Range-Fade Gated by Open-Interest Flatness, Single-Venue Long-Short (Single-Instrument BNBUSDT.BINANCE USD-M — Buy the Lower Band / Sell the Upper Band ONLY When Open Interest Is FLAT = a Balanced, Leverage-Neutral RANGING Market That Mean-Reverts; Stand Aside When OI Is EXPANDING = New Leverage Building a Trend; Mid-Band Exit + ATR Stop, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, SINGLE-VENUE RANGE mean-reversion strategy on BNBUSDT.BINANCE USD-M perpetual. This is a deliberately different STRATEGY TYPE from my in-flight ideas (all of which are momentum/continuation or extreme-fade reversals) and it uses OPEN INTEREST in a genuinely new ROLE: not as a directional-momentum confirm (OI expansion), but as a RANGING-REGIME CLASSIFIER via OI FLATNESS. Rationale for the regime gate: naive band mean-reversion (Bollinger/Keltner fade) fails in crypto because it gets run over in trends — but trends in perps are built by fresh leverage ENTERING (rising OI), whereas balanced, mean-reverting ranges occur when open interest is FLAT (no net new leverage, participants two-sided and rotating). So the strategy only fades the range when OI is flat (the market is genuinely balanced and reverts), and stands completely aside when OI is expanding (leverage is building a directional move — do NOT fade it). The OI-flatness gate is orthogonal to price and is the element that makes this not a dead pure-OHLCV mean-reversion (satisfies L42) — it is a regime filter, not another price indicator. Reliability note: this stays on BINANCE USD-M where OI is well-populated in the sandbox (unlike COIN-M, whose OI feed loops in verification), uses core OHLCV + the USD-M OI series only, and is kept to THREE real parameters after a prior strategy died at Layer-2 from parameter bloat. Single-venue, single-instrument — no cross-venue leg problem (L44), no options feed (L43), no taker/premium feeds (which have shown Layer-2 fragility). BNBUSDT chosen as a liquid major perp that spends long stretches range-bound (frequent fade setups) with wide-enough bands that reversion targets clear the 0.10% USD-M round trip, and it is a fresh instrument vs BTC/ETH/SOL/DOGE/AVAX/LINK already in flight. ENTRY (4h bars + same-venue OI): compute a `band_len`-period moving average and ±`band_k`×stdev bands, and oi_change = OI_now/OI[`band_len`]-1; require |oi_change| < a flatness threshold (derived from band_len, not a 4th free param — e.g. within ±3%) to classify a RANGING regime; then ENTER LONG on a close at/below the lower band, ENTER SHORT on a close at/above the upper band; skip all signals when OI is expanding (trend regime) or when OI is stale/missing. Risk 2% equity/trade, ATR(14) hard stop beyond the band extreme, max exposure 25%, leverage 2x. EXIT on first of: price reverts to the mid-band (primary target); opposite band reached (stop-and-reverse candidate); ATR stop beyond the entry band (range broke — likely a regime change to trend); 18-bar (~3 day) time stop. Instruments: ["BNBUSDT.BINANCE"]. venue_type: futures. data_requirements: ["ohlcv", "open_interest"].

