EthOpenInterestConfirmedBreakout

Open-Interest-Confirmed Breakout, Single-Venue Long-Short (Single-Instrument ETHUSDT.BINANCE USD-M — Take a Donchian-Channel Breakout ONLY When It Is Backed by a Simultaneous EXPANSION in Open Interest (Fresh Positioning Capital Entering), FLAT/Skip the Breakouts That Fire on FALLING OI = Short-Covering / Stop-Run Fakeouts, ATR-Trailing Exit, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, SINGLE-VENUE strategy on ETHUSDT.BINANCE USD-M perpetual. It is the orthogonal-feed upgrade of the factory's one proven survivor (BTC dual-timeframe momentum confluence LS): instead of confirming a momentum breakout with a second PRICE timeframe (pure OHLCV, the 0/213 dead class per L42), it confirms the breakout with OPEN INTEREST — a genuinely orthogonal, well-populated catalog feed whose mechanism is positioning-flow, not a re-derivation of trend. Core observation: a price breakout is only sustainable when it is backed by FRESH capital taking new positions. A breakout on RISING open interest means new longs (upside) or new shorts (downside) are entering and pushing price — it tends to follow through. A breakout on FALLING open interest is short-covering (upside) or long-liquidation (downside) closing existing positions into a vacuum — no new conviction, so it typically fails and reverts. The strategy therefore takes a Donchian-channel breakout ONLY when open interest is simultaneously EXPANDING, and SKIPS (stays flat on) breakouts that fire while OI contracts. This is distinct from the dead cross-sectional OI-confirmed momentum ROTATION family (whose edge was cross-sectional dispersion across 8 names ranked and dollar-neutralized — this is a single-instrument time-series breakout gate, no ranking, no cross-sectional book) and from an OI-FLUSH reversal (opposite sign: this requires OI to RISE with the move, not collapse). Single-venue, single-instrument — no cross-venue joint-availability problem (L44), no options/IV feed (L43). ETHUSDT chosen for deep liquidity, well-populated OI history, and high volatility so breakout follow-through clears the 0.10% USD-M round trip with margin; diversifies away from the 2%-survival BTC tag. 3 parameters to resist overfit. Instruments: ["ETHUSDT.BINANCE"]. venue_type: futures. data_requirements: ["ohlcv", "open_interest"].

Implements the hypothesis as specified: a time-series Donchian breakout with an orthogonal OPEN-INTEREST confirmation gate, single instrument, single venue, no ranking and no second leg. THE GATE IS VALIDATED, NOT ASSUMED. I measured it on the real ETHUSDT 1-HOUR panel (57,696 bars) with per-instrument OI merged from the Vision metrics and fapi mirrors (68.5% of bars carry an OI reading within 6 hours). Running the exact entry/exit rules the code implements, the OI-expansion gate improves the per-trade gross edge in EVERY cell tested: at oi_lookback=72h and atr_trail=3.0 the gated book returns +0.265%/+0.268%/+0.288%/+0.224%/+0.237%/+0.346% per trade for donchian_lookback 24/36/48/72/96/120 versus +0.059%/+0.077%/+0.079%/+0.126%/+0.193%/+0.309% ungated - a broad, stable region rather than one lucky cell. I set the defaults at donchian_lookback=48, oi_lookback_hours=72, atr_trail=3.0: 438 trades over ~6.6 years (~66/yr, ~45 inside the 365-day sandbox window), +0.288% gross per trade against the ~0.100% USD-M round trip, i.e. ~+0.19% net, above the 0.15% floor. HONEST CAVEAT, stated once: the sign of the OI effect depends on the confirmation window - at a 24-hour OI lookback the relationship INVERTS in the event study (expansion +0.068% vs contraction +0.633% forward 48h), while at 72 and 120 hours it runs the way the hypothesis predicts (expansion +0.432% vs contraction +0.097%). So the edge is about multi-day positioning build-up, not hour-to-hour OI ticks; oi_lookback_hours is one of the three searched parameters and the optimizer will see both regions. IMPLEMENTATION: the OI series is parsed once into sorted numpy arrays and queried with searchsorted strictly at-or-before the bar timestamp (no look-ahead, no per-bar rescan that could blow the sandbox wall clock); the Donchian channel excludes the current bar so the breakout test is not self-referential; the ATR trail is re-derived each bar from the bar buffer and the positions ts_opened, so it is calendar-anchored and restart-safe rather than a process-local counter; when no OI reading is available the strategy simply does not enter, with no price-only fallback, because the gate is the edge being tested. Futures venue because the book goes both long and short; leverage stays 1.0 and per-trade risk is capped at 1% of equity through the ATR-derived stop distance with notional independently capped at 1x equity.

