Trading Lab · Experiment brief
Wed Sep 9, 2026 · paper only · honest engine

Published 2026-09-09 13:54:16 CEST

Dropped Just hold Lab pick

The look-ahead fix kills the million-dollar paper KEEPs

This brief supersedes earlier Donchian and SMA paper dollars (BTC transfer, alts screen, Donchian-hold hybrid). Those numbers were produced by an engine that traded on the same bar it decided — peeking at the close, then filling as if you already knew it. After PR #16, a close[t] signal fills at open[t+1], and year-split / walk-forward keep strategy state across the boundary. Same $1,000 start, same 15 bps round-trip, same 95% size. The story changes completely.

BTC 2019-01-01 → 2026-09-08 ETH same window ALGO 2019-06-21 → 2026-09-08 15 bps RT · next-bar open · 95%
BTC Don 10/5 now
$2,598
Was (invalid)
$3.41M
BTC buy&hold
$19,879
ALGO 20/10 KEEP
$10,663
Read this first: do not use any KEEP/TWEAK/DROP from briefs dated Sep 6–9 that quote Donchian $3.41M / $8.8M / $81.4M or SMA $124k. Those are look-ahead artefacts. This file is the new scorecard.

What $1,000 becomes · full sample

Honest next-open fills. “Units” = how many coins your ending cash pile could buy at the last close.

BTC-USD · core sleeve

StrategyEnd $Max DDIn mktUnitsTrades
Buy & hold$19,879−76%100%0.251
Donchian 20/10$17,595−47%47%0.2241
SMA 5/20$12,163−61%54%0.1680
Donchian 10/5$2,598−70%43%0.0388

On BTC, hold wins. Donchian 10/5 — the old paper KEEP — now finishes with about one-eighth of buy&hold and far fewer coin-units. SMA 5/20 also loses to hold on dollars and units. Donchian 20/10 is the least-bad active rule: slightly under hold on $, softer crash (−47% vs −76%).

ALGO-USD · alts sleeve

StrategyEnd $Max DDIn mktTrades
Buy & hold$79−93%100%1
Donchian 10/5$886−82%33%70
SMA 5/20$1,177−89%44%86
Donchian 20/10$10,663−58%29%32

ALGO buy&hold is a wreck ($79 from $1,000). Old KEEP Donchian 10/5 still “beats hold” at $886 — but that is a loss vs start, not a growth strategy. The calmer 20/10 is the one that actually grows the paper $1,000.

ETH-USD · alts candidate

StrategyEnd $Max DDIn mktTrades
Buy & hold$17,691−79%100%1
SMA 5/20$13,054−61%51%79
Donchian 20/10$11,566−60%40%41
Donchian 10/5$31,808−50%44%72

ETH Donchian 10/5 still beats hold on the full sample ($31.8k vs $17.7k, milder dip). That is the only old KEEP that still looks good on dollars. Walk-forward, below, does not confirm it.

Old paper $ vs honest $

Same ideas, same fees, same 95% size. Only the fill timing changed.

Claim (old brief)Old $Honest $Now
BTC Don 10/5 KEEP$3.41M$2,598DROP
ALGO Don 10/5 KEEP$8.8M$886DROP
ETH Don 10/5 KEEP$81.4M$31,808TWEAK
BTC SMA 5/20 baseline$124k$12,163DROP vs hold
ALGO Don 20/10 TWEAK$1.0M$10,663KEEP
BTC Don hold 10/10 TWEAK$1.82M$10,389DROP vs hold

Year-split · each year starts at $1,000

Signals keep pre-January history (PR #16). 2026 is year-to-date through Sep 8. Beat = higher return than that year’s buy&hold.

BTC Don 10/5
2 / 8
BTC SMA 5/20
3 / 8
ALGO Don 20/10
6 / 8
ETH Don 10/5
5 / 8

BTC Donchian 10/5 only beats hold in 2022 (lost less in the crash) and 2026 YTD. It loses every bull year. Old brief said 8/8 — that was the look-ahead.

ALGO 20/10 is the clean alts stress: 6/8 years, avg +59%, worst dip −50% vs hold’s −93%. Lost only 2020 and 2021 (hold’s monster bounce years).

ETH 10/5 5/8 — still mixed. Beats in 2019, 2022, 2024–26 YTD; loses the 2020–21 and 2023 bulls.

Walk-forward (out-of-sample)

Train 365 days → test 90 → step 90 · fixed params · fresh $1,000 each fold. Same boundary-state rule as year-split.

