instead of optimising narrowly. A language model also proposes fresh, mechanism-grounded hypotheses to seed regions the search has not yet reached.
A strategy is treated as a hypothesis that is guilty until proven innocent. It only advances by surviving a sequence of independent tests, each designed to expose a different way a result can be a fluke.
History is partitioned into train, validation, a sign-only soft holdout, and a single-shot lockbox that can be consumed exactly once — so no decision is ever contaminated by data it has already seen.
Expanding-window folds confirm the edge persists through changing market conditions, not just one lucky stretch of history.
Because thousands of strategies are tried, statistical thresholds are raised per behaviour cell and a false-discovery-rate sweep demotes winners that do not hold up against the crowd of attempts.
Survivors must show enough independent trades, an acceptable drawdown profile, performance across volatility regimes, and a positive result after realistic option pricing and transaction costs — not just a favourable raw t-statistic.
Promising strategies are paper-traded automatically and must accumulate a real track record of closed trades before they can ever be considered for live capital — a final, unfakeable gate.
The operator manages the system from a phone-installable app — a progressive web app backed by a small, deliberately minimal API. It shows system health, per-strategy scoreboards and live status, and lets the operator restart the data collector or confirm a strategy with a single tap.
Sign-in uses device biometrics and passkeys, so no shared password ever crosses the network. The action surface is intentionally tiny: even a full compromise of the control app cannot place a trade or read sensitive data.
Discovery is only half the work. The system runs itself on a hardened cloud server, with the operational hygiene of production software rather than a research script.
The platform turns strategy research from a manual, bias-prone activity into an automated pipeline that runs every hour and improves as data accumulates. Survival is intentionally rare: the value is precisely that almost nothing passes, so what does pass is worth attention.
More importantly, it demonstrates the approach LambdaOrbit brings to any applied-ML problem — automate the search, but be ruthless about evidence, and engineer the whole thing to operate reliably without a human babysitting it.
Discuss a system like this