Thin, decaying momentum-breakout edge that will not clear the post-optimization robustness gates. Despite genuine positives — max_drawdown only 12.7%, four consecutive positive years (2022-2025), PF 1.24, and 434 trades — the governing metric is too weak: Sharpe 0.377 with a 95% CI [-0.329, 0.993] straddling zero (not statistically distinguishable from no skill and below the 0.5 OOS minimum), and information_ratio -0.67 (underperforms buy-and-hold ETH). The edge is outlier-carried (return_kurtosis 20.3) and the recent regime has decayed: 2026 is negative and the rolling Sharpe collapsed to ~-8 in July 2026. This is the OHLCV+open-interest breakout/momentum-continuation class (L9/L41, near-zero survival; its OI-backed-trend sibling was abandoned this session at Sharpe 0.17/PF 1.08). Entering the optimizer at Sharpe 0.38, best-of-225 selection would inflate it on noise and near-certainly fail the deflated-Sharpe gate (the selected Sharpe must sit well above the expected-max over ~225 trials, which a 0.38 base cannot), with the 15-day holdout falling in the decaying 2026 regime. avg_trade_return_pct 0.295% clears the fee floor, so this is a weak-OOS/overfit risk, not fee-fragility. The strong drawdown and multi-year consistency do not overcome a Sharpe indistinguishable from zero. Abandon at BACKTEST_REVIEW rather than spend 2 hours.

Single-instrument long-short breakout on ETHUSDT.BINANCE USD-M perp, 1-HOUR bars. calculate_signal returns the continuous position of price within a Donchian channel built from the prior donchian_lookback bars ((close-mid)/half-range, +/-1 at the channel edge). A breakout (|signal| > 1) is taken ONLY when open interest is simultaneously EXPANDING over the last oi_lookback_hours (OI(t)/OI(t-72h)-1 > 0), i.e. fresh positioning capital is entering behind the move; breakouts firing on CONTRACTING OI (short-covering on the upside, long-liquidation on the downside) are skipped entirely - not faded. OI is read per-instrument from supplementary_data[open_interest_by_instrument] (contract OI, hourly, multi-year) with the flat open_interest list as a fallback, always as-of the bar timestamp with a 6-hour staleness limit. Exit is an ATR trailing stop measured from the extreme reached since the positions own ts_opened. Sizing is risk-based: the initial trail distance costs 1% of equity, notional capped at 1x equity, leverage 1.0.

At backtest-review require full-history net per-trade above 0.15% and Sharpe CI clearly above zero before optimizing.

Marginal edge for the analyst (not a fault): sandbox avg_trade_return_pct is 0.104% NET over 83 trades — it clears the ~0.10% round-trip but is BELOW the 0.15% analyst abandonment floor, with Sharpe 0.0077, CI [-1.73, 1.57] straddling zero, PSR 0.50. Not structurally fee-doomed (variable/large moves, PF 1.035), so not a critical fee fail, but the analyst must confirm the thin edge survives the full backtest.

Constrain oi_lookback_hours to >=48-72h and require walk-forward to confirm the expansion sign is stable OOS.

Robustness fragility (developer-flagged): the SIGN of the OI effect depends on oi_lookback_hours — it inverts at 24h (expansion +0.068% vs contraction +0.633% fwd-48h) but runs as hypothesized at 72h/120h. Since oi_lookback_hours is searched, the optimizer can select an inverted-sign region. Code is correct; the risk is parameter selection.

Optionally persist a running extreme in position state instead of rescanning, for exactness on long holds.

Minor latent edge: should_exit rescans the bar buffer back to ts_opened for the trailing extreme; a hold longer than the retained buffer (~74 bars) would truncate the extreme window. Sandbox avg hold ~22h, well inside the buffer, so it doesn't bite in practice.

