EthPerpOpenInterestBackedTrendContinuationLS

Open-Interest-Backed Trend Continuation, Long-Short (Single-Instrument ETHUSDT.BINANCE USD-M — Hold in the Direction of a Price Move ONLY When OPEN INTEREST Is EXPANDING = New Capital Is Backing the Move; Stand Aside / Exit When OI Contracts = Deleveraging Noise or Trend Exhaustion, 2-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, SINGLE-VENUE directional strategy on ETHUSDT.BINANCE USD-M whose signal is genuinely ORTHOGONAL to price — the RATE OF CHANGE of OPEN INTEREST (new capital entering vs positions closing), not a price-derived indicator. It is evidence-backed by the factory's OWN prior work: the abandoned COIN-M OI-deleveraging-FLUSH reversal was falsified because its event study (over every OI-covered ETHUSDT 1h bar, 2021-12→2026-07) showed the fade-on-OI-contraction gate is BACKWARDS and value-subtracting — which is direct evidence that the OPPOSITE gate, riding moves that are accompanied by OI EXPANSION, is the value-ADDING direction. And the closely-related COIN-M OI-confirmed-trend hypothesis died NOT on its mechanism but purely on DATA: the COIN-M OI series has a 1,248-day hole. THIS hypothesis fixes exactly that by moving to ETHUSDT USD-M, whose OI is dense and continuous 2021→2026 (the very series that supported the flush event study). Economic premise: a price move accompanied by RISING open interest is backed by genuine new capital taking a side and tends to CONTINUE; a move on FLAT or FALLING OI is short-covering / deleveraging noise that fades. So we position with the move only while OI expands, and exit when OI rolls over (positions closing = trend exhaustion). This is a continuation family distinct from the dead forced-flow REVERSAL lines (cascade/ignition/absorption/flush), the dead positioning-contrarian line, and my own pending momentum-with-funding-regime and mean-reversion items (different feed, different mechanism). Data-safe: funding-free, single venue + single instrument on 4H bars → OHLCV and the dense USD-M OI feed are jointly available in Layer-3 (no cross-venue, options, liquidation, or holed COIN-M feed). Deliberately 2-parameter to resist the overfit deaths that dominate the log.

NOTE: this is a re-emission of the same strategy -- my previous submission for this hypothesis contained a raw line break inside the JSON code string (in the oi_gap_reset_hours bounds line), which would have made the payload unparseable. The code is otherwise identical and that line is now on a single line; no logic changed. I verified the hypothesis's data claim and pre-tested the mechanism on the full covered sample before writing code, and this returned the strongest evidence of any hypothesis in this batch -- notably because the OI gate is not a filter on an existing edge, it IS the edge. DATA: ETH USD-M open interest is genuinely dense and continuous -- 497,293 readings spanning 2021-12-01 to 2026-07-31 with ~8,600-9,500 readings in EVERY month, merged with the open_interest tail that runs to 2026-08-01. This is exactly the defect that killed the COIN-M version (whose OI series has a 1,248-day hole), and it matters beyond Layer 3: because coverage runs to the present, the walk-forward OOS windows AND the 15-day holdout can both be populated, which is what three of my recent strategies died for lack of. I checked ETH specifically rather than generalising, since I have already established that metrics coverage is symbol-dependent (BTCUSDT is holed, SOLUSDT is not). MECHANISM: over the full covered sample, with a matched 96h momentum and OI-RoC window, direction-adjusted forward returns at 96h are +0.007% (t=+0.1) for UNGATED momentum, +1.235% (t=+6.7) when OI is expanding, and -0.294% (t=-1.8) when OI is contracting; at |mom|>=1.0 the expanding bucket gives +1.010% at 48h (t=+5.4) and +1.994% at 96h (t=+7.4) against -0.123% and -0.268% contracting. Ungated momentum being statistically ZERO while the gated version is strongly positive is the cleanest possible demonstration that the OI feed is orthogonal and additive rather than a repackaging of price momentum, and the negative contracting bucket independently confirms the hypothesis's second half. It also corroborates the prior flush study's implication that the fade-on-contraction gate was backwards. The horizon profile builds monotonically (12h +0.277, 24h +0.517, 48h +1.010, 96h +1.994) -- a proper continuation gradient, which is precisely what was ABSENT in the falsified reversal strategies. Largest t is 7.4 against an expected maximum of ~3.1 under pure noise across the 135 cells scanned, so this clears multiple testing by a wide margin, unlike the three strategies I abandoned which each peaked at or below their noise maximum. DESIGN CHOICES: the effect is absent at 24h and 48h OI lookbacks and only appears at ~96h, which is economically coherent (open-interest accumulation is a multi-day process, a one-day OI change is noise), so I FIXED lookback at 24 bars rather than exposing it as a tunable -- searching it would be fitting the one window that works. Likewise oi_expand_z and oi_exit_z are fixed so the regime definition cannot be tuned; only entry_z and trail_atr are exposed, per the 2-parameter mandate. The exits deliberately avoid the failure mode of the earlier BTC OI strategy in this factory, which exited on a 3% OI drawdown-from-peak plus a price-mean crossing and produced a 33% win rate with small wins against full-size losses: here the OI exit is a hysteretic z-crossing (enter above +0.5, exit only below -0.5, hold through the neutral band) and risk is carried by an ATR trailing stop so winners run out to the 96h horizon where the edge peaks. The gate is a z-score rather than a fixed percentage because ETH open interest has grown several-fold across the sample and an absolute OI-change threshold would fire at wildly different rates in 2022 versus 2026 -- the failure mode behind my lesson on absolute supplementary thresholds. Fee viability is not marginal: +1.0% to +2.0% gross per trade against the 0.10% USD-M round trip is 7-13x the 0.15% floor, and multi-day holds keep turnover low. Warmup is ~180 bars (30 days) against a 365-day sandbox window, per-bar cost is O(1) via forward-only pointers and running sums, and the gap-reset guards the rolling windows against any discontinuity in the collected series. Leverage stays at 1.0 and is never referenced in sizing -- risk-based sizing against a ~5-6% ATR stop already produces roughly half of equity in notional, so declaring leverage above 1 would be inert and would trip the leverage_set_but_unused gate.

