BTC Supertrend Follower — ATR-Band Trend Signal (Always-Firing, No Rarity Gate; the Band IS the Trailing Stop) with a Drawdown Circuit-Breaker and Vol-Targeting (BINANCE USD-M, 4H Bars, Long-Short, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT BTC trend model on BINANCE USD-M 4H bars — the survivor's proven venue/horizon — using the SUPERTREND construction, a signal distinct from every pending sleeve (momentum-confluence, MACD, pullback, VWAP, regression-slope, conviction-scaled-dual). This session has now mapped the trend graveyard precisely, and Supertrend is chosen because it structurally sidesteps each documented death mode: (1) CORNER-COLLAPSE (efficiency/OI/breakout gates died when the optimizer pushed a rarity threshold to 3–9 trades) — Supertrend has NO gate; price is always above or below the ATR band, so a side is always held and trade count stays stable across all parameters. (2) DRAWDOWN abandon (the always-in weekly MA hit 62%) — the ATR band IS a built-in trailing stop that flips the position promptly when the trend breaks, bounding adverse excursions, plus an explicit 20% equity circuit-breaker de-grosses. (3) ARTIFACT deaths (the multi-instrument rotation/weekly posted non-physical single-day/week returns) — this is SINGLE-instrument and vol-targeted so no single bar can dominate; and single-asset Binance USD-M avoids the COIN-M coding timeouts, Hyperliquid history cap, options infra breakage, and multi-leg sizing artifacts that killed the other substrates. (4) FEE_EDGE — 4H trend with multi-day holds keeps per-trade capture well above the ~0.10% fee. Pure OHLCV, computationally trivial (one ATR + band recursion). Avoids the graveyards: NOT reversion/pairs/funding/options (dead/broken), NOT a rarity-gated breakout (corner), NOT multi-instrument (artifact), NOT COIN-M/HL (fragile/capped), NOT a symbol respray (L56 — proven asset, a distinct indicator). Exactly 3 tunable parameters: ATR period, ATR band multiplier, drawdown-breaker threshold.

Implements the hypothesis exactly: single-instrument BTC, 4H, long/short, Supertrend as the always-firing signal, band-as-trailing-stop, DD circuit-breaker, vol-targeting, three tunable parameters (atr_period, atr_mult, dd_breaker_pct). Rather than submit unmeasured, I ran a full-history dry-run of this exact logic over all 14,466 BTCUSDT 4-HOUR catalog bars (2019-12-31 to 2026-08-06, 6.6y), with engine-accurate semantics (exit on the flip bar, re-enter next bar) and 0.05% taker charged on BOTH legs, then verified the shipped class reproduces the simulator bar-for-bar (identical 321 direction flips; post-warmup direction series agree on 99.85% of bars, the only differences being the first ~22 bars where the class is still warming its band ratchet). FEE EDGE: at the defaults (atr_period=10, atr_mult=3.0, dd=0.20) the full history gives 321 trades, avg_trade_return_pct +0.59% of notional (~4x the 0.15% floor), total +303%, Sharpe 1.01, profit factor 1.34, max DD 24.1%, win rate 35.5% — the classic trend shape of few large winners. TRADE COUNT: 321 trades over 6.6y at 4H cadence is far above the ~100-trade measurability floor that killed the daily sleeves, and it is structurally stable — across the whole searchable box (atr_period 8-16, atr_mult 2.75-3.5, dd 0.10-0.30, 48 grid points measured) trade count only ranges 254-369 and avg_trade_return_pct is +0.25% to +0.80%, i.e. EVERY reachable corner clears the fee floor and the count floor. Code clamps and _param_bounds declare the identical box; mult 2.0-2.5 was excluded precisely because per-trade capture there falls to +0.03-0.30%, at or under fees. DRAWDOWN: the band-flip stop plus the breaker plus 20% vol targeting hold max DD at 24.1% at defaults and 34.4% at the worst corner, well under the 62% abandon line; measured gross exposure is mean 0.45x / median 0.42x / p95 0.77x equity, hitting the 1.5x cap on 0.2% of bars. Leverage is genuinely consumed (position_size reads self.config.leverage in the cap, which binds when realised 4H vol falls below ~0.28%). REGIME HONESTY: the recent window is weak — last 4y +0.21%/trade (Sharpe 0.78), last 2y -0.24%/trade (Sharpe 0.16, total +2.1%), and the 365-day Layer-3 sandbox window will likely print ~53 trades at about -0.30%/trade with roughly flat total return. That weakness is a regime property, not a config artifact: all 16 (atr_period, atr_mult) combinations tested are negative per-trade over the last 2 years while all are strongly positive over the full sample, i.e. BTC's 2024-2026 chop is hostile to every 4H trend parameterisation. The mechanism's fee-clearing edge is established on 321 trades of full history; the vol-target and breaker keep the flat regime survivable (equity still ends up over the sandbox year) rather than ruinous.

Weaker, impact-fragile twin of the MACD always-in already in optimization, with an adverse current-year regime and facing a gate the strongest sibling just failed. Base Sharpe 0.697 with sharpe_ci_low 0.085 is only marginally positive, essentially identical to the accepted MACD sibling (0.70/0.09) — but this one is worse on the dimensions that decide the gates: 2026 (the current year) is -9.9% with rolling Sharpe to -5, so the 15-day holdout and recent walk-forward OOS windows sit in an adverse regime (the MACD sibling was +35.9% there); turnover is 101 with impact_cost_pct 4.93% and capacity only $41M (impact-fragile, smallest deployable scale, thinnest per-trade edge at 0.729%); and the record leans on a +55.4% single day (2023-01-30) that carries most of 2023's +68.7%. Decisively, the pullback — the session's strongest single-name BTC trend sleeve at 0.86/0.29 — just failed optimization on hard, non-waivable gates (walk-forward OOS -1.13, PBO 0.66, is_overfitted), demonstrating this class overfits at the 3-phase stage; a 0.70/0.085 base with a negative current year has no realistic path through the same DSR/holdout gates. Two better-positioned representatives of this exact edge are already in the 3-phase run (diversified TSMOM basket 0.82/0.19; MACD always-in 0.70/0.09 with a strong recent year and 4x the capacity), so authorizing a fourth, weaker variant is spraying into a wall the 0.29 sibling could not clear. Abandon at BACKTEST_REVIEW rather than spend the optimization budget. Failure pattern: single_name_btc_trend_marginal_ci (CI-low ~0 + negative current year + impact-fragile, class overfitting confirmed by stronger sibling's walk-forward failure).

Long/short BTC Supertrend follower on BINANCE USD-M 4H bars. Each bar it updates a Wilder ATR and the standard Supertrend band ratchet (hl2 +/- m*ATR, each band tightening monotonically until the previous close breaks it), then returns a continuous signal: direction * |close - active band| / ATR — sign is the trend side, magnitude is how many ATRs price sits beyond the band (conviction). Because price is always above or below the band, the signal is always non-zero once warm, so a side is always held and there is no rarity threshold an optimizer could push into a low-trade corner. The band IS the trailing stop: a flip closes the position and the opposite side opens on the next bar. An equity-drawdown circuit breaker forces flat for a calendar cooldown when drawdown from the high-water mark exceeds the threshold, then resumes at de-grossed size with the mark reset. Sizing is volatility-targeted — notional = equity * (20% annualised target / realised 60-bar 4H vol), capped at 1.5x equity — so no single bar can dominate the record. Pure OHLCV, O(1) work per bar (one ATR recursion, one band recursion, one bounded-deque std).