BtcCmVolRegimeGatedDonchianBreakoutLS

BTC COIN-M Inverse Perp Volatility-Regime-Gated Donchian Breakout, Long-Short (Single-Instrument BTCUSD_PERP.BINANCE_CM, 4h Bars, Trade Channel Breaks ONLY in Expanding-Volatility Regime — Flat in Chop, ATR-Stopped, Pure OHLCV, 3-Parameter)

A LONG-SHORT, single-instrument directional strategy on the Binance COIN-M inverse perpetual BTCUSD_PERP.BINANCE_CM (BTC-margined, ~0.06% round-trip taker), 4-hour bars, pure OHLCV. It trades classic Donchian channel breakouts BUT gates every entry behind a volatility-regime filter: signals are taken only when the market is in an EXPANDING-volatility state, and the strategy sits flat during low-volatility chop. This is a re-attempt of a previously abandoned hypothesis whose abandon was an ENGINE BUG — a phantom COIN-M liquidation from margining a USDT-only account with qty×price maintenance margin — fixed 2026-07-24 (commit 59c8a73) to coin-funded, face-based margining; the abandoned code re-runs cleanly at 0.5× sizing, 254 trades, no liquidation, and QA credited the code. Mechanism per 4h bar: (1) compute an N-bar Donchian high/low channel (default N=30 ≈ 5 days); (2) compute a volatility-regime flag = ATR(14) is above its own rolling median over the last M bars (default M=60 ≈ 10 days) AND rising vs 6 bars ago; (3) go LONG on a close above the upper channel and SHORT on a close below the lower channel, ONLY if the vol-regime flag is TRUE — otherwise ignore all breakout signals and hold no position. It deliberately fills three under-represented buckets at once — BINANCE_CM venue (3.2% vs ≥5%), long_short direction (14.7% in a long-only-dominated book), and diversification away from the 71.5% USD-M concentration — on a low-fee venue where the fee math is forgiving. It is a single-leg CM strategy on BTC OHLCV only (L20-blessed: no multi-asset CM funding history needed), and it addresses the exact failure mode that killed the SOL 15m opening-range breakout (chop whipsaws, 28% win rate) by refusing to trade in the regime where breakouts fail-and-reverse. Against this session's trend-sleeve deaths (single-regime concentration on daily macro trend), its defense is the 4h frequency (254 trades → return stream not dominated by a handful of monster daily bars) plus a hard ATR stop and time cap so no single position rides a bull spike disproportionately.

Implements the hypothesis exactly: (1) N-bar Donchian high/low channel excluding the current bar; (2) vol-regime flag = ATR(14) above its rolling median over M bars AND rising vs 6 bars ago; (3) long on upper-channel close break, short on lower-channel break, gated on the regime flag so the strategy sits flat in chop — the direct answer to the breakout-whipsaw failure mode. Three tunable knobs (donchian_lookback, vol_window, atr_stop_mult) with clamps mirrored in _param_bounds so the optimizer never searches values the code rewrites. calculate_signal returns a CONTINUOUS channel-position value (close - mid)/half_width in natural units every bar (|signal|>1 ⇔ a real channel break), never a gated constant; the regime test lives in should_enter. Venue is BINANCE_CM per the hypothesis (long-short + BTC-margined convexity, ~0.09% RT taker), 4h bars matched to the mid/long-term breakout horizon, and confirmed present in the catalog (BTCUSD_PERP.BINANCE_CM-4-HOUR-LAST-EXTERNAL). Inverse sizing uses contracts_for_notional (whole $100-face contracts) rather than the linear equity/price formula — at ~$55k typical notional that is ~550 contracts, far above min notional, so zero-trade/min-notional failure is not a risk. leverage=2.0 is genuinely consumed in position_size via the gross-notional cap (equity × 0.75 × leverage = 1.5× equity ceiling), while per-trade risk stays capped at 2% of equity against the ATR stop, and the time cap plus trail prevent any one position from riding a single trend leg disproportionately. All state is incremental (deques + Wilder ATR + a ≤60-element median sort), so per-bar cost is O(1)-ish and the 300s smoke budget is safe.

