BtcOiFueledTrendContinuationLS

BTC Open-Interest-Fueled Trend Continuation, Long-Short (Single-Instrument BTCUSDT.BINANCE Perp — NEW Information Source: Ride the Price Trend ONLY While a Multi-Day OPEN-INTEREST UPTREND Confirms New Leveraged Capital Is Fueling It, ATR-Trailing, 3-Parameter)

A LONG-SHORT directional strategy on the BTCUSDT.BINANCE USD-M perpetual that rides the price trend ONLY while a sustained, MULTI-DAY OPEN-INTEREST UPTREND confirms fresh leveraged capital is fueling the move — a new-information-source (OI) idea used in the CONTINUATION direction, per the meta-learner's L38 mandate and aligned with the one edge that survives (momentum), unlike the many BTC fades that have all died (liquidation, OI-divergence, DVOL, skew, false-breakout, deleveraging). It is distinct from the two dead OI constructions: the OI-CONFIRM-momentum (single-bar 'OI up this bar') and the OI-DIVERGENCE FADE (fade moves where OI falls). This uses a SLOWER, REGIME-level signal: OI in a persistent multi-day uptrend (new-money leverage entering) marks a DURABLE trend regime worth riding, while a trend on flat/contracting OI is unfueled and avoided. The thesis is that leverage-fueled trends persist because new positions keep pushing price, whereas trends on stagnant OI are exhaustion-prone. It trades the liquid BTC perp (dev-safe, signal=execution), single instrument, long-short (fills the 13.5% bucket vs 86.5% long-only), 4H cadence for a measurable ~100-200 trade sample, ATR trail to ride continuation. Open-interest is a WELL-COVERED Binance feed (avoiding the taker-ratio/L-S coverage ceilings, the options no-trades/no-edge failures, and the DVOL history limits). 3 parameters.

ITERATION 3 -- fixes exactly what QA measured (negative expectancy: 30% win rate against 1.56 reward:risk) with a four-number recalibration; NO structural change, no new imports, same signal, same four exits, same sizing, same three tunables. METHOD: I rebuilt the strategy's exact decision path offline against the real 4H catalog bars and the real OI series (metrics/open_interest_value, 5-min, 2020-09 to 2026-07) and REPRODUCED QA's sandbox numbers before changing anything -- my replay of the iteration-2 config over the sandbox's last-365-day window gives n=32, total_return -7.9%, PF 0.66, avg_trade_return_pct -0.68%, 17 long (QA saw 30, -7.61%, 0.670, -0.637%, 16 long). That validated the harness, so the changes below are measured, not guessed. DIAGNOSIS: the OI gate was NOT the problem -- three exits were firing ahead of the trail and cutting BOTH tails, which hurts winners more than losers. FIX: (1) exit_frac 0.5 -> 1.0, so the reversal exit needs a FULL flip to the opposite entry level rather than a mere halving of conviction (a rolling momentum reading decays toward zero mechanically on any pause, so the 0.5 level was closing winners on window arithmetic); (2) trail_atr_mult 3.0 -> 4.0, since a 3-ATR trail on 4H BTC sits inside the ordinary noise band of a live trend -- this also widens the sizing denominator max(trail, gap) so positions get ~25% smaller, i.e. risk per trade is unchanged rather than increased; (3) mom_lookback 12 -> 10 and entry_thresh 0.55 -> 0.4, keeping the sample at the ~35 entries/year cadence QA asked for now that the exits no longer churn it. RESULT (same replay): full history 2020-2026 n=221, PF 1.25, win_rate 0.45, avg_trade_return_pct +0.497% (3.3x the 0.15% fee floor), profitable in 6 of 7 calendar years (only 2022 negative) and on BOTH sides (long +28.8k, short +10.9k); the QA sandbox window flips from -7.9%/PF 0.66 to +13.6%/PF 1.77/avg +0.70%. NOT AN ISLAND: every neighbour cell (mom_lookback 10, entry_thresh 0.35-0.45, trail 3.5-4.5) is positive in both windows, and perturbing the LOCKED constants (oi_trend_days 3/5/10, oi_ref 0.3-1.0, max_hold 30-120, oi_exit_z 0.75-1.5) keeps full-history PF 1.15-1.30. ANSWERING QA'S ADDITIVITY QUESTION DIRECTLY: holding this exit structure fixed and forcing the fuel weight to 1.0 (pure price momentum, gate off) drops per-trade edge from 0.497% to 0.133% and PF from 1.25 to 1.13 -- the OI gate roughly triples the per-trade edge, it just trades a third as often, so it IS additive on the metric the analyst floors on. Invariant preserved: the unfuelled branch is clipped at 0.9 x 1.0 x 0.4 = 0.36, strictly below both the 0.4 entry threshold and the 0.4 reversal-exit level. leverage 1.0 and no sizing path reads leverage; BINANCE USD-M because a fuelled downtrend is ridden short.