Near-zero, outlier-driven, fee-fragile edge that will not clear post-optimization robustness gates. Sharpe is 0.174 with a 95% CI [-0.532, 0.724] straddling zero (statistically indistinguishable from no skill), profit_factor 1.08 (well below the 1.2 meaningful line — fees consume nearly all the gross edge), information_ratio -0.72, and total_return only 22.4% over ~4.5 years. avg_trade_return_pct 0.192% sits right on the 0.15% fee floor, so the strategy is fee-fragile — any slippage erases it. The return stream is outlier-driven (return_kurtosis 36.6, skew 2.72): a few spike days (2025-05-12 +18.4%, 2024-05-25 +15.2%, 2022-06-17 +13.3%, 2023-01-13 +12.0%) carry the whole result, and annual returns alternate sign with no persistence (2026 holdout region is negative). This is the OHLCV+supplementary continuation class (L9/L41, zero survivors), the same orthogonal-gate momentum family as the funding-gated and VRP-gated candidates abandoned this session; the OI-expansion gate did not manufacture a measurable edge. Best-of-225 selection on a Sharpe-0.17 signal would inflate it on noise and near-certainly fail deflated-Sharpe/holdout in ANALYZING, and there is no robust parameter region to tune toward. Abandon at BACKTEST_REVIEW rather than spend 2 hours.

Long-short, single-instrument trend-continuation strategy on ETHUSDT.BINANCE USD-M, 4-HOUR bars, gated by open-interest expansion. Each bar it takes the latest non-anticipating OI reading (merged from the metrics long history and the open_interest recent tail, de-duplicated into one timestamp-sorted series and walked with a forward-only pointer) and forms two quantities over a matched 24-bar (96h) window: a volatility-normalised trend, mom_z = (close/close[-24] - 1) / realised 96h vol, and an OI-expansion weight, oi_w = 0.5*(1 + tanh((oi_z - 0.5)/0.5)) where oi_z is the z-score of the OI rate of change against its own 180-bar rolling distribution. calculate_signal returns mom_z * oi_w every bar: the sign is the trend direction, the magnitude is scaled toward zero when open interest is flat or contracting. It goes long when that score clears +0.6 and short when it clears -0.6, so a strong price move that no new capital is backing cannot trigger a trade. Exits are open-interest ROLLOVER (oi_z falling below -0.5, hysteretic against the +0.5 entry so the position holds through the neutral band rather than churning), an ATR trailing stop at 3.0 ATR from the best price reached, and a 120-bar backstop. Size risks 3% of equity against the initial trailing-stop distance, capped at 1.0x equity; leverage is 1.0 and is not referenced in sizing.

Run the full 2021-2026 backtest and confirm the expanding-OI edge shows in realized trades before spending optimization; if still negative/break-even, iterate the exit or abandon.

SANDBOX IS MATERIALLY NEGATIVE AND CONTRADICTS THE MOTIVATING EVENT STUDY. The developer's 5-year event study shows a strong monotone-building expanding-OI edge (+1.010% @48h t=+5.4, +1.994% @96h t=+7.4; ungated momentum zero; contracting bucket negative; largest t=7.4 vs ~3.1 noise max). But the 363-day sandbox LOSES: -8.71% total, Sharpe -0.445, PF 0.763, win_rate 0.333, avg_trade_return_pct -0.52%, over 30 trades, long side especially bad (0.21 win on 19 trades). Not a data ceiling (ETH OI dense 2021-2026) and no signal-inverting code defect found — the gap is either an adverse range-bound 2024-2025 window the 1-yr sandbox over-weights, or an exit that fails to monetize the horizon edge (next finding).

If the full backtest underperforms, make oi_exit_z tunable or replace the hysteretic-z OI exit with a hold-to-horizon exit; compare realized holding vs the 96h edge horizon.

