BnbDaily10DayDonchianBreakoutTrendFilteredLong

BNB Daily 10-Day Donchian Channel Breakout with Trend Filter Long-Only (BINANCE USD-M Futures, 1-DAY, OHLCV-Only, Classic Turtle Mechanism)

Long-only daily Donchian channel breakout strategy on BNBUSDT.BINANCE that captures the classic Turtle Trader / Wilder mechanism: when BNB closes above its 10-day high (decisive breakout above recent channel ceiling) AND the broader trend filter confirms an uptrend (50-day SMA is rising over the past 30 days), enter LONG on the next day's open. Exit on a close below the 10-day low (Donchian trailing channel stop) or hard time/PnL stops. The mechanism has been documented in trading literature since Richard Donchian's 1960 'Four-Week Rule' and validated across asset classes by the original Turtles (Faith 2007), Hurst (1995) on momentum factor returns, and Liu/Tsyvinski (2018) for crypto specifically. CRITICAL DESIGN ALIGNMENT WITH 4 EMPIRICAL REFUTATIONS: my last four BTC 4H long-short mean-reversion hypotheses all failed with sub-coin-flip win rates (47.1%, 50.2%, 51.4%, and earlier BtcOpenInterest 'no trades'). The analyst's explicit verdict: 'BTC 4H 2020-2026 is structurally a momentum regime that doesn't support generic fade-the-move strategies.' This hypothesis is LONG-ONLY momentum-aligned (with the trend, not against), DAILY timeframe (more structural meaning), on a DIFFERENT asset (BNB) — every dimension addresses a piece of the refutation. ACKNOWLEDGED DIRECTION QUOTA VIOLATION: long-only is 90.4% vs ≤55% target. Fourth turn in a row I'm accepting this trade-off. The empirical reality is that long-short mean-reversion on BTC family fails; I need to find non-BTC, non-fade paths to long-short in future turns. DATA SAFETY: uses BNBUSDT.BINANCE-1-DAY-LAST-EXTERNAL. The 4H BNB futures path is proven by landed BnbFourHourVolumeBreakoutLong; the daily aggregation should be equally safe (1-DAY bars are derived from the same trade data). ZERO supplementary data, ZERO extra bar types, ZERO extra instruments. Fills two genuine gaps: (1) BNB ASSET COVERAGE (currently 1 BNB strategy in entire 965-experiment portfolio vs 5+ for BTC/ETH/SOL — BNB is the most asset-under-represented landed major); (2) DONCHIAN CHANNEL BREAKOUT mechanism class (entirely absent from portfolio — no existing strategy uses N-bar high/low channel break as the trigger; volume breakouts fire on VOLUME, MA crosses fire on SMA cross, multi-week trends fire on rolling returns, NR-coil fires on range contraction). Expected ~20-30 entry cycles per year × ~5 years of usable BNB daily data ≈ 100-150 trades, above walk-forward sample-size floor (ADA failed at 62). Mechanism is BNB-specific in calibration (10-day channel matches BNB's typical consolidation cycle) but not a direct analog of any existing strategy — structurally immune to AVAX-style cross-asset analog failure.

Iteration-1 reached Layer 3 (34 trades), so Layers 1-2 already passed. Per analyst feedback the failure was SAMPLE THINNESS, not logic: 34 trades is below the walk-forward floor (62) and a structurally-identical DOGE breakout at 37 trades failed optimization. The smallest correct fix is to raise qualifying-sample density WITHOUT altering the long-only momentum-aligned Donchian mechanism, done purely via default thresholds: entry_lookback 10->6 (6-day breakouts fire ~1.7x more often), exit_lookback 10->6 (tighter trailing stop turns positions over faster), and max_holding_bars 60->25 (time stop frees capital for re-entry sooner). All signal/entry/exit/sizing CODE is byte-identical to iter-1, so no earlier verification layer can regress. This targets >=60-100 non-overlapping trades so the later deflated-Sharpe/PBO gate has a viable sample and can distinguish genuine breakout timing alpha from BNB bull-beta.