Marginal edge in the momentum/continuation class (L9, zero survivors at PF < 1.2). Despite a genuine per-trade edge (avg_trade_return_pct 0.635%, above fee floor) and correct long-short mechanism, the composite risk-adjusted profile is too thin to justify 2 hours of optimization: profit_factor 1.16 (below the 1.2 meaningful line and below the developer's own claimed 1.25, so the OI gate under-delivers live), Sharpe 0.289 with a 95% CI [-0.30, 0.96] that already straddles zero, and information_ratio -0.672. Optimization selecting the best of ~225 trials on a signal this weak inflates Sharpe on noise; the realistic OOS ceiling won't clear the 0.5 minimum, and the near-certain outcome is an overfit-abandon in ANALYZING (the exact fate of the momentum/continuation siblings this session). Not a code or premise bug to iterate on — the edge is simply below what the promotion gates require, so abandon rather than burn the optimization budget.

Long/short BTCUSDT.BINANCE perp that rides the price trend only while a multi-day open-interest UPTREND confirms fresh leveraged capital is fuelling it. Every 4H bar: price momentum over mom_lookback bars is expressed in random-walk sigma units (ATR*sqrt(n)); the standardized 5-day log change in open interest is turned into a direction-agnostic fuel weight w = clip(oi_z/0.5, 0, 1); the signal is momentum x fuel. Long on a fresh cross up through entry_thresh sigma, short on a cross down through -entry_thresh. Below the fuel floor the signal is damped and clipped strictly below the exit level, so an unfuelled trend cannot be entered and fading fuel cannot by itself close a position. Exits: 4-ATR trailing stop from the best close since entry (intrabar-detected, booked at the close), a trend FLIP to the full opposite threshold, confirmed deleveraging (oi_z <= -1), and a 7-day max hold. Gap-aware risk sizing off max(trail, 3) ATRs at 2% equity risk, capped at 50% of equity notional, leverage 1.0. If the OI feed is missing/stale/too short no position is taken -- no price-only fallback.

Verification failed (Layer 4 — QA review): - MEASURABLE NOW, BUT NET-NEGATIVE -- the OI-fuel gate is non-additive, answering the developer's own flagged additivity question. The iteration-2 recalibration (oi_ref 1.0->0.5, entry_thresh 1.0->0.55, deleveraging exit decoupled) correctly lifted the count 8 -> 30 and symmetrized it (16 long / 14 short), so the sample is now readable. And it reads decisively negative: total_return -7.61%, profit_factor 0.670, Sharpe -0.731 (CI [-2.42, 0.99]), win_rate 0.30, avg_trade_return_pct -0.637% (negative and below the fee floor), expectancy -$254/trade. The reward:risk is favorable (avg_win $1,715 vs avg_loss $1,098, ~1.56) but a 30% win rate needs RR > ~2.3 to break even, so the gated trend-follower loses. This is exactly the risk the developer pre-flagged: OI builds and price trends are positively correlated by construction, so the fuel gate is non-additive -- it does not improve trend-following, and here it subtracts.

At BACKTEST_REVIEW / walk-forward: confirm the gate-on minus gate-off per-trade edge is materially positive on the full sample; verify the +7% is not one-trend-dependent (avg_trade above the fee floor with the top 1-2 wins excluded -- the sandbox +6.80% is ~97% one $6,567 win); check OOS Sharpe > 0, IS/OOS ratio < 3.0, holdout >= 70% of WF-OOS Sharpe. The developer's parameter-neighbourhood check (neighbours positive in both windows, locked-constant perturbations keep full-history PF 1.15-1.30) is encouraging but must be reproduced by the sensitivity/walk-forward phases, not taken on trust.

VERIFY THE GATE ADDITIVITY AND THE FULL-HISTORY EDGE ON THE REAL BACKTEST -- the No.1 analyst item, because it is offline evidence I cannot independently confirm. The developer answered my iteration-2 additivity concern directly: with the exits fixed, forcing the fuel weight to 1.0 (gate off, pure momentum) drops per-trade edge 0.497% -> 0.133% and PF 1.25 -> 1.13 across 2020-2026, i.e. the OI gate roughly triples the per-trade edge -- which, if it reproduces, overturns my iter2 'non-additive' read (that read was confounded by the winner-truncating exits, not a clean gate test). He also reports full-history n=221, PF 1.25, avg_trade +0.497%, profitable in 6 of 7 years and on both sides. The 3-phase optimization must reproduce this on real data: run the gated vs UNGATED trend over the full span and confirm the gate lifts per-trade edge and PF, and that avg_trade stays above the 0.15% floor with both legs contributing.

The full multi-year backtest (the developer's 221-trade span) is the arbiter -- the 363-day window is one year and understates or overstates the per-trade edge depending on regime. Do not promote on the sandbox alone; require the fee-floor margin and the tail-recurrence to hold across the full history and the holdout.