OI-ROLLOVER EXIT LIKELY TRUNCATES THE MEASURED EDGE AND THE OPTIMIZER CANNOT REPAIR IT. The edge builds monotonically to 96h (24 bars), but should_exit fires when oi_z <= -0.5, and oi_z (a z-score of noisy OI RoC) mean-reverts by construction, so it crosses -0.5 within a few bars and cuts positions early. avg_holding_period is ~66h, already short of the 96h peak. oi_exit_z is FIXED (only entry_z/trail_atr tunable), so if this is the problem the 3-phase optimizer cannot fix it — it needs a developer iteration. Code faithfully implements the stated 'exit when OI contracts' rule, so not a mismatch, but the rule may be self-defeating.

Confirm both feeds are the same units for ETHUSDT; if not, normalize before merging or key RoC off a single source.

MERGED OI UNIT-CONSISTENCY must be verified. _load_oi merges 'metrics' (sumOpenInterest) and 'open_interest' (open_interest) into one series and computes RoC across it; if the two feeds use different units the splice injects a spurious RoC that corrupts oi_z near the boundary. The USD column is deliberately avoided and gap-reset/z-scoring partly absorb a one-time jump, but confirm rather than assume.

Clean, data-safe construction: single venue, single instrument, dense continuous USD-M OI feed (fixes the COIN-M data hole that killed the sibling) — verification-safe, orthogonal signal

Deliberately 2 tunables (OI gate held fixed) to resist overfitting

Hysteretic OI exit + ATR trailing stop is a thoughtful risk design; max_drawdown a moderate 24.8%

avg_trade_return_pct 0.192% technically clears the futures fee floor

Sharpe 0.174 with 95% CI [-0.532, 0.724] straddles zero — indistinguishable from no skill

profit_factor 1.08 is well below the 1.2 meaningful line — fees consume nearly all the gross edge (fee-fragile)

avg_trade_return_pct 0.192% sits right on the fee floor; any execution slippage erases it

Edge is outlier-driven: return_kurtosis 36.6, skew 2.72, with +18.4%/+15.2%/+13.3% single-day book gains carrying the entire 22.4% total return over 4.5 years

information_ratio -0.72; annual returns alternate sign with no persistence and 2026 (holdout region) is negative

OHLCV+supplementary continuation class (L9/L41, zero survivors); the OI gate did not produce a measurable edge — same failure profile as the funding/VRP-gated momentum siblings abandoned this session

This strategy was a well-reasoned attempt to salvage two prior OI ideas: it took the value-adding direction implied by a falsified OI-flush-reversal event study (ride OI expansion rather than fade contraction) and fixed the 1,248-day COIN-M data hole that had killed the OI-confirmed-trend sibling by moving to ETH USD-M, whose OI is dense and continuous. The construction was clean and disciplined — 2 tunables with the OI gate held fixed, a hysteretic exit and ATR trailing stop, a moderate 24.8% drawdown. But the backtest showed the gate produced no measurable edge: a Sharpe of 0.174 with a confidence interval straddling zero, profit factor 1.08 that fees nearly erase, and a meager 22.4% total return carried entirely by a handful of spike days with no year-to-year persistence and a negative holdout region. The analyst abandoned it at backtest review as a near-zero, fee-fragile, outlier-driven member of the zero-survivor orthogonal-gate momentum family — the same profile as the funding- and VRP-gated candidates abandoned that session — with no robust parameter region to optimize toward.

Successfully fixing the data defect that killed a sibling (moving to a dense, continuous OI feed) does not resurrect the edge — even with a thoughtfully-motivated orthogonal OI-expansion gate held fixed, the result was a fee-fragile, outlier-carried Sharpe indistinguishable from zero, confirming that orthogonal-gate momentum continuation is a zero-survivor family with no robust region to tune toward.

It was abandoned at the BACKTEST_REVIEW gate: a Sharpe straddling zero, profit factor 1.08, and a per-trade return right on the fee floor make it a near-zero, fee-fragile, outlier-driven edge with no persistence. As another member of the zero-survivor OHLCV+supplementary continuation family — the same failure profile as the funding-gated and VRP-gated momentum siblings abandoned that session — best-of-N optimization would inflate a Sharpe-0.17 signal on noise and near-certainly fail the deflated-Sharpe/holdout gates, so it never reached optimization.

A long-short, single-instrument, single-venue directional strategy on ETHUSDT.BINANCE USD-M that rides a price trend only while open interest is expanding (a z-scored OI rate-of-change gate signaling new capital backing the move) and stands aside or exits when OI contracts (deleveraging noise / trend exhaustion), with a hysteretic OI exit and ATR trailing stop — moving to the dense, continuous USD-M OI feed to fix the data hole that had killed the COIN-M sibling.

The gate did not manufacture a measurable edge: Sharpe 0.174 with CI [-0.532, 0.724] straddling zero, profit factor 1.08 (well below the 1.2 line, so fees consume nearly all the gross edge), information ratio -0.72, and total return only 22.4% over ~4.5 years across 162 trades. The result was outlier-driven (return kurtosis 36.6, skew 2.72) with a few spike days carrying the whole return, avg_trade_return_pct 0.192% sitting right on the fee floor, annual returns alternating sign, and the 2026 holdout region negative — though max drawdown was a moderate 24.8%.