EthOpenInterestConfirmedBreakoutLS

ETH Open-Interest-Confirmed Volatility Breakout — Trade Donchian Range Breakouts ONLY When Open Interest Expands (Leverage-Backed, Not a Stop-Run Fakeout), Ride With a Chandelier Trailing Stop (BINANCE USD-M, 4H Bars, Long-Short, Asymmetric Let-Winners-Run, 3-Parameter)

A LONG-SHORT, single-instrument, ASYMMETRIC-PAYOFF breakout strategy on ETHUSDT.BINANCE (USD-M perp). It is a direct response to the single clearest pattern in this session's results: every SYMMETRIC mean-reversion / contrarian-fade sleeve (ETH funding-contrarian, LTC-BCH pairs, SOL-ETH pairs) converged to breakeven BELOW the fee floor, while the ONE survivor is a directional trend construction that lets winners run with a trailing stop (Sharpe 1.99). The structural lesson: fixed-target reversion cannot clear fees here, but an asymmetric right-tail can. This sleeve harvests that tail differently from the sprayed 4H+1D momentum-confluence template (L56) — it is a VOLATILITY BREAKOUT with an OPEN-INTEREST CONFIRMATION filter, which is a genuinely different mechanism. The novel edge: a Donchian range breakout accompanied by EXPANDING open interest is a leverage-backed move (new money entering in the breakout direction → follow-through), whereas a breakout on flat/falling OI is typically a stop-run / short-cover fakeout that fails. Requiring OI expansion filters the fakeouts that make naive breakouts unprofitable. It avoids the graveyards: NOT symmetric mean-reversion/fade (just died fee_edge), NOT the dual-TF momentum-sign clone (L56 — this is range-breakout + OI, not momentum confluence), NOT funding carry/differential (L57), NOT pairs/cross-sectional (dead), NOT options (infra-broken), NOT a liquidation feed (L58 — uses the deep OI metrics series, not per-liquidation prints). Computationally TRIVIAL: a rolling max/min (Donchian), an ATR, and an OI percent-change — a few scalars per bar, no full-history rescans, no sorted()/min() over long lists in the hot path. Fills the under-represented long-short bucket and uses supplementary data. Exactly 3 tunable parameters: Donchian lookback, OI-expansion threshold, and chandelier ATR multiple.

Implements the hypothesis exactly: Donchian range breakout gated on expanding open interest (leverage-backed follow-through vs stop-run fakeout), long-short on ETH 4H, asymmetric payoff via a chandelier trailing stop and no fixed target, with the three named tunables (donchian_lookback, oi_expansion_z, chandelier_atr); every clamp is mirrored in _param_bounds. Two deviations worth naming: (1) the OI-expansion threshold is a rolling Z-SCORE of the OI percent change rather than a fixed percentage, because fixed absolute supplementary thresholds have repeatedly produced regime-dependent trade counts (past lesson L5) - a z-gate keeps firing rate stable across OI regimes and it also tested better; (2) the deep OI history comes from the 'metrics' feed, since the dedicated open_interest feed only starts 2026-03 for ETHUSDT (both are merged). The exact submitted code was dry-run bar-by-bar on real ETH 4H bars plus real OI, 2021-12 to 2026-08: 130 trades (~28/yr, 30 in the most recent 365 days, so the Layer-3 >=1-trade gate is safe), mean +1.08% NET of a 0.10% round-trip taker charge - about 7x the 0.15% fee floor - 37.7% win rate with a +56.8% best trade against a -13.7% worst, i.e. the intended right-tail asymmetry; both legs work (long +0.84%/trade over 78 trades, short +1.43% over 52). The OI filter is doing real work rather than decorating: at the same lookback and stop, removing it drops mean per-trade from +1.43% to +0.60% while adding fakeouts (282 vs 130 trades), and the filter improves the per-trade mean at every lookback tested (15/20/30/40). Caveats for the analyst: yearly means are +0.73 (2022), -0.14 (2023), +3.01 (2024), +1.54 (2025), -0.21 (2026 partial) - the edge is real but lumpy and concentrated in trending years, as any breakout sleeve is; and chandelier_atr has a genuine cliff below ~2.0 (at 2.0 the same config nets 0.0%/trade because the trail is inside normal 4H noise), which is why the declared lower bound is 1.5 and the default sits at 3.0 on the plateau. Venue is BINANCE USD-M futures (MARGIN) because the strategy shorts; leverage stays 1.0 and is deliberately not referenced in sizing.