THIN SANDBOX MARGIN + ONE-WIN DEPENDENCE. In the 363-day sandbox, avg_trade_return_pct is 0.195% -- only ~1.3x the 0.15% fee floor and well below the developer's claimed 0.497% full-history figure -- and total_return +6.80% is dominated by a single trade (largest_win $6,567 ~= 97% of the ~$6,780 net profit; remove it and the year is near-breakeven). Sharpe 0.554 has a CI [-1.14, 2.10] straddling zero at n=28. For a trailing-stop trend follower a fat right tail with a ~50% win rate is the EXPECTED shape (as with the KAMA and DVOL-regime trend strategies passed this session), not a defect, but at this thin sandbox margin the full-sample edge must clear the floor with room, not just in one favourable year. Capacity $57.9M / impact 4.16% are healthy.

Mechanism matches hypothesis: 183 trades, 117 long / 66 short, genuine long-short continuation with direction from price and OI as a fuel gate

avg_trade_return_pct 0.635% is 4x the futures fee floor — NOT fee-fragile; per-trade edge is real

Healthy sample (183 trades) and acceptable max_drawdown 17.3%; positive in 6 of 7 years

profit_factor 1.16 is below the 1.2 meaningful-edge line; developer's own rationale claimed ~1.25, so the OI gate is less additive live than modeled

Sharpe 0.289 with sharpe_ci_low -0.30 — the CI straddles zero before optimization even runs; risk-adjusted edge is not distinguishable from zero

information_ratio -0.672 and ~3.6% CAGR; rolling Sharpe deeply negative through 2022 (to -5.8) — inconsistent across regimes

This is the OHLCV momentum/continuation class (L9) with PF < 1.2, which has no survivors — optimization selecting best-of-225 trials will overfit a marginal signal

BtcOiFueledTrendContinuationLS tried to distinguish itself from two dead OI constructions by using open interest as a slow, regime-level fuel gate — riding BTC price trends only while a multi-day OI uptrend confirmed fresh leveraged capital was pushing the move. Three iterations of recalibration (loosening the entry gate, then fixing exits that were truncating winners) produced a clean 183-trade sample with a genuine per-trade edge of 0.635%, but the composite stayed marginal: profit factor 1.16, Sharpe 0.289 with a CI straddling zero, and an information ratio of -0.672. The analyst abandoned it at backtest-review rather than optimize, judging it another marginal entry in the momentum/continuation class that has no survivors below a 1.2 profit factor. The decision was explicitly budget-driven: the edge was simply below what the promotion gates require, so optimization would only inflate Sharpe on noise before an inevitable overfit-abandon.

A real, fee-clearing per-trade edge is not enough — if the risk-adjusted profile (PF < 1.2, Sharpe CI straddling zero, negative information ratio) is marginal before optimization, running best-of-225 trials will just overfit noise, so it's cheaper to abandon at backtest-review than to burn optimization on a signal below the promotion floor.

It was abandoned at the pre-optimization BACKTEST_REVIEW gate: despite a real per-trade edge and correct mechanism, profit factor 1.16 sat below the 1.2 meaningful-edge line and the Sharpe CI already straddled zero, so the reviewer judged the momentum/continuation class (with no survivors at PF < 1.2) too weak to justify spending the optimization budget, expecting a near-certain overfit-abandon.

Ride the BTCUSDT.BINANCE perp long/short with the price trend, but only while a sustained multi-day open-interest uptrend confirms fresh leveraged capital is fueling the move — OI used as a regime-level 'fuel' gate rather than a directional signal.

The backtest gave a healthy two-sided sample (183 trades, 117 long / 66 short) with a genuine per-trade edge — avg per-trade return 0.635% (~4x the fee floor) — but a thin risk-adjusted profile: profit factor 1.16, Sharpe 0.289 (CI [-0.30, 0.96] straddling zero), information ratio -0.672, ~+26% total return, max drawdown 17.3%, with rolling Sharpe deeply negative through 2022.

Verification failed (Layer 4 — QA review): - UNMEASURABLE AND NET-NEGATIVE -- the OI-fuel gate is far too binding and adds no edge on the sample it produces. 8 trades over 363 days (~48 over a multi-year span, below the ~100 floor, L16) against the developer's own modeled 30-60/year: the conjunction of a confirmed multi-day OI uptrend (oi_z >= 0.25 x oi_ref over 5 days) AND a >=1-sigma price-trend cross fires only 8 times a year. And that tiny sample is decisively negative: total_return -3.35%, profit_factor 0.631, Sharpe -0.479, win_rate 0.25, avg_trade_return_pct -1.151%, expectancy -$419/trade -- with long_win_rate 0/6 (every one of the six longs lost; only the two shorts won). return_skew 3.08 / kurtosis 42 confirm one or two trades dominate. This is the developer's own flagged risk realized: OI builds and price trends are positively correlated by construction, so the gate mostly re-selects 'strong trend' rather than adding independent information -- and here it is both over-selective and non-additive.