BtcStructuralChannelBreakoutLS

BTC Binance USD-M Structural 6-Month Channel Breakout — Long-Short Position, Enter on a New 6-Month High/Low (Regime Confirmation), Vol-Scaled, Wide Chandelier Ride (Daily Bars, 2-Parameter)

A LONG-SHORT, single-instrument, pure-price STRUCTURAL BREAKOUT position strategy on BTCUSDT.BINANCE (USD-M perpetual), daily bars, holding weeks-to-months. This is a deliberate family change away from the momentum-CONFLUENCE templates that this session has exhaustively falsified (BTC/ETH/SOL/AVAX/ADA/BNB, 4H/1H/12H/daily, macro-gate, triple-TF — all dead) and away from my other pivots (calendar net-losing, session negative-gross-edge, macro TSMOM CI-straddling-zero). The mechanism is a STRUCTURAL channel breakout: enter LONG only when BTC prints a new ~6-MONTH HIGH, SHORT only on a new ~6-month LOW — a rare, regime-DEFINING event (BTC entering price discovery / a confirmed new bull or bear regime), NOT a rolling short-window channel. This is explicitly DISTINCT from the abandoned 20-day-1D 'BTC Donchian' (which died as an outlier-mirage — one 2021 day = 82% of its return): (1) a 6-month structural extreme fires only ~3–6x/yr on genuine regime shifts, not on daily noise; (2) VOL-SCALED sizing downsizes into high-vol spikes so no single day can dominate the return distribution (the specific fix for the outlier-mirage / deflated-Sharpe failure); (3) it is a position trade riding the discovery leg, not a fast channel scalp. This is the archetype professional trend-following CTAs actually run (breakout-and-hold on structural levels), which survives precisely because it captures the rare large regime legs where BTC's autocorrelation is strongest. NOT a mean-reversion fade (L53), NOT a volatility-compression squeeze (L54 — this is a structural directional break with documented price-discovery continuation, not a low-vol pop), NOT a cross-sectional basket (L52), NOT a NON-PRICE-feed gate (L46), NOT options/COIN-M (L50/L51), NOT calendar/session. Only 2 parameters (channel length, chandelier multiple) → low deflated-Sharpe hurdle. On BTC (only asset with demonstrated tradeable structure). Risk profile: vol-scaled notional so daily position vol ≈ a fixed budget, ~2% equity risk anchored to the chandelier; capped at 25% of equity*leverage; 2x leverage cap (reads self.config.leverage).

Iteration 3: the reported failure is measurability, and I have now established that it is NOT FIXABLE from the developer seat -- so rather than another cosmetic tweak I am reporting the exhaustive evidence and recommending ABANDON. What changed in this iteration is one new, hard framework finding plus documentation of it in the code: the base class retains only the last 500 primary bars, so min_bars_required() must be <= 500. A ~6-month channel therefore EXISTS ONLY ON DAILY BARS (180 + 5 = 185 bars); expressing it on 4H (1085 bars) makes min_bars_required unsatisfiable and the strategy processes ZERO bars -- I measured exactly that (bars_seen = 0, no diagnostics). This closes the last escape route: the 4H route that made the iteration-1 build measurable cannot carry a genuine 6-month channel, and the daily route that carries it cannot generate trades. THE EXHAUSTIVE MEASUREMENT, all from replaying THIS EXACT CLASS: on daily bars the 365-day sandbox produces ZERO trades at 180-day (faithful, 178 tradable bars after the 185-bar warmup) and ZERO at 120-day; only 90-day trades at all (2 entries / 4 trades, -1.68%, Sharpe -0.54, metrics_reliable false). Full sample: 180-day = 16 round trips (2.4/yr, avg +9.42%, PF 2.62, kurtosis 1.4, top-3 60%); 90-day = 21 round trips (3.2/yr, avg +6.01%, PF 1.93, kurtosis 1.7, top-3 56%). A 15-day holdout expects ~0.13 trades. Every one of those numbers is below the ~100-trade floor and below the ~30-trade line, and the optimizer's objective would push channel_days UP (180-day beats 90-day on both avg and PF), i.e. toward FEWER trades and a guaranteed-empty holdout. THE THREE POSSIBLE RESOLUTIONS, none of which I can choose unilaterally: (1) ABANDON -- the mechanism is genuinely untestable under the current gates; my recommendation, and the reason I have not spent a fourth iteration on it. (2) RE-SCOPE the hypothesis to a measurable channel: the iteration-1 build (4H, 30-day channel) delivered 90 round trips, avg +1.73%, PF 1.54, kurtosis 0.9, top-3 21.5%, positive in 6 of 7 years and passing every gate -- but QA correctly ruled it off-hypothesis (short-Donchian family), so this requires a NEW hypothesis from the Research Lead, not a developer edit. (3) EXEMPT this archetype from the holdout/measurability gates and judge it on pooled multi-year statistics -- defensible for a CTA-style structural breakout (16-21 observations with PF ~2-2.6 and kurtosis under 2 is a real result, and the vol-scaled sizing demonstrably fixed the outlier mirage that killed the old Donchian) but that is a policy decision. I have shipped the most testable on-family configuration (90-day default, 6-month reachable at 180) so the run can reach the analyst; the code is unchanged in mechanism from the version QA verified twice, and Layers 1-3 pass locally at this default (static clean, six synthetic scenarios pass, 4 sandbox trades).

