aiarena · ← the arena · rule playground · blog

Break the Winner

A two-act exercise built entirely on this platform's real, archived research. Act one: make a strategy look good with two moves that feel innocent. Act two: freeze your kill conditions, then meet the history nobody searched on. There is no score, no leaderboard, and deliberately no deploy button.

What this is. Every number below comes from our published research records: a rule that silently never traded, a candidate selected from a 72-configuration exploratory grid, and its real test on untouched 2020–2023 history. Nothing is simulated for drama. The point is not to teach you to find pretty curves — it is to teach you why we refuse to believe them. Exploratory, post-selection data; Binance USDT-perp, 4h bars; backtests charged taker fees 0.05%/side + real per-period funding. Failure disclosure earns inspection, not belief.

Act I·A
One line of code

This breakout condition looks obviously correct: close > highest high of the last 20 bars. Toggle whether "the last 20 bars" includes the current bar (this is exactly the bug we shipped to production in July 2026 — full story):

Act I·B
One choice of window

Now meet a real candidate from our grids: Donchian-55 breakout, pullback entry to EMA20, channel exit, EMA200 regime filter — BTC, 4h bars, tested on the 2023–2026 search period. Here are all its windows. Which one would you put in the pitch deck? Click it.

full 3y: +16.3% year 1: +6.2% year 2: +1.5% year 3: +3.5%
MDD −33.0% · 58 trades · win rate 32.8% · every window positive. Looks robust, doesn't it?

Act II
Now try to kill it

That candidate was selected after searching 72 configurations, then selected again for its low-drawdown profile. All its numbers so far live inside the search period. The only honest next step is the one we actually took: test it on history that no search ever touched (2020-07 → 2023-07), with failure conditions frozen before looking.

Your turn. Freeze your three kill conditions first:

Net return on untouched history (fees + real funding) must be
No single trade may contribute more than of gross positive PnL
Max drawdown must stay under

What actually happened

This candidate is real. It was one of two "low-drawdown overlay" components we tried to qualify on 2026-07-12 for a portfolio hypothesis (B10). Both died on untouched history under pre-frozen kill conditions — the one above failed both the net-return condition and the low-drawdown identity condition. The portfolio hypothesis was archived the same day as rejected-at-component-qualification, with no component swapping allowed. Its search-period "every window positive" profile did not survive. That is not a story about one bad strategy; it is the reason search-period consistency is a screening tool, not evidence.

Canonical summary (quote this, not the drama): We tested a breakout-pullback candidate (Donchian-55/EMA20 pullback/channel exit/EMA200 regime, BTC 4h), selected post hoc from a 72-configuration exploratory grid, against untouched 2020–2023 history under kill conditions frozen before the test (positive net return incl. fees and funding; no single trade >50% of gross positive PnL; drawdown below 75% of the reference strategy's). It failed the net-return and drawdown conditions. This applies to the tested configuration; it does not establish that pullback entries or overlays are universally ineffective. Exploratory, post-selection. Full research notes are in the platform's public write-ups.
Read a full autopsy → Try rulelint → Write your own rule →

No login, no leaderboard, no deploy button — a rule that just demonstrated bias should not walk into a deployment funnel. The hash above is a local, in-browser commitment for this session only; it is not a public timestamp and not a platform preregistration. aiarena is paper-only: no exchange connections, no API keys, no real-money execution. Nothing here is investment advice.