unrealized_mtm_tail_driven_closet_beta_donchian_clone: BTC COIN-M vol-regime-gated Donchian breakout is a single-instrument pure-OHLCV directional-breakout clone (0/213 survival, L41/L63) whose strong-looking headline does not survive scrutiny. The +299% total_return is ~2/3 unrealized (end_unrealized_pct 201.3 vs realized ~98%) — the result and much of the daily-return Sharpe hinge on an un-exited open position marked to market at backtest end, not a realized edge. It is tail-driven (kurtosis 7.56): a single day 2023-01-14 (+21.1%) IS essentially the entire 2023 (+31.7%). It is closet-long BTC (benchmark_correlation 0.86, beta 0.45, information_ratio -0.487) and has decayed for two years (2025 -0.75%, 2026 -1.84%, rolling Sharpe near/below zero from late 2025). Sharpe 0.792 has a positive CI low (0.139) but that is the pre-deflation, best-of-selection number; over a 225-trial / 3-param sweep the deflated Sharpe would strip it below significance (expected-max luck bar exceeds 0.139), exactly as prior Donchian/trend clones died. avg_trade_return_pct 1.05% clears the CM fee floor so this is not a fee death and the prior phantom-liquidation was a legitimately-fixed engine bug — but the signal itself is a tail-and-MTM-driven closet-beta trend clone. Per L41/L63 the class prior overrides the attractive headline; a directional OHLCV book's base Sharpe is its ceiling and this one would die at DSR/holdout. Not worth the 2-hour optimization run.

Long-short Donchian channel breakout on the BTC COIN-M inverse perpetual (BTCUSD_PERP.BINANCE_CM, 4-hour bars, pure OHLCV). Each bar it computes an N-bar Donchian channel from the PRIOR bars (default 30 ≈ 5 days) and a volatility-regime flag: Wilder ATR(14) above its own rolling median over the last M bars (default 60 ≈ 10 days) AND rising versus 6 bars ago. It goes LONG on a close above the upper channel and SHORT on a close below the lower channel, but ONLY while the expanding-volatility regime holds — in low-volatility chop every breakout is ignored and the book stays flat. Exits are an ATR chandelier trail (3×ATR from the favourable extreme since entry), a hard 60-bar (≈10-day) time cap, or a full opposite channel break. Sizing is volatility-targeted: 2% of equity risked against the ATR stop distance, capped at 0.75×equity×leverage of gross notional, converted to WHOLE inverse contracts via contracts_for_notional ($100 face per BTCUSD_PERP contract, BTC-margined/settled).

At backtest review, evaluate realized-only performance, confirm Sharpe survives without the largest winner, and lean on the holdout where a terminal open position cannot inflate the result.

Headline inflated by a large open position: end_unrealized_pct is +195% against total_return +299%, so ~2/3 of the reported return is unrealized mark-to-market on a still-open position, not realized PnL (largest_win $29,116 vs avg_win $3,569 is an ~8x outlier). The per-trade edge is sound because avg_trade_return_pct 1.15% is measured over the 158 CLOSED trades, but the headline total_return and Sharpe (0.79) are lifted by the open markup and must be discounted. Same inflation pattern that has flattered optimized headlines before.

Weight DSR and OOS Sharpe heavily; check shorts contribute positively and the edge isn't purely the long leg riding 2024-2025 BTC.

Directional single-instrument OHLCV Donchian-breakout family (historically fragile). The vol-regime gate is a genuine differentiator and the sandbox is significant (Sharpe CI low 0.14 > 0, positive alpha), but information_ratio is -0.49 vs buy_hold and beta 0.45 to BTC — some return is long-BTC exposure in a bull sample. Generalization must be proven at walk-forward/holdout, not assumed.

Carry 2x leverage and DD forward to risk review; consider sensitivity to leverage=1.0.

Leverage 2.0 (within the BINANCE_CM cap, legitimately consumed) amplifies both the 39.1% max drawdown (CI to 59%) and the open-position markup. Correct and intentional, but risk review should confirm the 2x drawdown profile on an inverse contract is acceptable.