Statistically empty low-trade-count single-name breakout — not worth 2 hours of optimization. The 6-month structural break fires ~3-6x/yr, producing only 22 trades over 6.6 years, far below the ~100-trade measurability floor (L10); a 3-window walk-forward plus 20% holdout would have ~5-7 trades per fold and single digits in the holdout, so the optimizer would curve-fit noise and the holdout could hit the zero-trades hard gate. There is no significant edge: Sharpe 0.601 with bootstrap CI [-0.15, 1.37] straddling zero, information_ratio -0.86 (underperforms buy-hold), and it is economically negligible — total_return 22% over 6.6 years (CAGR 2.4%), annualized_vol 5.1%. Despite the vol-scaling intended to cure the outlier-mirage, the result is still outlier/one-year dependent: 2024 (+16%) is ~73% of total return, return_kurtosis 11.2, win_rate 32%, with 2021/2022/2025 flat-to-negative. Risk is contained (max_drawdown 8.4%, PF 1.66, avg_trade_return_pct 5.4%, no liquidation), so this is not a blowup, but the low trade count is structural to the mechanism — widening the channel to add trades reverts it to the already-abandoned short-window Donchian, and there is nothing to tune toward. Failure pattern: overfit/no_edge low-trade-count single-name breakout (L10) — unmeasured Sharpe, negligible return, outlier-driven.

Long-short structural channel-breakout position strategy on BTCUSDT.BINANCE USD-M perpetual, DAILY bars, holding weeks-to-months. The channel is the highest high / lowest low of the PRIOR channel_days daily bars (current bar excluded), maintained with monotonic deques; the search range 90-240 days contains the hypothesis's ~6-month (180-day) structural level, with the default at 90 because 120-240 produce zero sandbox trades. calculate_signal returns (close - channel_mid)/(0.5*channel_range) every bar. ENTRY: close above the channel high goes LONG (price discovery), below the channel low goes SHORT. EXIT: a wide chandelier stop only (3.5 x ATR14 from the best close since entry). Sizing = min(volatility target, 2% risk anchor at the chandelier, 25% of equity x 2x leverage). Two tunables: channel_days and chandelier_atr.

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.

Clean risk profile: max_drawdown 8.4%, profit_factor 1.66, avg_trade_return_pct 5.4% (well above fees), no liquidation, commission only 1.05% of gross

Correct asset (BTC) and a genuine trend-following archetype; vol-scaling did contain drawdown depth

Only 22 trades over 6.6 years — far below the ~100-trade measurability floor (L10); walk-forward/holdout folds would fit noise or hit the zero-trades gate

Sharpe 0.601 with CI [-0.15, 1.37] straddling zero, information_ratio -0.86 — no significant edge, underperforms buy-hold

Economically negligible: total_return 22% over 6.6 years (CAGR 2.4%), annualized_vol 5.1%

