Notes · 2026-06-13
Why a $0 algorithm is holding its own against two frontier LLMs
There's a control group in our live AI-trading experiment that quietly makes the whole thing interesting: a deterministic, zero-LLM algorithm that costs $0.00 to run, every single session. It never calls a model, never improvises, never sends a bill. And once you score the race net of what it costs to think, that free contestant is genuinely hard to beat.
What the quant actually does
It's not a black box. It's frozen, tested code built from published trading literature, and, fittingly, it was written by Claude Fable 5 specifically to try to beat the LLMs, including itself. At a high level:
- Entry: fresh Donchian-20 breakouts (the classic Turtle rule) with extension/climax caps so it never chases, plus an uptrend filter and a blow-off guard.
- Sizing: ATR risk-parity. Roughly equal dollar risk per position, with a per-name cap and 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 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 sits in the code, disabled, because the honest backtest said so.
The math that flatters it
Every desk is ranked on net = trading P&L − cost to decide. The LLM desks pay real API money per session; the quant pays nothing. So the quant starts every race with a structural head start: its net is its P&L. An LLM doesn't just have to trade well. It has to out-trade its own bill.
That surfaced a result we love to point at: in the crypto race, the Claude desk was the only desk actually making money trading… and it finished last among the crypto desks, because its API spend turned a real gross profit into a net loss. The quant, meanwhile, sat flat. Its regime gate kept it out of a choppy tape, and not trading cost it nothing.
Capability isn't the bottleneck anymore. Unit economics is. A model that's right but expensive can lose to a free heuristic once you count the compute.
The honest caveats
This is a measurement in progress, not a proven edge. The sample is still small, and a $0 algorithm "holding its own" over a few weeks is not the same as an edge over hundreds of trades. We label that clearly, and the ledger is append-only so the record can't be quietly cleaned up. The quant could absolutely fall behind as the sample grows; that's the whole point of running it in the open.
What's not in doubt is the framing: the cheapest contestant changes how you read every other row.
Watch it play out → Read the methodology
100% simulated. $1,000 of pretend money per desk, live quotes, honest fills. Not financial advice.