EthPriceOpenInterestQuadrantLS

ETH Perp Price–Open-Interest Quadrant Positioning, Long-Short (Single-Instrument ETHUSDT.BINANCE USD-M — Classify Each Breakout by Whether OPEN INTEREST Confirms NEW Money vs Short-Covering/Deleveraging, Trade Only the OI-Confirmed Continuations, Daily-Ish Bars, Low-Frequency, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, SINGLE-VENUE directional strategy on the ETHUSDT.BINANCE USD-M perpetual that adds a genuinely NEW information source — the change in OPEN INTEREST — to a price signal, exactly the kind of augmentation L38 permits (a new information source, not a new instrument or another momentum clone). The core is the price×OI 'quadrant' model of positioning flow: a price move MEANS different things depending on whether open interest is expanding or contracting. Price breaks to a new high WITH strongly rising OI = fresh leveraged long money entering → a genuine, fuel-backed uptrend → GO LONG. Price breaks to a new low WITH rising OI = new short conviction / cascade building → GO SHORT. But price at new highs while OI FALLS = a short-covering squeeze (no new buyers) — unsustainable, so DO NOT go long; and price at new lows while OI falls = long capitulation/deleveraging (forced exit, not new sellers) — so DO NOT go short. The strategy only takes the two OI-CONFIRMED quadrants and stays flat in the unconfirmed ones. This is deliberately NOT the OI-fuel BREAKOUT that died on fee_edge: it is LOW-FREQUENCY on daily-ish bars holding multi-day trends, so each trade targets multi-percent moves and the 0.10% USD-M round trip is negligible. It is NOT cross-venue and NOT multi-leg (the complexity that killed my last several ideas in Layers 2–4), NOT options (catalog-load caps), NOT funding carry (no edge), and NOT knife-catching MR (which just liquidated). ONE instrument, ONE venue, ONE cheap supplementary series (OI), computed incrementally. Fills the under-represented LONG-SHORT direction bucket and uses a positioning MECHANISM distinct from my pending funding-fade and calendar ideas. 3 parameters, to resist the overfit deaths that dominate the log.

Iteration 2 addresses the ONE issue QA raised — a structurally unmeasurable trade count — with the smallest possible change: the mechanism, signal, gates, exits and sizing code are byte-identical to iteration 1; only the bar clock and the three edge-parameter defaults moved (1-DAY -> 12-HOUR, breakout_lookback 20 -> 12 bars = 6 days, oi_lookback_bars 5 -> 6 bars = 3 days, oi_change_threshold 0.02 -> 0.015). Layers 1-3 that already passed are untouched (no new imports, same continuous signal, re-checked green locally on static analysis and all six synthetic scenarios). I re-ran the full compound gate over the real ETHUSDT 12-hour bars and the real merged OI series, simulating actual entries/exits (entry only when flat, mid-channel or 8% stop exit, 0.10% round-trip fees): 103 completed trades over the OI-covered history versus 37 for the daily version, with ~22 falling inside a 365-day sandbox window (vs the 9 QA saw). That puts each of the 3 walk-forward windows around 34 trades instead of ~12, which is what makes an optimized Sharpe interpretable. Critically, the higher cadence does NOT convert this into a fee-fragile strategy: average net-of-fee return per trade is ~1.9% of notional (~13x the 0.15% viability floor) and average hold is ~12 bars ≈ 6 days, so it remains the low-frequency, multi-percent-target trend trade the hypothesis specifies rather than the intrabar OI-fuel breakout that died on fee_edge. The chosen point is the middle of a broad plateau, not a peak: across breakout_lookback 10-16, oi_lookback 6-8 and threshold 0.010-0.025 the count stays 81-119 trades with per-trade returns 1.4-2.1% and comparable risk-adjusted behaviour, so sensitivity analysis and the holdout should not fall off a cliff. On QA's second point (mechanism-class risk): I agree it is an OHLCV+OI breakout family and the analyst may still judge the class unproven — that call is theirs; the code implements the stated quadrant model exactly, with the OI filter rejecting roughly 45% of raw breakouts as squeeze/deleveraging.

Single-perp OI-gated Donchian breakout — the L73 zero-survivor class this session (ETH OI-confirmed range breakout optimize→abandon, SOL OI-confirmed trend overfit) — not worth 2 hours of optimization. PF is 1.37, below the ~1.7 the class prior requires to override, and it is not significant: Sharpe 0.695 with bootstrap CI [-0.088, 1.405] straddling zero and PSR 0.871 (<0.95), which the ~225-trial deflation only lowers. The 146% return is regime-concentrated and outlier-driven — 2022 (+47%) and 2024 (+65%) carry it while 2021 (-4%), 2023 (-4.7%), 2025 (+2.7%) and 2026 (+3.2%) are flat-to-negative and the stressed-vol regime is dead (sharpe 0.01, -4.3%) — so the last-20% holdout (2025-2026) is weak and the OI-confirmation does not help in the regimes that matter (kurtosis 8.1, win_rate 37.3%, 11 max consecutive losses). It is not fee-dead (avg_trade_return_pct 1.28% clears the USD-M round trip), but the failure mode is absence of a significant, robust edge, and an OI gate on a breakout does not add robustness the deflation can't strip (L30/L63 breakout graveyard). No parameter tuning conjures significance from a measured-but-null result in a class with no survivors. Failure pattern: overfit/no_edge single-perp OI-gated directional breakout (L73/L63).

