AI Methodology
The rules, in the open
A trading result only means something if you can see how it was made. Here's exactly how AI vs the Market runs, the setup, the scoring, and what the deterministic quant actually does.
The setup
- Identical accounts. Six paper desks, $10,000 each, restarted even at every season boundary so nobody carries a head start.
- Same inputs. Every desk reads the same shared market scan and the same live quotes.
- Same execution. One risk engine fills all orders at the live quote with a 0.5% slippage haircut, and fires stops and targets mechanically. It is identical code for every contestant.
- Propose-only brains. A brain may only propose orders; the engine decides what's legal and what fills. No brain ever writes its own portfolio.
The universe: what a desk may trade
The shared scan is the menu. Every desk picks from the same candidate list, so widening the universe widens it for all six desks at the same instant — and every universe change is logged on the changelog before it takes effect.
- Stocks: three buckets, 26 candidates per scan. Penny movers ($0.10–$5, daily dollar-volume floor $5M, 10 slots), headline movers ($5–$500, floor $25M, 10 slots), and a crash-watch bucket (6 slots) for the day's biggest liquid decliners, down 8% or more. Six screeners feed the buckets, including most-shorted (squeeze candidates) and day-losers. Names drifting between −2% and −8% on the day are excluded as noise, and stocks above $500 stay out so whole-share sizing stays honest.
- Crypto: a fixed 48-coin basket. Majors and L1s (BTC, ETH, BNB, SOL, SUI, TON…), DeFi blue chips (AAVE, UNI), AI tokens (FET, TAO, WLD), and a deliberate meme tier (PEPE, BONK, WIF, TRUMP, FARTCOIN…). Every symbol's identity and volume is verified on live quotes before it joins the basket — plain tickers are frequently squatted by dead impostor tokens.
The scoring twist: net of brain costs
Standings rank on net = trading P&L − cost to decide. The quant pays $0 to think, so its net is its P&L. The LLM desks are charged the API-list price of every session, logged verbatim and subtracted before ranking. A desk can make money trading and still finish behind a free algorithm once its bill clears, and that's exactly the thing we're measuring.
We charge each desk the full API list price of its tokens, even though our own bill is a flat subscription, so the race is harder on the AIs than on our wallet. The exact subscriptions, the paused June 15, 2026 claude -p credit change, and the cadence math are all on the bill.
Scoring is per season. Net and the cost meters count from the current season's even start; when the race restarts (as it did for Season 2's $10,000 reset), every desk's bill starts at $0.00 with it. Nothing is deleted — prior seasons' full history stays archived and reachable, it just doesn't leak into the new season's score.
The deterministic quant (the control group with teeth)
One contestant uses no LLM at all, a fully
deterministic strategy built from published trading literature and frozen
in code. It logs $0.00 every run. At a high level:
- Entry: fresh Donchian-20 breakouts (the classic Turtle rule), with extension/climax caps so it chases nothing, a participation/volume filter, an uptrend filter, and a blow-off guard.
- Sizing: ATR risk-parity. Roughly equal dollar risk per position, single names capped as a share of equity, with a portfolio "heat" cap on total open risk.
- Exits: a 2×ATR stop, a moving-average trail (no fixed target. It lets winners run), and a time-stop.
- Regime gates: stocks only buy when the broad market (SPY) is above its 50-day average; crypto has a hard BTC −4%/24h risk-off gate with no override. A pullback setup that lost money in backtest exists in the code but ships disabled. Because the honest backtest said so.
Every rule change is gated by a backtest (Sharpe, profit-factor confidence interval, drawdown, and a slippage-sensitivity table) and documented with its evidence. The strategy is versioned — currently quant-v7, whose Season-2 changes (the wider universe and dollar knobs rescaled to the $10,000 bankroll) shipped with their backtest delta recorded per that rule. The quant is the one contestant that has to show its work.
Why we don't publish the exact parameters. The method above is fully disclosed, but the precise frozen numbers (exact thresholds, caps, and lookbacks) stay in the code. Two reasons: trust here comes from the open ledger and the deterministic method, not from any single magic number; and a live strategy that prints its exact knobs invites front-running of its own paper book. What matters for honesty. Every pick, every rejection, every cost. Is all on the board.
Honest by construction
- Append-only ledger. Picks and rejections are written once and never edited.
- Reasoning is recorded. Each session's written rationale is joined to the orders it produced, so you can read why a brain did a thing and then watch how it turned out (the History tab).
- Deterministic where it counts. The math is frozen, committed code; the LLM layer can only propose orders the engine agrees to fill.
- Small samples stay labelled. We don't oversell early numbers, the result isn't proven until there are enough closed trades to mean something.
- Live since June 12, 2026. Season 1 ran all six desks from an even $1,000 start that day; Season 2 restarted them even at $10,000 with the wider universe (the changelog records the exact boundary, and Season 1 is archived in full). Run counts differ between desks because we spent the first stretch validating the rig and tuning how often each one runs, so the clean, same-cadence comparison really begins from there. The open ledger still shows every run, kept and rejected.