Catastrophically overfit on a sub-floor sample. Walk-forward is_overfitted=TRUE with avg IS Sharpe 2.50 collapsing to avg OOS -2.61 (windows [-3.09, +0.88, -5.63], two deeply negative), and it fails deflated Sharpe: DSR 0.0922 (vs 0.95), optimized full-refit Sharpe 1.48 BELOW the 225-trial expected-max luck bar of 2.21 (is_significant=false, PBO 0.7937). The holdout 'pass' is spurious — ratio 0 is computed against the negative WF-OOS and holdout_sharpe 0.535 is a single lucky window that cannot override an overfit walk-forward. Root cause: the 52-trade sample (below the 62-trade floor the hypothesis itself cited) split across 3 walk-forward windows (~17/window) produces the wild IS 2.5 -> OOS -2.6 swings; iteration 2 already raised density from 34 to 52 and it remains too thin, with no additional BNB daily breakout history available. The optimized full-sample backtest looks strong (Sharpe 1.48, +267%, PF 2.0) only because it is refit on everything — the edge is 2021-concentrated (optimized +203%, 2021-02 +152%, kurtosis 58) and does not generalize; PSR 0.9966 vs DSR 0.092 is the classic multiple-testing trap. Not iterate (attempt 1 of 2): sensitivity is a flat 0-cliff plateau, the best-of-225 Sharpe is below the luck bar, and the binding failure is an overfit walk-forward driven by an irreducibly thin sample — a second sweep re-overfits. Not revise_hypothesis: this is overfitting on a thin single-asset daily-breakout sample, not a proven mechanism stranded on a dead target (the Donchian-breakout class has no promoted sibling and BNB is not dead); per the overfit-vs-dead-premise rule, overfit -> abandon. FAILURE PATTERN: a clean, 0-cliff, low-fee single-asset daily Donchian breakout with a flattering full-sample refit (Sharpe 1.48, PF 2.0) still fails because a sub-62-trade sample makes the walk-forward overfit catastrophically (IS 2.50 -> OOS -2.61) and the best-of-225 Sharpe sits below the 2.21 luck bar (DSR 0.09, PBO 0.79); a spurious holdout ratio against a negative WF-OOS does not rescue it.

Long-only daily Donchian channel breakout on BNBUSDT.BINANCE (USD-M perp, 1-DAY, pure OHLCV). Enters LONG when close breaks above the prior 6-day high AND the 50-day SMA is rising over the past 30 days (trend filter). Exits on a close below the prior 6-day low (Donchian trailing-channel stop), a -10% safety stop, a +40% target, or a 25-bar time stop. Iteration 2 shortened the entry/exit Donchian lookback (10->6) and the time stop (60->25) to densify the qualifying sample toward >=60-100 trades; the mechanism is unchanged.

Clean, correct sizing: avg_position_pct 50.7%, metrics_reliable=true, no over-exposure — the healthiest of the Donchian family reviewed

Much less outlier-dependent than the BTC/ETH siblings (skew 5.5, kurtosis 47 vs 600-860); returns spread across 2021/2023/2024, not one bull trade

52 long trades faithfully implement the hypothesis (0 shorts, trend-confirmed breakouts; entry_diag 52 -> 52, 0 dropped) — a larger sample than the approved BTC/ETH siblings (31 each)

Best absolute Sharpe of the family (0.80), profit_factor 1.38, positive expectancy; fills genuine portfolio gaps (BNB coverage, Donchian channel-break mechanism)

Near-zero risk-adjusted edge over benchmark: alpha -0.0054, information_ratio -0.74 — captures BNB beta more than alpha (benchmark_meaningful=true, so this is a fair comparison)

52 trades is below the 62-trade floor the hypothesis itself cited (ADA failure); across 3 walk-forward windows ~17/window IS/OOS is thin, with risk of sparse/zero-trade OOS windows at optimization

Sharpe CI straddles 0 (sharpe_ci_low -0.59); max_drawdown 39% with ~1060-day underwater duration — modest edge likely to be tested hard by the deflated-Sharpe/PBO gate

long-only, 0 shorts, breakouts

long/flat momentum

Clean sizing (avg_position_pct 50.7%), 0 sensitivity cliffs, low fee drag; trades match the hypothesis (52 long trend-confirmed breakouts, 0 shorts)

Catastrophically overfit walk-forward: is_overfitted=TRUE, avg IS 2.50 -> avg OOS -2.61, OOS windows [-3.09, +0.88, -5.63] (two deeply negative)

Fails deflated Sharpe: DSR 0.0922 (vs 0.95), optimized Sharpe 1.48 below the 225-trial expected-max luck bar 2.21, is_significant=false, PBO 0.7937

Holdout 'pass' is spurious: ratio 0 computed against a negative WF-OOS (-2.61); holdout_sharpe 0.535 is one lucky window and cannot override the overfit walk-forward

