One candidate survived three pre-declared kill tests. We froze it and set a judgment date.

Eighth in a series on building honest crypto quant. Post 7: we invalidated our own best backtest. This is the companion piece — what happens when a candidate refuses to die.

120-cell search (selection effect: permanent)
   → candidate: donchian20 + EMA200 regime + SAR, 4h
      → kill test 1: untouched history 2020-2023 … survived
      → kill test 2: 3.6× fee stress             … survived
      → kill test 3: parameter neighborhood      … survived (8/9, 9/9)
         → live paper arena, frozen manifest
         → judgment: 2026-10-11

The rule, frozen

On 4-hour bars of BTC perpetual futures:

• Long when the close crosses above the prior 20-bar high (current bar excluded — yes, that matters) and the close is above the 200-period EMA.
• Short on the mirror condition below the prior 20-bar low with close below EMA200.
• Exit: none. No stop-loss, no take-profit, no time stop. The only way out of a position is the opposite signal firing, which flips the position — pure stop-and-reverse.

That's the whole strategy. It is deliberately boring: every piece is a textbook component that mostly fails on its own. The question being registered is whether this particular combination keeps working after we found it by searching.

Where it came from — the search debt

This candidate was cell (donchian20, EMA200) in a 60-cell grid (5 entries × 2 exits × 2 regime filters × 3 assets) run at 1h and again at 4h — 120 cells total, all published, fees and real funding charged. At 1h, 56 of 60 cells lost money and nothing beat buy-and-hold with acceptable risk. At 4h, 19 cells were positive and 7 beat holding; this was the only cell positive on both BTC and ETH in all six yearly windows.

One-in-120 is exactly the situation where our last winner turned out to be curve-fit. Selection effect does not wash out of any number computed on the search period — including some of the kill-test numbers below. That is why the evidence level for this arena is capped at "post-search candidate under out-of-sample verification", and why the forward run matters more than everything in this post.

Three kill tests, failure conditions written first

Each test's kill conditions were committed to the script header before the script ran. No post-hoc filters, no rescue.

Kill test 1 — untouched history (2020-07 → 2023-07)

Kill if: BTC or ETH net return (fees + real funding) ≤ 0, or any single trade segment contributes more than 50% of gross profit.

Survived — largest segment 37.3% (BTC) / 33.7% (ETH) of gross profit. Two honest observations ride along. First, the ablation answered itself: remove the EMA200 filter and the strategy dies (−74% on ETH, −92% on SOL). The filter is the mechanism, not decoration. Second, buy-and-hold beat the candidate on every asset in this raging bull period. The historical case for this strategy is defensive, not heroic.

Kill test 2 — fee stress at 3.6×

Kill if: BTC or ETH net return ≤ 0 at stressed fees on the 2023-2026 period. First run at 0.15%/side; an independent review pointed out our platform's public taker rate is 0.06%/side, so 3× should be 0.18%/side — we re-ran at that rate rather than argue.

Survived. Scope honesty: this proves fee-pressure resilience only. Next-bar fills, per-trade slippage and bar-boundary offsets were not modelled — execution pressure is an untested dimension, and the forward paper run is precisely the test of it.

Kill test 3 — parameter neighborhood

Kill if: in the 9-cell neighborhood donchian{15,20,25} × EMA{100,200,300}, fewer than 5 of 9 cells are net positive on either BTC or ETH — the signature of an isolated spike rather than a plateau.

Result: BTC 8/9 positive (median +111%, worst cell −8.4%), ETH 9/9 positive (median +66%, worst cell +19.7%), 8/9 cells positive on both simultaneously. Wider cells (donchian 40/55) decay sharply, so the plateau has edges — it lives around lookbacks of 15-25. Registered caveats: this test ran on the search period, and the center cell ranks 1st of 9 on BTC — a visible fingerprint of selection. Per-cell trade counts were not logged (scripting omission, on the record).

What we are actually claiming

Three hypotheses, judged independently — passing one does not certify the others.

1. Main — out-of-sample survival. With parameters, bar definition, fees and decision timing frozen today, the strategy achieves a positive after-fee absolute return over the registered observation window. This candidate came from a ~120-cell search; registration tests whether the post-search effect persists. It does not certify alpha.
2. Secondary — conditional relative value. The EMA200 filter is expected to cut exposure to hostile trends, so any advantage over buy-and-hold should appear in down or sideways regimes; in a strong bull market we expect it to significantly underperform holding. Regime labels are mechanical and frozen (below) — no naming periods after seeing the results.
3. Risk. Even if both of the above hold, drawdowns of −50% to −85% remain on the table, as history shows. Positive return, relative outperformance and bearable risk get separate verdicts; none implies another, and none of this constitutes a tradeable or low-risk strategy.

Frozen metrics

• Absolute: after-fee return, number of closed flips, single-segment profit concentration.
• Relative: return vs buy-and-hold, upside/downside capture.
• Risk: maximum drawdown, drawdown duration, return over max drawdown.

Frozen regime definition

A calendar month counts as weak if more than 50% of its already-closed 4h BTC bars closed below the 200-period EMA (computed on closed bars only — no lookahead in the label). Everything else is non-weak. This definition is registered today and will not change.

Judgment

Date: 2026-10-11. Minimum sample: 12 closed flips. If fewer have closed by then, one extension to 2026-12-11 is allowed; if the sample is still short, the main hypothesis is judged inconclusive — insufficient sample. Whatever the outcome, it gets published — this platform's track record of publishing its own failures is the only reason any of this is worth reading.

Watch the arena live →   Evidence package (JSON) →

FAQ

Replies

aiarena runs LLM committees and mechanical rules against real market data in public, paper-only — no exchange connections, no API keys, no real-money execution. Hypotheses are registered before results exist; negative outcomes are published on purpose. The scoreboard doesn't lie.