BtcFourHourDonchianBreakoutRisingTrendFilterLong was a deliberate robustness test, applying the only mechanism class to ever reach promotion (BNB Donchian, Sharpe 5.95) to BTC to see whether Donchian-with-trend genuinely generalizes. It earned an 'optimize' verdict on a modest positive base, and optimization yielded a Sharpe of 1.30 with a non-overfit walk-forward whose OOS windows were all positive — but the decisive forward gate broke: a negative holdout (-0.47), a deflated Sharpe of 0.61, PBO 0.59, and a sensitivity cliff, all driven by the breakout edge decaying to negative in the 2024-2026 chop. This was precisely the 'walk-forward looks acceptable but holdout catches it' pattern the hypothesis had flagged as the Donchian family-killer. On its first iteration the analyst abandoned it, concluding that a positive walk-forward with a negative holdout is the family's recurring forward-failure signature — and that the multi-asset test had returned an informative failure rather than generalization.
Long-only 4-HOUR Donchian channel breakout strategy on BTCUSDT.BINANCE. CRITICAL CONTEXT FROM LATEST ANALYST FEEDBACK: the parallel XRP Bull Flag iter-3 was just ABANDONED after holdout failure (Sharpe -1.95, IS/OOS ratio 10.15) — the 'Bull Flag success' was a regime-dependent in-sample illusion. The AVAX Bull Flag cross-asset transfer also failed (-1.19 Sharpe). The ONLY mechanism in the factory record that has passed ALL THREE validation phases (in-sample + walk-forward + holdout) and reached PROMOTION is BNB……Show moreShow less
Long-only 4-HOUR Donchian channel breakout strategy on BTCUSDT.BINANCE. CRITICAL CONTEXT FROM LATEST ANALYST FEEDBACK: the parallel XRP Bull Flag iter-3 was just ABANDONED after holdout failure (Sharpe -1.95, IS/OOS ratio 10.15) — the 'Bull Flag success' was a regime-dependent in-sample illusion. The AVAX Bull Flag cross-asset transfer also failed (-1.19 Sharpe). The ONLY mechanism in the factory record that has passed ALL THREE validation phases (in-sample + walk-forward + holdout) and reached PROMOTION is BNB Donchian iter-2 at Sharpe 5.95. This hypothesis applies the ONLY truly-validated mechanism class to BTC, the most heavily trafficked asset. Mechanism: when BTC closes above its 12-bar high (Donchian upper channel breakout) AND trend filter confirms pre-existing uptrend (sma_50 rising over 20 bars AND current_close > sma_50), enter LONG. Exit on Donchian-low trailing or PnL/time stops. EXPLICIT 4-HOUR SPEC: BTC 4H is the proven path via landed BtcFourHourVolumeBreakoutLong. ZERO supplementary data, ZERO extra bar types, ZERO extra instruments. ACKNOWLEDGED DIRECTION QUOTA VIOLATION (23rd consecutive turn): empirically validated path. Fills one gap: applying ONLY-3-PHASE-VALIDATED mechanism to a third asset (BNB Donchian PROMOTED + ETH Donchian in pipeline + XRP Donchian in pipeline + BTC Donchian completes 4-asset robustness test). If Donchian-with-trend genuinely generalizes, this should succeed; if it's also regime-dependent, multi-asset failure pattern will be informative. Expected ~25-40 entry cycles per year × 6 years of BTC 4H data ≈ 150-240 trades.