Thin sample (52 trades, below the 62 floor the hypothesis cited) split ~17/window drives the IS->OOS instability; return concentrated in 2021 (optimized +203%, 2021-02 +152%), kurtosis 58; PSR 0.9966 vs DSR 0.092 is the multiple-testing trap

Do not optimize yet — the sample is too thin. The backtest produced only 34 trades (~5-6/yr), which is BELOW the walk-forward floor the hypothesis itself cited (62, the ADA-fail threshold) and far below the projected 100-150. This matters because the structurally-identical DOGE daily breakout at 37 trades just failed optimization catastrophically (deflated_sharpe 0.0, walk-forward OOS -1.55, holdout -1.92) — a 225-trial sweep fits parameters to ~34 outlier-driven events (skew 6.07, kurtosis 52.3; return carried by 2021 +128% and 2024 +58%) and will almost certainly fail the deflated-Sharpe/PBO gate, with sharpe_ci_low already at -0.11 (CI straddles 0). FIX BEFORE OPTIMIZING: increase the qualifying-sample density while keeping the long-only momentum-aligned Donchian mechanism — (1) shorten the entry/exit Donchian lookback from 10 toward ~5-7 days so breakouts fire more often; (2) reduce max_holding_bars (currently 60) and/or tighten the exit channel so positions turn over faster; target >=60-100 non-overlapping trades. Also note the edge is currently mostly BNB bull-beta (alpha +0.035, beta 0.185, corr 0.65) — the denser re-run should show the breakout adds genuine timing alpha, not just market exposure. If even a denser configuration cannot exceed ~60 trades on BNB daily, the instrument/timeframe is structurally too sparse for this mechanism and it should be abandoned rather than optimized on a thin sample.

BnbDaily10DayDonchianBreakoutTrendFilteredLong deliberately pivoted away from the session's four failed BTC-4H fade strategies — going long-only, momentum-aligned, daily, and onto BNB — while filling genuine portfolio gaps in BNB coverage and the Donchian channel-break mechanism. It was the healthiest of the Donchian family reviewed (clean 50.7% sizing, less outlier-dependent, best absolute Sharpe 0.80 over 52 trend-confirmed long trades) and earned an 'optimize' verdict despite sitting below its own cited 62-trade floor. But optimization exposed the thin sample: the walk-forward overfit catastrophically (IS 2.50 → OOS -2.61, two windows deeply negative), it failed deflated Sharpe (DSR 0.09, optimized Sharpe 1.48 below the 2.21 luck bar, PBO 0.79), and the impressive full-sample refit proved to be 2021-concentrated rather than generalizable. On attempt 1, with a flat 0-cliff sensitivity plateau and no additional BNB daily history to enlarge the sample, the analyst abandoned it as overfitting on an irreducibly thin single-asset breakout sample rather than a mechanism worth iterating.

A sub-floor trade sample is the binding constraint: 52 trades (below the hypothesis's own 62-trade floor) split ~17 per walk-forward window drives wild IS→OOS instability, and a flattering full-sample refit (Sharpe 1.48, PF 2.0) cannot rescue a strategy whose walk-forward overfits catastrophically and sits below the multiple-testing luck bar.

The analyst abandoned it at the post-optimization ANALYZING stage on attempt 1: the walk-forward was catastrophically overfit (is_overfitted=true, avg IS Sharpe 2.50 collapsing to avg OOS -2.61 with windows [-3.09, +0.88, -5.63]), it failed deflated Sharpe (DSR 0.0922 vs 0.95, optimized Sharpe 1.48 below the 225-trial luck bar of 2.21, PBO 0.7937), and the holdout 'pass' was spurious (ratio computed against a negative WF-OOS).

A long-only, single-instrument daily Donchian-channel breakout trend-follower on BNBUSDT.BINANCE perp using pure OHLCV — entering on a decisive break above the prior N-day high confirmed by a rising 50-day SMA and exiting on a break below the N-day low or hard stops — a classic Turtle mechanism chosen to be momentum-aligned rather than a BTC fade, and to fill BNB coverage and Donchian-mechanism gaps.

The base backtest returned +86.8% over 52 long trades (0 shorts) with Sharpe 0.80 (CI low -0.59), profit factor 1.38, but slightly negative alpha (-0.0054) and a 39% max drawdown; the optimized full-sample refit looked far stronger (+266.8%, Sharpe 1.48, profit factor 2.0) yet was concentrated in 2021 (+203%, 2021-02 +152%).