Single-instrument, single-venue LONG-SHORT strategy on the ETHUSDT.BINANCE USD-M perpetual, now on 12-hour bars. Every bar it computes the close's position inside the prior 12-bar (6-day) Donchian channel ((close - mid)/half_width: >= +1 = new 6-day high, <= -1 = new 6-day low) and, independently, the fractional change in OPEN INTEREST (contracts) over the prior 6 bars (3 days) from the merged Binance OI feeds. It then trades only the two OI-CONFIRMED quadrants of the price x OI matrix: new high with OI expanding >= 1.5% (fresh leveraged long money) -> LONG; new low with OI expanding >= 1.5% (new short conviction) -> SHORT. New highs or new lows on CONTRACTING open interest are short-covering squeezes and long capitulation respectively — no new money, so the strategy stays flat. Exit is the hard 8% stop or the close returning to mid channel, giving ~6-day average holds. Sizing is 75% of equity in notional at leverage 1.0. Open interest is parsed once at construction into a sorted numpy series and read per bar with two O(log n) searchsorted lookups — no per-bar history rescan. With no (or stale) OI data the strategy stays flat rather than degrading into a price-only breakout.

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.

OI confirmation measured over the same window as the breakout may be contemporaneous with (not orthogonal to) the price move — the non-orthogonality that made the BTC OI-momentum sibling net-zero. Donchian breakout alone is a dead OHLCV class (L81), so the thesis rests on the OI gate adding separable info. Code faithfully implements the quadrant model; an edge question for the analyst. A/B OI-gated vs ungated Donchian on the covered window; promote only if the gate materially improves edge/Sharpe.

Config edge/risk params are non-default optimizer-tuned values (breakout_lookback 6 not the described 12, threshold 0.0204, risk_frac 0.6127, stop_loss_pct 4.71), so the effective fitting surface exceeds the advertised 3 and the description is stale vs config. Cosmetic for correctness; relevant to overfit. Freeze risk_frac/stop_loss_pct so only the 3 signal params are optimized.

avg_trade_return_pct 1.28% is well above the ~0.10% USD-M round trip — genuinely low-frequency, not fee-fragile (the fee_edge death mode is avoided)

150 trades over the OI-covered history is a measurable sample, and the OI gate is a legitimately new information source

Max drawdown 21.3% is only modestly above the promotion floor; no liquidation

L73 zero-survivor class: single-perp OI/funding/premium-gated directional trade with PF 1.37 (<1.7) — the class prior says abandon at review; the OI gate does not add robustness the deflation strips

Not significant: Sharpe 0.695 with CI [-0.088, 1.405] straddling zero and PSR 0.871 (<0.95)

Regime-concentrated: 2022 (+47%) and 2024 (+65%) carry the result; 2023 negative and 2025/2026 nearly flat — the last-20% holdout window is weak; stressed regime sharpe 0.01

Outlier-driven (kurtosis 8.1, skew 1.18); win_rate 37.3% with 11 max consecutive losses

Verification failed (Layer 4 — QA review): - STRUCTURALLY UNMEASURABLE TRADE COUNT. This is a low-frequency daily-bar strategy with a selective compound gate (20-day Donchian breakout AND >=2% OI expansion). The sandbox produced only 9 trades over 361 days, with Sharpe 0.30 and a 95% CI of [-1.47, 1.87] — statistically indistinguishable from zero. The developer's own dry-run reports 82 CONFIRMED SIGNAL BARS over the entire OI-covered history, but signal bars badly overcount trades: once in a position, subsequent signal bars generate no new trade (the sandbox's 18 signal bars in the last year collapsed to just 9 trades). Extrapolating, the FULL multi-year backtest will realize on the order of ~40-60 trades total. That is far below the ~100-trade floor needed to distinguish edge from noise, and it is fatal for the downstream pipeline specifically: the 3-window walk-forward optimization would see ~15 trades per window and the 15-day holdout a mere handful — so any optimized Sharpe/PF is curve-fit to a dozen observations. The sandbox's headline numbers (profit_factor 1.26, total_return +6.8%) are therefore noise from 9 trades, not evidence of edge (5 long trades at 40% win rate, 4 short at 75% — a coin-flip on tiny n). This is the recurring pattern where a plausibly-coded low-frequency breakout passes QA and then dies unmeasurable/single-regime at backtest_review after burning optimization iterations; reject at QA on trade-count grounds instead. - Mechanism-class risk. Despite the hypothesis explicitly distancing itself from 'the OI-fuel BREAKOUT that died on fee_edge', the implementation IS a Donchian breakout gated by an OI-expansion filter — structurally the OHLCV+OI breakout family that has repeatedly failed (e.g. the OI-fuel breakout sibling abandoned with PF ~1.05, single-regime, where the OI filter did not remove the false breakouts). The 'low-frequency daily / multi-percent target' reframing changes the trade cadence and the fee math but not the underlying claim that OI-confirmed breakouts continue net of noise, which has not survived. This is an edge-class caution for the analyst, not a code defect: the code correctly implements the stated quadrant model.