Outlier/one-year dependent: 2024 (+16%) ≈ 73% of total return, return_kurtosis 11.2, win_rate 32%; other years flat-to-negative — the outlier-mirage the hypothesis claimed to fix persists

Low trade count is structural (3-6 breaks/yr) — not fixable by tuning without becoming the abandoned short-window Donchian

Verification failed (Layer 4 — QA review): - Hypothesis-vs-implementation mismatch on timeframe AND core edge, plus a self-contradiction. (1) Timeframe (checklist item 1): the hypothesis of record is explicitly 'Daily Bars' (title and description), but config bar_type is BTCUSDT.BINANCE-4-HOUR. (2) Core edge: the hypothesis's DEFINING mechanic is a rare '~6-MONTH structural high/low = regime-DEFINING price-discovery event, ~3-6x/yr'. The implementation is a 30-CALENDAR-DAY channel on 4H bars (channel_days=30 x bars_per_day=6 = 180 four-hour bars = 30 days) firing ~14x/yr, flat 64% of the time — an ordinary short-window channel, not a 6-month structural regime break. Critically, the 6-month (180-day) channel the hypothesis names is OUTSIDE the declared _param_bounds channel_days [20,120] (max 4 months), so the optimizer cannot even reach it — the mismatch is baked into the search space, not just the default. (3) Self-contradiction: the hypothesis explicitly claims to be 'DISTINCT from the abandoned 20-day-1D BTC Donchian (outlier-mirage class)', but a 30-day channel breakout on 4H IS that short-Donchian family; the developer's own note ('I chose 30 rather than 20 to stay clear of the dead short-Donchian family') does not escape it — 30-day is squarely in it. The one feature that distinguished this from the dead family (the 6-month structural length) was dropped for measurability, so the measurable version collapses into the family the hypothesis set out to avoid. - CREDIT: the code is CORRECT and the outlier-mirage fix genuinely works. The monotonic-deque channel correctly excludes the current bar (channel read at the top of calculate_signal from the prior-bar deque state before this bar is rolled in, so the break is a genuine new extreme over the prior channel_bars bars), breakout polarity is correct (LONG on close>channel_hi, SHORT on close<channel_lo — WITH the break), the chandelier-only exit matches the hypothesis (ride the leg), and the vol-scaled min(vol_target, risk_anchor, leverage_cap) sizing delivers the promised outlier-mirage fix (full-sample trade-return kurtosis 0.9, top-3 21.5%, versus the old Donchian's one-day-82%). All zero/finite guards present, BTC size_precision 5, leverage consumed in the cap. This fail is about hypothesis alignment, not implementation — the code would be a valid 30-day-4H channel strategy if the hypothesis were re-scoped to that. - Even setting the mismatch aside, measurability of the faithful version is the same trap the siblings hit: the 6-month daily channel fires ~5x/yr (35 round trips, top-3 38%), below the ~100-trade floor with a foreordained-empty 15-day holdout; the 30-day 4H version reaches ~14x/yr (90 trades) only by abandoning the structural premise. The sandbox is also thin and outlier-flavoured in the recent year (10 trades, 5 entries, Sharpe 0.43 with CI [-1.37, 1.89] straddling zero, skew 2.8, sandbox-year kurtosis 29, long_win_rate 0.0). So the measurable configuration is off-hypothesis and the on-hypothesis configuration is unmeasurable — the analyst/Research Lead must pick which, not QA.

