People ask what "an AI team researches strategies" means concretely, so here's one hypothesis walked end to end. The important thing to hold onto: this is a funnel that exists to reject. Most ideas die. That's the product working.
Which questions people actually ask about it
Who comes up with the ideas?
A research agent that reads what the data library can support — funding rates, open interest, liquidation tapes, options surfaces, news sentiment, Fear & Greed — and writes hypotheses with a thesis, a venue, a timeframe and a risk profile. It's explicitly fee-aware: an idea whose plausible edge is under the venue's round-trip cost is malformed on arrival. It also reads the graveyard. Every abandoned strategy's failure reason is embedded and searchable, so "funding momentum on hourly bars, third attempt" gets flagged before it wastes a cycle.
Who writes the code, and why should anyone trust it?
A developer agent codes the strategy against a fixed base template — capital-relative sizing, explicit stop logic, declared data needs. Trust comes from the gauntlet, not the author. Four layers: static analysis (AST checks for the anti-patterns we've catalogued); synthetic scenarios (six market shapes with fabricated supplementary data — a strategy that buys into a crash scenario it was designed to avoid dies here); a sandbox backtest on real data that must produce at least one trade; and a QA agent doing code review for signal polarity, edge cases and fee viability. Layer 2 catches the most embarrassing class: code that runs perfectly and does the opposite of the hypothesis.
What does the backtest actually charge?
Real maker/taker per venue, funding accrual on the mark, square-root market impact, margin and liquidation for leveraged books, session calendars for equities. Mark-to-market equity including unrealized PnL, so drawdowns aren't hidden between entry and exit. If that sounds like table stakes, it isn't; each item got added because its absence produced a specific wrong promotion.
Doesn't optimization just overfit?
It tries to. The pipeline assumes so: parameters get a sensitivity sweep first (a strategy that only works at exactly lookback=14 is a coincidence, not a strategy), then walk-forward optimization across three windows with in-sample/out-of-sample tracking, then a final holdout the strategy sees exactly once. Out-of-sample Sharpe below half the in-sample number is a rejection. The holdout has to retain at least 70% of the walk-forward out-of-sample Sharpe. These are code-enforced gates, not analyst discretion — the analyst can waive a soft floor with a logged reason, but can't waive the overfit checks.
What's the pass rate?
Low, and it should be. The modal death is at analyst review: the edge is real but under 0.15% per trade after costs. The second most common is the sensitivity cliff. Survivors go to a risk officer, then to paper trading, where a daily replay-parity check keeps them honest against live data indefinitely. Live capital sits behind a separate, human-controlled gate.