ContenderFolds beat holdMean / foldWorst DD
BTC SMA 5/209 / 27 (33%)+11%−29%
BTC Don 10/58 / 27 (30%)+4%−39%
BTC Don 20/1013 / 27 (48%)+13%−25%
ALGO SMA 5/2014 / 25 (56%)+9%−50%
ALGO Don 10/515 / 25 (60%)+8%−46%
ALGO Don 20/1017 / 25 (68%)+14%−37%
ETH SMA 5/209 / 27 (33%)+13%−48%
ETH Don 10/511 / 27 (41%)+15%−28%
ETH Don 20/109 / 27 (33%)+14%−35%
Truth serum: a KEEP needs to beat hold outside the one long path. Only ALGO Donchian 20/10 clears that bar (68%). ETH Donchian 10/5’s full-sample win does not show up in rolling tests (41%). BTC Donchian 10/5 is worse than a coin flip (30%).

Secondary · Donchian-hold grid on BTC

Same 12-cell exit-stickiness grid as the Sep 9 hybrid brief, re-run on the fixed engine. The old “winner” (10/10, confirm=1, gap=0) now loses to hold ($10.4k vs $19.9k; 2/8 years; WF 11/27).

Honest-engine $ / Sharpe winner is entry=10, exit=15, confirm=2, gap=0: $26,148, units 0.33 vs hold 0.25, dip −46% vs −76%, 26 trades. Year-split 4/8. Walk-forward still only 11/27 (41%).

TWEAK, not KEEP. Full sample can beat hold with slower exits + a 2-bar confirm. Out-of-sample does not. Also: confirm_bars=2 is no longer the “trap” the old engine claimed — that was look-ahead talking. Do not promote a 12-cell winner that fails walk-forward.

Keep · Tweak · Drop

Dropped
donchian_breakout 10/5 on BTC — prior KEEP. Now $2,598 vs hold $19,879; 2/8 years; WF 8/27. Sells often and still eats a −70% dip. Demoted.
Dropped
donchian_breakout 10/5 on ALGO — prior KEEP ($8.8M). Now $886: beats a crashed hold but loses vs the $1,000 start. Not a growth rule. Demoted.
Dropped
sma_crossover 5/20 on BTC and ETH — old “baseline to beat.” Loses to buy&hold on dollars and coin-units. Family may still be searched on the honest engine (sticky SMA was never re-run here).
Dropped
donchian_hold 10/10 confirm=1 — prior grid winner / TWEAK. Now trails BTC hold on $ and units; 2/8 years.
Lab pick
buy_and_hold on BTC-USD — new paper baseline for the core sleeve. Deepest dip (−76%), but the active rules we trusted did not grow more coins after honest fills.
Lab pick
donchian_breakout 20/10 on ALGO-USD — only active rule that grows the $1,000, beats hold 6/8 years and 17/25 walk-forward folds, softer crash than hold. Alts-sleeve paper KEEP. Still 32 round-trips and a −58% dip — paper-live before any real talk.
Still testing
donchian_breakout 10/5 on ETH — still beats hold on the full sample ($31.8k vs $17.7k) and 5/8 years, but WF 11/27 fails truth serum. Not a KEEP until OOS improves.
Still testing
Donchian family on BTC (20/10, and hold 10/15 confirm=2) — softer dips than hold; 20/10 almost matches $; hold-hybrid beats $ in-sample. Both fail year-split / WF as KEEPs. Search knobs on the fixed engine if we keep this family. Do not reuse old million-dollar grids.
Still testing
sma_crossover 5/20 on ALGO — beats crashed hold ($1,177 vs $79) but trails Donchian 20/10 on every score that matters.

Paper-live books

The Donchian 10/5 paper-live ledgers (ALGO / ETH / BTC) were started from the old KEEP. They are path-step logs, not proof, and they do not inherit a KEEP after this revalidation. Do not read a few days of marks as confirmation.

Path to real

Screen
Stress
Paper-live
Real (needs OK)

BTC core is back at Screen/Stress: hold is the bar. ALGO 20/10 is the only rule that finished Stress. No live orders without an explicit per-trade OK.

Suggested next experiment

On BTC, treat buy&hold as the number to beat and either (a) stop trading overlays until a rule beats it on walk-forward, or (b) re-grid Donchian-hold / sticky SMA on this engine with a tiny knob set — then stress the winner. On alts, start (or restart) paper-live for Donchian 20/10 on ALGO, not 10/5. Leave ETH 10/5 as a TWEAK until OOS is honest.