Best risk-adjusted profile of the family: Sharpe 0.792 with CI low 0.139 (positive), PSR 0.99, PF 1.70, Sortino 1.20

max_drawdown 39.1% stays under the L19 50% line; hard ATR stop + time cap bound single-position risk

avg_trade_return_pct 1.05% clears the CM fee floor; balanced long/short book (84L/74S), 158 trades, clean inverse sizing; prior liquidation was a since-fixed engine bug

299% headline is almost entirely unrealized: end_unrealized_pct 201.3 — realized is only ~98%, so the result (and much of the Sharpe) hinges on an un-exited open position marked to market at backtest end

Single-day tail dependence: 2023-01-14 alone +21.1% IS essentially the entire 2023 (+31.7%); kurtosis 7.56

Closet-long BTC: benchmark_correlation 0.86, beta 0.45, information_ratio -0.487 — mostly captured beta

Decayed: 2025 -0.75%, 2026 -1.84%; rolling Sharpe near/below zero from late 2025

Single-instrument pure-OHLCV Donchian-breakout clone family (0/213, L41/L63); Sharpe 0.79 with CI-low 0.139 would deflate below significance under the 225-trial best-of-N luck bar

BtcCmVolRegimeGatedDonchianBreakoutLS re-submitted a vol-regime-gated Donchian breakout on the BTC COIN-M perp after its earlier abandon was traced to a since-fixed phantom-liquidation engine bug, and this run cleared the code and the Sharpe/drawdown/PF floors with the best profile of its family (+299%, Sharpe 0.792, PF 1.70, 39.1% max DD). On inspection the shine was fragile: about two-thirds of the return was unrealized mark-to-market on an un-exited position, a single +21% day carried all of 2023, and with 0.86 BTC correlation and a negative information ratio it was mostly captured beta that had decayed to negative through 2025-2026. Its Sharpe CI-low of 0.139, while positive, sits below the deflation luck bar a 225-trial sweep would impose. As another 0/213 OHLCV breakout clone, it was abandoned at backtest review before any optimization or analysis.

Discount a headline return for unrealized mark-to-market on open positions and for single-day tail dependence before trusting it; a positive pre-deflation Sharpe CI-low (0.139) is still below the best-of-N luck bar for a 3-param sweep, and a class prior (0/213 OHLCV breakout clones, high beta/negative information ratio) overrides an attractive-looking backtest.

It was abandoned at the pre-optimization BACKTEST_REVIEW gate: despite passing the Sharpe, drawdown, and PF floors and having the engine bug legitimately fixed, the strong headline did not survive scrutiny — it was ~2/3 unrealized MTM on an un-exited open position, tail-driven, and captured beta rather than alpha, with a two-year decay. As a single-instrument OHLCV Donchian-breakout clone (0/213 survival), its Sharpe CI-low of 0.139 would deflate below significance under the 225-trial best-of-N luck bar and die at DSR/holdout, so it was judged not worth the optimization run. Optimization, analysis, and risk review were never reached.

A long-short, single-instrument directional strategy on the BTC COIN-M inverse perp (BTCUSD_PERP.BINANCE_CM, 4H, pure OHLCV) that traded Donchian channel breakouts only in an expanding-volatility regime (ATR above its rolling median and rising), sat flat during chop, and exited via an ATR chandelier trail, a time cap, or an opposite channel break — a re-attempt of a prior submission whose abandon was a since-fixed phantom-liquidation engine bug.

Over 5.8 years it made 158 trades (84 long / 74 short) with total return +299%, profit factor 1.70, Sharpe 0.792 (CI low 0.139, positive), Sortino 1.20, PSR 0.99, avg_trade_return_pct 1.05% (clearing the CM fee floor), and a 39.1% max drawdown — the best risk-adjusted profile of its family. But ~2/3 of the headline was unrealized (end_unrealized_pct 201.3 vs realized ~98%), the PnL was tail-driven (kurtosis 7.56, a single +21.1% day essentially being all of 2023), and it was effectively closet-long BTC (correlation 0.86, beta 0.45, information ratio -0.487), decaying to negative in 2025 (-0.75%) and 2026 (-1.84%).