Genuinely orthogonal OI-confirmation gate (not a second price series), single venue/instrument — verification-safe

Excellent risk control: max_drawdown only 12.7% (CI high 38.7%)

Positive in four consecutive years (2022 +4.7%, 2023 +12.2%, 2024 +16.9%, 2025 +20.3%) — better consistency than most of the class

avg_trade_return_pct 0.295% clears the fee floor; 434 trades is a solid sample; PF 1.24 above the 1.2 line

Sharpe 0.377 with 95% CI [-0.329, 0.993] straddling zero — risk-adjusted edge not statistically distinguishable from zero, below the 0.5 OOS minimum

information_ratio -0.67 — underperforms buy-and-hold ETH on a risk-adjusted basis

Edge is outlier-carried (return_kurtosis 20.3) and the recent regime has decayed: 2026 negative, rolling Sharpe collapsed to ~-8 in July 2026

OHLCV+OI breakout/momentum-continuation class (L9/L41, near-zero survival); the OI-backed-trend sibling was abandoned this session at Sharpe 0.17/PF 1.08

15-day holdout falls in the decaying 2026 regime; best-of-225 selection on a Sharpe-0.38 signal would near-certainly fail deflated-Sharpe/holdout

This strategy was a thoughtful orthogonal-feed upgrade of the factory's one proven survivor: rather than confirming a breakout with a second price timeframe (a dead pure-OHLCV class), it gated an ETH Donchian breakout on simultaneous open-interest expansion — trading only breaks backed by fresh positioning capital and skipping the short-covering fakeouts on falling OI. It showed genuine strengths the class rarely has: a tiny 12.7% max drawdown, profit factor 1.24, a solid 434-trade sample, and four consecutive positive years from 2022 to 2025. But the risk-adjusted edge was too thin to matter — a Sharpe of 0.377 with a confidence interval straddling zero, a negative information ratio against buy-and-hold ETH, outlier-carried returns, and a 2026 regime that turned negative with rolling Sharpe collapsing to ~-8. The analyst abandoned it at backtest review as a weak, decaying member of the near-zero-survival OHLCV+OI breakout class whose sibling had already died that session, judging that a Sharpe-0.38 base would near-certainly fail the deflated-Sharpe and holdout gates and that strong drawdown control could not rescue an edge indistinguishable from zero.

Excellent drawdown control and four straight positive years are not enough — if the Sharpe is statistically indistinguishable from zero, the information ratio is negative versus buy-and-hold, and the edge is outlier-carried and decaying in the recent regime, an OHLCV+OI breakout enters the robustness gauntlet too weak to survive best-of-N deflation.

It was abandoned at the BACKTEST_REVIEW gate: a Sharpe indistinguishable from zero and below the 0.5 OOS minimum, a negative information ratio, and a decayed recent regime with the 15-day holdout falling in negative 2026. As an OHLCV+OI breakout/momentum-continuation strategy (near-zero survival class, whose OI-backed-trend sibling was abandoned that session at Sharpe 0.17), entering the optimizer at Sharpe 0.38 would near-certainly fail the deflated-Sharpe/holdout gates, so its strong drawdown and multi-year consistency could not overcome the weak Sharpe. It never reached optimization.

A long-short, single-instrument, single-venue strategy on the ETHUSDT.BINANCE USD-M perp that takes a Donchian-channel breakout only when it is confirmed by a simultaneous expansion in open interest (fresh positioning capital entering) and skips breakouts firing on falling OI (short-covering / stop-run fakeouts), exiting via an ATR trailing stop — an orthogonal-feed upgrade of the factory's dual-timeframe momentum survivor that confirms with OI rather than a second price series.

It had unusually good risk control but a thin, decaying edge: max drawdown only 12.7% (CI to 38.7%), profit factor 1.24, avg_trade_return_pct 0.295% (above the fee floor), 434 trades, and positive in four consecutive years (2022 +4.7% through 2025 +20.3%). But the governing metric was weak — Sharpe 0.377 with CI [-0.329, 0.993] straddling zero, information ratio -0.67 (underperforming buy-and-hold ETH) — and the edge was outlier-carried (kurtosis 20.3) with the recent regime decayed (2026 negative, rolling Sharpe collapsed to ~-8 in July 2026).