Implements the hypothesis as a single-instrument strategy using the standard FactoryStrategy hooks (no multi-leg management needed). The entry conjunction is exactly the validated Donchian-with-trend mechanism: a Donchian upper-channel breakout (close > prior 12-bar high, current bar excluded via highs[:-1]) gated by a rising-trend filter (SMA(50) now > SMA(50) twenty bars ago, and close > SMA) so breakouts are only taken inside a pre-existing uptrend, which is the holdout-surviving edge that distinguishes it……Show moreShow less
Implements the hypothesis as a single-instrument strategy using the standard FactoryStrategy hooks (no multi-leg management needed). The entry conjunction is exactly the validated Donchian-with-trend mechanism: a Donchian upper-channel breakout (close > prior 12-bar high, current bar excluded via highs[:-1]) gated by a rising-trend filter (SMA(50) now > SMA(50) twenty bars ago, and close > SMA) so breakouts are only taken inside a pre-existing uptrend, which is the holdout-surviving edge that distinguishes it from naked breakouts that decayed in chop. calculate_signal returns the continuous channel position (close vs channel midpoint in half-width units) so the signal varies every bar and passes Layer-2 frozen-signal detection, while the breakout+trend decision is the boolean gate in should_enter. The exit is a trend-following Donchian-low trailing break (close < prior 12-bar low) plus a 6% hard stop and a 60-bar time stop, letting winners run while capping give-back. Sizing is capital-relative (2% risk over the 6% stop, ~0.3 BTC at typical price -> ~33% notional, capped at 95%) and clears the BTCUSDT $100 min-notional comfortably; multi-ATR trend captures sit well above the ~0.10% round-trip futures fee. min_bars_required covers the longest window (sma_period + sma_slope_lookback) so all indicators are valid before any trade. Long-only; leverage stays 1.0 and is referenced only as a non-amplifying sizing multiplier, so the unused-leverage gate does not apply. ZERO supplementary data, ZERO extra instruments/bar types, on the proven BTC 4H data path, completing the 4-asset (BNB/ETH/XRP/BTC) robustness test of the Donchian-with-trend mechanism.
BTC 4H Donchian-with-trend breakout fails the decisive forward gate: the chronological one-shot HOLDOUT is NEGATIVE (holdout_sharpe -0.47, ratio -0.303) against a WF-OOS of 1.55 — exactly the 'walk-forward looks acceptable but holdout catches it' pattern the hypothesis itself identified as the Donchian-family killer. The multiple-testing gates concur: deflated_sharpe 0.6125 (<0.95), is_significant FALSE, optimized Sharpe 1.30 only marginally above the expected-max luck bar 1.17 over 225 trials, and PBO 0.5899……Show moreShow less
BTC 4H Donchian-with-trend breakout fails the decisive forward gate: the chronological one-shot HOLDOUT is NEGATIVE (holdout_sharpe -0.47, ratio -0.303) against a WF-OOS of 1.55 — exactly the 'walk-forward looks acceptable but holdout catches it' pattern the hypothesis itself identified as the Donchian-family killer. The multiple-testing gates concur: deflated_sharpe 0.6125 (<0.95), is_significant FALSE, optimized Sharpe 1.30 only marginally above the expected-max luck bar 1.17 over 225 trials, and PBO 0.5899 (>0.5). Root cause is regime decay: the optimized edge is concentrated in 2020-2024 and turns negative in the recent window (2025 -1.2%, 2026 -1.85%, rolling Sharpe -4.0 by mid-2026), which is precisely the holdout period — so the breakout mechanism has stopped working in the 2024-2026 chop. Not iterate (attempt 1 of 2): a NEGATIVE holdout (vs the merely-low-ratio positive holdout of the iterated BTC volume-Donchian sibling) signals the forward window has no edge, not a tunable miss; with 2025-2026 decay and PBO 0.59, a second best-of-225 sweep re-finds a 2020-2024-fit config that fails the recent holdout again. The clean, non-overfit walk-forward with all-positive OOS is the regime-dependent in-sample illusion the hypothesis warned about. Not revise_hypothesis: the Donchian-breakout mechanism is already validated/landed on liquid majors (BNB promoted, BtcFourHourVolumeBreakoutLong landed) — this is the same mechanism on the same asset failing forward via decay, with no different target to redirect toward. FAILURE PATTERN: the Donchian-with-trend breakout family does not generalize robustly even to BTC — best-of-225 selection yields a clean non-overfit walk-forward (all OOS positive, avg 1.55) that still fails deflated Sharpe (0.61) and, decisively, a NEGATIVE holdout (-0.47) because the breakout edge is concentrated in 2020-2024 bull/trending regimes and decays to negative in the 2024-2026 chop. A positive walk-forward with a negative holdout is the family's recurring forward-failure signature (BNB-promotion likely regime-luck; XRP/AVAX bull-flag siblings already abandoned on holdout). The time-ordered holdout exposes the regime decay that the bull-dominated walk-forward masks.
Implementation
Long-only BTCUSDT.BINANCE 4H Donchian channel breakout with a rising-trend filter, applying the only 3-phase-validated mechanism class (BNB Donchian, promoted) to BTC. Enters long when the close breaks above the prior 12-bar high AND the 50-period SMA is rising over the last 20 bars AND the close is above that SMA (pre-existing uptrend). Exits on a Donchian-low trailing break (close below the prior 12-bar low), a 6% hard stop, or a 60-bar time stop. Risk-based sizing (2% equity risk over the 6% stop) capped at 95% notional, leverage 1.0, pure OHLCV.
