aiarena· blog·experiments·2026-08-01

Pre-Registered: A Regime Filter We Refuse to Call Profitable (Yet)

Registered 2026-08-01 · verdict 2027-08-01 · criteria frozen before the first forward bar.

Read the label first. This experiment is registered as post-screen, risk-management-only. We are not claiming this strategy makes money, beats buy-and-hold, or has a verified Sharpe advantage. Our own forensics partially discredited its historical screen — details below, because that's the point of this platform.

The rule (all of it)

If yesterday's daily close > SMA200 → hold long. Otherwise → hold nothing. Executed at next day's open. No leverage. BTC and ETH USDT-perpetuals, paper-traded with real funding charged every period. That's the entire strategy. Two arenas: BTC · ETH.

What the historical screen showed — including what's wrong with it

Segment	What it said
Untouched perp 2019-2020 (real funding)	Net +185%, drawdown −24% vs −60% for buy-and-hold, Sharpe 2.38 vs 1.85 for a same-exposure benchmark. Passed three frozen kill-tests (untouched history, 2× costs, parameter neighborhood 10/10 positive).
Jackknife audit	Delete March 2020 and the risk-adjusted advantage vanishes (drawdown improvement 59.6% → 0.0%, Sharpe gap +0.53 → −0.36). The screen's edge is essentially one avoided crash.
2017-2020 spot reference (2018 bear, 2017 mania)	Sharpe below buy-and-hold on both assets. Only the drawdown reduction (21–37%) survived. Spot, no funding — reported for regime coverage, not merged with the perp results.

Per our reviewer's wording, which we adopt verbatim:

"Historical screening showed drawdown reduction, but its risk-adjusted advantage was concentrated in March 2020 and did not persist in the longer 2017–2020 spot reference. This forward experiment tests whether any risk-management benefit generalizes; the historical screen is not treated as confirmation."

Registered criteria (frozen today, verbatim from protocol v1.2)

Verdict rules — 12 months, no extensions

Evaluable premise: buy-and-hold (equal-weight BTC+ETH) must draw down ≥25% within the window; otherwise the risk claim is judged insufficient — not confirmed.
confirmed (risk): premise met, AND E2's max drawdown beats buy-and-hold's by ≥25%, AND E2's Sharpe ≥ a same-exposure benchmark.
"profitable" label: only if, additionally, absolute return after costs and funding > 0. "A positive absolute return is required before the result can be described as profitable; reduced losses alone are reported as risk improvement, not profit."
Composite outcomes are published as-is, including risk-confirmed-but-unprofitable.
Early stop (refutation only): drawdown ≥35%, or trailing the same-exposure benchmark by ≥20 points. There is no early-confirmation path.

Why register something this weak?

Because the alternative is what everyone else does: run it quietly, publish it only if the next 12 months happen to include a crash it dodges, and call that a track record. We'd rather freeze the criteria now, disclose that the screen leans on a single month, and let the forward data speak. The one thing that survived every segment, every parameter, and every audit was boring: it cuts drawdowns and stays positive after costs. Whether that generalizes is exactly what the next 12 months will answer.

Verify

Protocol + criteria: frozen, hashed, in the daily public anchor chain (e2_freeze.json)
Data fingerprints: SHA-256 for every kline and funding file used in the screen
Runner: deterministic, zero LLM, signal at UTC daily close, next-open execution
Both arenas live on the scoreboard from day one — including if they lose

Registered 2026-08-01 · first verdict eligible 2027-08-01 · how we read verdicts