ITERATION 2 — single-line fix for the reported `leverage_set_but_unused` gate, nothing else touched. The previous submission DID consume leverage in position_size, but it read it as `getattr(self.config, "leverage", 1.0)`; the static check looks for a direct `self.config.leverage` reference and did not recognise the getattr form, so it flagged the leverage as inert. The fix is to read it directly: `lev = max(1.0, float(self.config.leverage))`, plus a comment spelling out that this is what makes the 2x real (the gross-notional cap becomes 0.25 * 2 = 0.50 of equity, exactly twice the 1x notional, so the backtest is genuinely not identical to 1x). This is the same one-line change that cleared this gate on the LINK strategy. Nothing else changed: imports are untouched (no banned modules), the class structure, the continuous band-z signal, the OI-flatness regime classifier, every entry/exit condition, all thresholds, and the risk-side sizing term are byte-identical to previous_code — so the earlier layers that already passed stay passing. Behaviourally the sizing is unchanged from what was intended all along (getattr on a config that always defines `leverage` returned 2.0 anyway); this only makes the leverage consumption visible to the verifier. The rest of the design is as previously accepted: OI used as a ranging-regime classifier rather than a directional confirm (the orthogonal, non-price element per L42), the flatness cut derived as a rolling self-quantile rather than a fourth free parameter and therefore scale-invariant per L5, OI merged from the `metrics` Binance Vision mirror (244,790 5-minute rows for BNBUSDT from 2021-12 to 2026-05) and the recent-months `open_interest` fapi mirror so the full 365-day sandbox window is populated per L35, a one-time numpy series build with O(log n) at-or-before lookups (no lookahead, no per-bar cost that scales with history), FLAT on missing/stale OI with no price-only fallback, and double-capped sizing per L15.

Decisive net loser with no edge and a catastrophic drawdown. profit_factor 0.70 (<1.0), total_return -51.6% over ~5 years, Sharpe -0.211 (CI [-0.89, 0.44], probabilistic_sharpe 0.26), information_ratio -0.91, and avg_trade_return_pct -0.289% (NEGATIVE) with expectancy -$170/trade across 303 trades. max_drawdown is 54.2% (CI high 85.8%) — a hard-floor violation (L19) that directly contradicts the 'balanced ranging regime that mean-reverts' thesis. It loses in 5 of 6 years (2022 -14.5%, 2024 -27.0%, 2025 -22.2%) with avg_loss ($1048) exceeding avg_win ($894). The hypothesis — that gating band mean-reversion to flat-OI ranging regimes avoids the trend run-overs that kill naive fades — is falsified: the OI-flatness-gated fade still gets run over. This is the range/mean-reversion fade family, repeatedly dead here (multiple funding/OI/VWAP-gated reversion and fade candidates abandoned this session for the same negative-edge reason — the regime/confirmation gate never rescues crypto mean-reversion). A negative avg_trade_return_pct (not merely sub-floor) means the signal is wrong-signed, so reducing frequency or retuning thresholds cannot help — there is no parameter region to tune toward, and reframing the same losing mechanism onto another instrument would lose again. Per L8 (never advance when total_return ≤ 0 or PF < 1.0) and L19 (DD > 50% hard abandon), abandon at BACKTEST_REVIEW rather than spend 2 hours optimizing.

Long/short range mean-reversion on BNBUSDT.BINANCE USD-M perpetual (4-HOUR bars, 2x leverage) whose trades are gated by an OPEN-INTEREST FLATNESS regime classifier. Every bar it computes a `band_len`-period moving average and standard deviation and returns the negated band standard-score -(close - ma)/sd as the continuous signal, so the sign is the trade direction (price stretched above the mean -> negative -> sell the upper band; stretched below -> positive -> buy the lower band) and |signal| >= band_k is exactly 'closed at or beyond the band'. In parallel it computes oi_change = OI_now / OI[band_len bars ago] - 1 and classifies the market as RANGING when |oi_change| sits at or below the rolling median of its own recent |oi_change| history. Entry requires both: the OI regime is flat AND price has closed beyond the band. When OI is expanding or contracting sharply (leverage building a directional move) or the OI feed is stale/missing, the strategy takes no trade at all. Exits, first to fire: (1) an ATR stop placed beyond the entry band by stop_atr_mult x ATR(14), frozen at entry and tested intrabar — the range broke, which is the regime turning to trend; (2) the mid-band target, the moving average touched intrabar; (3) the opposite band touched intrabar; (4) an 18-bar (~3 day) time stop. Sizing risks 2% of equity at that stop distance, hard-capped at 25% x leverage (= 50% at 2x) of equity notional.