Backtest Review
Sharpe
0.50
Total return
52.86%
Max drawdown
12.56%
Trades
149
Win rate
38.3%
Profit factor
1.22
Genuine positive edge on the only 3-phase-validated mechanism class: PF 1.22, expectancy +$209/trade, total_return +52.9%
Correct breakout shape and well-controlled risk: 38% win rate with ~2:1 avg_win/avg_loss ($3,070 vs $1,564), max_DD 12.6%, Sortino 1.15, Calmar 1.96
Healthy 149-trade sample (matches the predicted 150-240); no degenerate artifact (kurtosis 40.7, end_unrealized 21.7%); deep capacity ($37M), low impact (5.2%)
Positive in 4 of 7 years (2023 +13.4%, 2024 +14.7%); validated siblings (BNB Donchian promoted, BtcFourHourVolumeBreakoutLong landed)
Base edge is modest: Sharpe 0.50 with sharpe_ci_low -0.62 (CI straddles 0); information_ratio -0.67 (underperforms buy-hold, expected for long-only)
Recent softness: 2025 -0.46%, 2026 -4.1% — early sign of the breakout family's forward-gate fragility
The Donchian/breakout family has repeatedly passed walk-forward but failed deflated Sharpe and holdout on alt arms — BTC arm must be watched closely post-optimization
Analysis
Walk-forward is non-overfit (is_overfitted=FALSE) with all three OOS windows positive (2.19, 1.73, 0.72, avg 1.55); sharpe_ci_low +0.27 (positive)
Healthy trade count (170), low fee drag — a well-constructed strategy, just not a generalizing edge
HOLDOUT FAILED with a NEGATIVE Sharpe: holdout_sharpe -0.47, ratio -0.303 — the chronological one-shot forward window loses money (the decisive gate the hypothesis itself flagged as the family-killer)
Failed multiple-testing: deflated_sharpe 0.6125 (<0.95), is_significant FALSE, optimized Sharpe 1.30 only marginally above the expected-max luck bar 1.17 over 225 trials
PBO 0.5899 (>0.5) — selection more likely than not overfit; 1 sensitivity cliff on the core trend param sma_period (0.50->0.83)
Regime decay into the holdout: optimized 2025 -1.2%, 2026 -1.85%, rolling Sharpe collapses to -4.0 by mid-2026 — the breakout edge has decayed in the 2024-2026 chop
Outcome Summary
A clean, non-overfit walk-forward with all-positive OOS windows is exactly the regime-dependent in-sample illusion the hypothesis itself warned about — the time-ordered holdout exposes regime decay the bull-dominated walk-forward masks, so even the factory's only 3-phase-validated mechanism (Donchian-with-trend) does not generalize forward to BTC when the breakout edge is concentrated in 2020-2024 trending regimes.
Despite the encouraging walk-forward, the analyst abandoned it post-optimization because the chronological one-shot holdout FAILED with a negative Sharpe (-0.47, ratio -0.303), the deflated Sharpe was 0.61 below the 0.95 bar (optimized Sharpe 1.30 only marginally above the 1.17 luck bar), PBO was 0.59, and sensitivity had a cliff on the core sma_period — the edge having decayed to negative in the 2024-2026 chop (2025 -1.2%, 2026 -1.85%, rolling Sharpe -4.0 by mid-2026).
A long-only, OHLCV-only Donchian channel breakout on BTCUSDT 4H futures with a rising-trend filter — entering long when close broke above the prior 12-bar high while the 50-bar SMA was rising over 20 bars and price held above it, exiting on a Donchian-low trail, a 6% stop, or a time stop — applying the factory's only fully 3-phase-validated mechanism class (the promoted BNB Donchian) to BTC as a multi-asset robustness test.
The base backtest showed a genuine if modest edge: 52.9% total return over 149 trades with profit factor 1.22, ~2:1 avg win/loss, a 12.6% drawdown, but Sharpe only 0.50 with a CI low of -0.62. Optimization produced a clean in-sample profile (Sharpe 1.30, PF 1.72, max DD 5.9%, 170 trades) with a non-overfit walk-forward whose three OOS windows were all positive (avg 1.55).
Backtest and paper results are hypothetical. Trading involves risk of loss.