Statistically empty single-perp OI-gated breakout — not worth 2 hours of optimization. The OI-expansion gate (oi_expansion_z ~2.49, near the 2.5 max) almost never fires, producing only 9 trades over 6.6 years with exposure_pct 0.7% — far below the ~100-trade measurability floor (L10; single-digit trades are unmeasured per L27). A 3-window walk-forward plus 20% holdout would leave 1-2 trades per fold and near-certainly zero trades in the holdout, tripping the zero-trades hard gate; a win rate on 9 trades is noise. There is no edge: Sharpe 0.10 with bootstrap CI [-0.76, 0.73] straddling zero, PSR 0.52, information_ratio -0.93, total_return 3.3% over 6.6 years (CAGR 0.79%). The recent regime is negative (2024 -3.6%, 2025 -4.2%, 2026 -2.3%); the entire positive contribution is 2022 (+9.6%), so the holdout window would be negative even if populated. This is the L73 single-perp OI-gated directional family (edge-dead this session) crossed with the L30/L63 pure-breakout graveyard. Risk is contained (max_drawdown 9.7%, no liquidation), but the low trade count is structural: loosening oi_expansion_z to add trades removes the OI-confirmation filter that IS the hypothesis, reverting it to the naive price-only breakout the hypothesis itself says loses money — so there is nothing to tune toward. Failure pattern: overfit/no_edge single-perp OI-gated breakout, unmeasurable sample (L10/L73).

Long-short volatility-breakout strategy on ETHUSDT.BINANCE USD-M perp, 4-HOUR bars, with an open-interest confirmation filter and an asymmetric let-winners-run exit. Each bar it computes a continuous range-position signal (close - mid) / (0.5 * (upper - lower)) from the Donchian channel of the PRIOR donchian_lookback bars (current bar excluded, no look-ahead): ~0 mid-range, +-1 exactly at the range edge, beyond +-1 on a breakout. Entry requires BOTH a breakout (signal >= break_thresh for long, <= -break_thresh for short) AND open-interest expansion: the oi_window-bar percent change in OI, z-scored against a 180-bar rolling baseline, must be >= oi_expansion_z. OI is read from the supplementary 'metrics' feed (Binance Vision 5-minute metrics, open_interest field: 473k rows, continuous 2021-12 to now, 96% bar coverage) merged with the dedicated 'open_interest' feed; both are iterated as lists of row-dicts and merged once into a sorted numpy index consumed by one binary search per bar. There is NO take-profit: the sole exit is a CHANDELIER trailing stop (highest high since entry minus chandelier_atr x ATR for longs, mirrored for shorts), so losses are cut near chandelier_atr ATR while winners run. Sizing is capital-relative: risk_frac (1%) of equity divided by the chandelier stop distance, capped at 1x equity notional (measured 0.19x average, 0.59x max); leverage 1.0. If OI is missing or its baseline is not warm the strategy does not enter (never an unconfirmed breakout), while the price-only chandelier exit keeps managing any open position. All state is incremental (Wilder ATR, bounded deques) - a full 10,260-bar run takes ~1.0s.

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.

Computationally clean, OI loaded, no liquidation; contained drawdown 9.7%, avg_trade_return_pct 0.99% above fees

Asymmetric let-winners-run design is the right structural instinct

Only 9 trades over 6.6 years (exposure 0.7%) — far below any measurability floor (L10/L27); holdout would near-certainly hit the zero-trades gate

Sharpe 0.10 with CI [-0.76, 0.73] straddling zero, information_ratio -0.93, total_return 3.3% (CAGR 0.79%) — no edge, negligible

Recent regime negative: 2024 -3.6%, 2025 -4.2%, 2026 -2.3%; the entire positive is 2022 (+9.6%)

OI gate at z~2.49 is the reason for the tiny sample — loosening it removes the confirmation filter that is the hypothesis, reverting to the naive breakout that loses (L73/L30)