Verification failed (Layer 4 — QA review): - STRUCTURALLY UNMEASURABLE — now that the hypothesis mismatch is fixed, the faithful 6-month strategy cannot be tested under the factory's gates. The developer's own runs prove it: at the hypothesis's 180-day (6-month) channel the 365-day sandbox produces ZERO trades, at 120-day ZERO, and only the 90-day floor (the shipped default) produces any at all -- 2 entries / 4 trades with total_return -1.68%, Sharpe -0.54, PF 0.0, win_rate 0.0, avg_trade_return_pct -3.79%, metrics_reliable=FALSE. Full sample: 90-day = 21 round trips (3.2/yr), 180-day = 16 (2.4/yr); a 15-day holdout expects ~0.13 trades. This is decisively below the ~100-trade measurability floor (L16) and below the ~30-trade line at the on-hypothesis 180-day (16) -- fewer than the sibling tail-regime iteration-1 (27) that I rejected at Layer 4, and with ZERO sandbox trades at the faithful setting. The PF 0.0 / win_rate 0.0 are a 2-losing-trade small-sample artifact (metrics_reliable=FALSE), not an L17 code defect. Passing would route this into optimization, where the objective favours longer channels (per the developer's sweep 180-day +9.42%/PF 2.62 beats 90-day +6.01%/PF 1.93), pushing toward FEWER trades and an empty holdout -> zero-trade auto-abandon, burning the 3-phase budget -- exactly the waste L16 rejects at Layer 4. - CREDIT: the iteration-1 hypothesis-vs-implementation mismatch is RESOLVED. bar_type is now BTCUSDT.BINANCE-1-DAY (matching the hypothesis's 'Daily Bars'), bars_per_day=1 so channel_days maps 1:1 to daily bars, and _param_bounds widened to [90,240] so the hypothesis's ~6-month (180-day) structural level is now INSIDE the search space (it was capped out at 120 before). The code faithfully implements a multi-month daily structural breakout rather than the 4H short-Donchian I flagged. The one honest wrinkle: the shipped DEFAULT is 90-day (not 180) because 120-180 produce zero sandbox trades -- disclosed and reasonable as 'the most testable on-family setting', but it means the default is a 3-month, not 6-month, channel; that is a defaults choice, secondary to the measurability block above. - The code is CORRECT and unchanged in mechanism from the version I verified in iteration 1: the monotonic-deque channel correctly excludes the current bar (channel read at the top of calculate_signal from the prior-bar deque state before this bar is rolled in), breakout polarity is correct (LONG on close>channel_hi, SHORT on close<channel_lo), the chandelier-only exit matches the hypothesis, and the vol-scaled min(vol_target, risk_anchor, leverage_cap) sizing genuinely delivers the low-kurtosis outlier-mirage fix (full-sample kurtosis 1.4-1.7, no single-day mirage). All zero/finite guards present, BTC size_precision 5, leverage consumed in the cap. This fail is purely measurability, not implementation.

Verification failed (Layer 4 — QA review): - STRUCTURALLY UNMEASURABLE and now proven UNFIXABLE — abandon, do not iterate a fourth time. Iteration 3 is byte-for-byte identical to iteration 2 (which I failed on this exact ground) and the developer concurs with abandonment. The new, decisive evidence closes the last escape route: the base class retains only the last 500 primary bars, so min_bars_required() must be <=500; a 6-month channel needs 185 DAILY bars (works) but 1085 4H bars (unsatisfiable -> ZERO bars processed, measured). So the faithful 6-month channel is expressible ONLY on daily bars, and on daily bars it produces essentially no trades: sandbox ZERO trades at 180-day and 120-day, only 2 entries / 4 trades at the 90-day floor (total_return -1.68%, Sharpe -0.54, PF 0.0, win_rate 0.0, metrics_reliable=FALSE); full sample just 16 round trips (2.4/yr) at 180-day and 21 (3.2/yr) at 90-day, with a ~0.13-trade 15-day holdout. Every count is far below the ~100 floor and below the ~30 line, and the optimizer's objective pushes channel_days UP (180-day PF 2.62 beats 90-day PF 1.93) toward FEWER trades and a guaranteed-empty holdout. The PF 0.0 / win_rate 0.0 are a 2-losing-trade artifact, not an L17 defect. Per L16/L26, reject at Layer 4 rather than burn optimization budget on a run that will auto-abandon on an empty holdout. - The code is CORRECT and unchanged in mechanism from the version I verified in iterations 1-2, and the hypothesis mismatch from iteration 1 remains fixed (daily bars, bars_per_day=1, 180-day reachable in _param_bounds [90,240]). The monotonic-deque channel correctly excludes the current bar, breakout polarity is correct, the chandelier-only exit matches the hypothesis, and the vol-scaled sizing genuinely delivers the low-kurtosis outlier-mirage fix (full-sample kurtosis 1.4-1.7). This fail is purely measurability; there is no implementation defect to repair. The developer has now exhaustively and honestly mapped the frontier across three iterations.