COIN-M BTC Daily Donchian – Backtest

Hypothesis

A LONG-ONLY (LONG/FLAT, never short), SINGLE-INSTRUMENT, DAILY Donchian-channel breakout trend strategy on BTCUSD_PERP.BINANCE_CM (COIN-M inverse perpetual). Rationale: the engine over-exposure defect is now PROVEN specific to the BINANCE USD-M perp position_size->order-qty->equity pipeline (two clean long/flat 1x reproducers, BTC 429% and ETH 1062%, independent of flips/resizing), and BINANCE spot-CASH (MTM defect), multi-instrument (aggregation defect), and census signals (liq/OI/account-ratio/taker data wall)……Show moreShow less

A LONG-ONLY (LONG/FLAT, never short), SINGLE-INSTRUMENT, DAILY Donchian-channel breakout trend strategy on BTCUSD_PERP.BINANCE_CM (COIN-M inverse perpetual). Rationale: the engine over-exposure defect is now PROVEN specific to the BINANCE USD-M perp position_size->order-qty->equity pipeline (two clean long/flat 1x reproducers, BTC 429% and ETH 1062%, independent of flips/resizing), and BINANCE spot-CASH (MTM defect), multi-instrument (aggregation defect), and census signals (liq/OI/account-ratio/taker data wall) are all blocked. HYPERLIQUID sizes correctly but is data-length-limited (my HL BTC/ETH/SOL trend sleeve is pending on that risk). COIN-M is the one untested venue that combines (a) a DIFFERENT account/contract path (inverse, BTC-margined, dapi) that may sit OUTSIDE the corrupted USD-M position_size pipeline, with (b) DEEP multi-year history (BTCUSD_PERP since 2020) — unlike HL. COIN-M's only documented engine limitation is that inverse FUNDING is excluded; this book is PURE OHLCV (no funding), so that limitation does not apply. It applies the factory's VALIDATED daily trend edge on BTC (cleanest trender) in the simplest, bug-avoiding form: single-instrument, OHLCV-ONLY, LONG/FLAT (no flip-stacking), ONE fixed-fractional size at entry with NO dynamic resizing, leverage 1.0, close-to-flat before re-entry. It also fills the MOST under-represented venue (BINANCE_CM 1.8% vs 5% target) and has the lowest fees (~0.06% RT). PRIMARY PURPOSE: a clean-sizing probe of a deep-history venue — if COIN-M sizes correctly it is the first deep-history venue capable of validly backtesting the trend edge while USD-M is broken. PRE-REGISTERED ACCEPTANCE / SIZING CHECK: avg_position_pct ~<=100% and single-day equity returns within ~+/-15-20% (BTC at 1x; inverse PnL is mildly nonlinear but still bounded near BTC's daily move) — if violated, the over-exposure defect extends to COIN-M too (escalate). Low parameter count.

Iteration 2 makes the single change the QA verdict demands and nothing else. The blocking finding was measurability, not mechanics: 27 trades over 5.8 years put only ~5 in the last-20% holdout, so the >=10-trade holdout gate could never pass, and with that sample one $271k leg (7x avg_win) drove the +378.5% headline, PF 3.61 and the marginal Sharpe CI. The mechanism is unchanged - same Donchian breakout, same signal formula, same long/flat shape, same inverse contract sizing that the pre-registered probe……Show moreShow less

Iteration 2 makes the single change the QA verdict demands and nothing else. The blocking finding was measurability, not mechanics: 27 trades over 5.8 years put only ~5 in the last-20% holdout, so the >=10-trade holdout gate could never pass, and with that sample one $271k leg (7x avg_win) drove the +378.5% headline, PF 3.61 and the marginal Sharpe CI. The mechanism is unchanged - same Donchian breakout, same signal formula, same long/flat shape, same inverse contract sizing that the pre-registered probe validated (avg_position_pct 22.35%, no over-exposure on the CM account path). Only the sampling frequency moves: DAILY -> 12-HOUR with the channels shortened in calendar terms (entry 25d -> 10d, exit 10d -> 4d), which should give roughly 4-5x the trades so the holdout window carries a real sample and no single leg can dominate the distribution; per-trade edge shrinks proportionally but a 10-day BTC breakout swing is still an order of magnitude above the ~0.10% COIN-M round trip. target_frac 0.75 -> 0.60 pulls the flagged 64.4% drawdown (CI to 81%) toward the promotion floor without touching sizing shape. Layers 1-3 stay green: identical imports, class, bounded-deque O(1) state and contracts_for_notional path. Honest caveat the Research Lead owns: a long-only BTC trend book is beta by construction (benchmark_correlation 0.97, alpha -0.0051) and the faster channel will lower beta but cannot manufacture alpha - if the 100+ trade run still shows negative alpha, the family is refuted and should be abandoned rather than re-tuned.

data_unavailable: backtest data for BTCUSD_PERP.BINANCE_CM could not be collected after 5 retries: - bars (bar_type=BTCUSD_PERP.BINANCE_CM-12-HOUR-LAST-EXTERNAL): got 0, needed 50, or 90+ days span The data layer (not the strategy) is the blocker — re-evaluate this hypothesis once collection is fixed / the symbol is backfilled.

Implementation

Long/flat Donchian-channel breakout trend strategy on the BTC COIN-M inverse perpetual (BTCUSD_PERP.BINANCE_CM), now on 12-HOUR bars: enters long when the close breaks above the prior 20-bar (10-day) high, exits when the close breaks below the prior 8-bar (4-day) low. Pure OHLCV, single instrument, never short, one fixed-fractional notional (60% of equity) sized at entry via contracts_for_notional (whole $100 inverse contracts), leverage 1.0, no resizing or pyramiding.

Verification Results

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Iteration History

Verification failed (Layer 4 — QA review) [class=no_edge]: - UNMEASURABLE + OUTLIER-DRIVEN + NO ALPHA (identical to the ETH COIN-M Donchian twin). (a) 27 trades over 5.8 years: the hard holdout gate needs >=10 trades in the last-20% holdout, but ~5 land there, so it cannot pass by construction and optimization dies there (same trade-starvation as ETH twin [26], BTC CM Donchian [36], capitulation [18]). (b) The +378.5% headline is dominated by ONE trade: largest_win $271,411 vs avg_win $39,776 (~7x),……Show moreShow less

Verification failed (Layer 4 — QA review) [class=no_edge]: - UNMEASURABLE + OUTLIER-DRIVEN + NO ALPHA (identical to the ETH COIN-M Donchian twin). (a) 27 trades over 5.8 years: the hard holdout gate needs >=10 trades in the last-20% holdout, but ~5 land there, so it cannot pass by construction and optimization dies there (same trade-starvation as ETH twin [26], BTC CM Donchian [36], capitulation [18]). (b) The +378.5% headline is dominated by ONE trade: largest_win $271,411 vs avg_win $39,776 (~7x), end_unrealized_pct +91% (huge open winner); PF 3.61 collapses without that leg. (c) BTC beta, not skill: alpha -0.0051 (negative), IR -0.61, beta 0.77, benchmark_correlation 0.97. (d) Sharpe 0.685 with CI lower 0.0214 is 'marginally significant' only because of the single giant winner on 27 trades; max_drawdown 64.4% (CI to 81%). avg_trade 14.19% / PF 3.61 are tiny-sample tail artifacts. - Diagnostic outcome (the pre-registered sizing probe) PASSES: avg_position_pct 22.35% (<=100%), no over-exposure, no unbounded equity swings — the COIN-M inverse account path sizes correctly via contracts_for_notional (consistent with the ETH twin's 20.6%). The USD-M over-exposure defect does NOT extend to CM. Useful venue info, but doesn't make the book promotable.