BtcCmUsdmPerpBasisReversionNeutral

BTC Inverse Perp vs Linear Perp Basis Mean-Reversion (COIN-M vs USD-M, 4H, Long-Short, Cross-Venue)

Market-neutral cross-venue mean-reversion strategy exploiting transient price divergences between BTCUSD_PERP.BINANCE_CM (COIN-M inverse perpetual, BTC-margined) and BTCUSDT.BINANCE (USD-M linear perpetual, USDT-margined). Both contracts reference the same BTC/USD spot price but trade in structurally different venues: COIN-M is dominated by BTC holders willing to take BTC-denominated P&L (and is non-linear in BTC), while USD-M is the global dollar-quote benchmark used by most institutional and retail traders. During volatility spikes, liquidation cascades on one venue, or asymmetric funding cycles, the implied USD prices of the two contracts can diverge by 15–60 bps for 1–3 four-hour bars before cross-venue arbitrage flows close the gap. Strategy: compute pct_spread = (cm_close - um_close) / um_close on each 4H close, maintain a rolling 50-bar z-score, and when |z| > 2.0 AND |pct_spread| > 0.0018 (above combined ~0.16% RT fees of CM 0.06% + USD-M 0.10%), long the cheaper contract and short the richer one in equal USD-notional terms. Exit when |z| < 0.3 (mean reversion) or after 6 bars (24h) time-stop. This is NOT a classical chart pattern (avoided after 8+ pattern hypotheses failed), NOT directional, and targets the most-empty quotas: BINANCE_CM (0.3% vs 5% target — largest relative gap), cross-venue (4.7% vs 15%), long-short (9.2% vs target), and pairs scope simultaneously. Uses ONLY 4H bars on BTC — the safest data combination per the failure analysis (avoids the 8-HOUR data-unavailable trap that killed 7+ recent hypotheses).

Implements the hypothesis exactly: cross-venue, market-neutral, two-leg, 4H BTC only, with two tunable knobs (entry_z, min_spread) whose clamps are mirrored in _param_bounds; the COIN-M leg is primary so the engine's inverse contract handling and BTC-settled accounting apply to it, and the linear leg is derived from the same USD notional so leg sizing cannot leave residual directional exposure. I verified the data plumbing before coding: both legs have 4H bars on the identical 00/04/08/12/16/20 UTC grid with 12,635 overlapping bars (2020-08 to 2026-05), so the base class's cross-leg alignment barrier is satisfiable on every primary bar and this will not stall the way mismatched-grid pairs do. One honest and important caveat from measuring the actual spread series: the CM-vs-USD-M divergence is far tighter than the hypothesis assumes - standard deviation 7.1 bps, p99 19.5 bps, and |spread| reaches the stated 18 bps on only 1.8% of bars - so at the default gate the strategy fires roughly 8 times a year and each convergence captures at most ~18 bps against a two-leg round-trip cost of about 20 bps (CM 0.10% + USD-M 0.10%), i.e. the mechanism is fee-negative at the specified threshold and only becomes viable in the rare 30 bps-plus dislocations, which is why min_spread is the tunable knob with an upper bound of 60 bps. Loosening the gate to |z| >= 2 alone would give ~110 events a year but on a ~7 bps median divergence that is guaranteed fee death, so I kept the hypothesis's fee gate rather than manufacturing trade count. All state is incremental (a 50-element spread deque with running sum and sum-of-squares), per-bar cost is O(1), and leverage stays 1.0 with 0.35x equity per leg so combined margin is ~0.7x equity - a genuinely neutral book with no path to the -100% leg-sizing blowups that plague delta-neutral pairs.

Negative-expectancy cross-venue basis reversion in the zero-survivor L57/L44 family — not worth 2 hours of optimization. Over 74 trades in 5.8 years: profit_factor 0.953, avg_trade_return_pct ~0.0005% (net negative), expectancy -$28.3/trade, total_return -18.6%, Sharpe -0.26 with CI [-1.08, 0.56] and PSR 0.268, losing in every year 2020-2026, with a 34.8% drawdown on a supposedly market-neutral book. The mechanism is structurally weak: a same-asset BTC COIN-M-vs-USD-M basis is a few bps, so the 0.18% min_spread fee gate rarely fires legitimately (hence only 74 trades), and a persistent >0.18% divergence between two contracts on the same reference is far more likely a bar-alignment/stale-leg artifact than a real premium — the captured 'edge' is phantom and still loses net of the two-leg round-trip. The book is cleanly built and genuinely delta-neutral (this is not a code defect), but the economics are absent: negative per-trade, PF below 1, negative every year, on a spread too small to clear its own cost. No tuning of entry_z/min_spread converts that into an edge. This is the same cross-venue basis/spread family (SOL/BTC spot-perp and CM-USDM) abandoned repeatedly this session; benchmark_meaningful is correctly false. Failure pattern: no_edge/fee_edge cross-venue basis reversion, same-asset spread too small / artifact-prone (L57/L44).

Market-neutral cross-venue mean reversion between the two Binance BTC perpetuals: BTCUSD_PERP.BINANCE_CM (COIN-M inverse, BTC-margined, primary/clock leg) and BTCUSDT.BINANCE (USD-M linear), both on 4-hour bars. Each aligned bar it computes pct_spread = (cm_close - um_close) / um_close and its 50-bar rolling z-score. When |z| >= 2.0 AND |pct_spread| >= 18 bps it shorts the richer contract and longs the cheaper one at equal USD notional - the COIN-M integer contract count is chosen first and the USD-M BTC quantity is derived from that same notional, so net delta is ~0 by construction. It exits when the divergence closes (|z| <= 0.3), on a 6-bar (24h) time stop, or immediately if only one leg ends up filled (a half-open book is naked delta). Both legs open and close on the same bar, so the book is managed in calculate_signal and the base class's single-leg entry path is inert.

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.

Analyst to confirm whether any min_spread is simultaneously fee-positive and frequent enough for the >=10-trade holdout; the spread measurement suggests not — abandon as fee-dead rather than optimize.

Fee viability: min_spread default 18 bps is BELOW the ~20 bps two-leg cross-venue round-trip, so the reversion cannot pay fees — fee-negative at defaults (sandbox avg_trade_return_pct ~0). It clears the literal 0.15% single-leg floor of check 8 and is tunable to 0.60%, so not a hard-wired breach, but the fee-viable region (30 bps+) is too rare (~1.8% of bars reach even 18 bps) to populate the holdout. Classic L6/L16 fee-graveyard-vs-too-rare bind.

No fix required; if promoted, note the ~spread% residual BTC delta and inverse convexity.

Delta neutrality is USD-NOTIONAL-matched. For the INVERSE CM leg BTC delta is notional/price, so the legs' BTC deltas differ by the spread (~7 bps) plus inverse convexity on large moves — small residual, not exactly zero. The large symmetric per-leg PnL swings (avg_win ~$1180 vs avg_loss ~$1161) are per-leg accounting on the hedged pair netting to the small aggregate — a working hedge, not a neutrality bug.

Cleanly built and genuinely delta-neutral (contract-count-first sizing, contemporaneous-bar barrier, half-open-book flatten); low fee drag (6.1% of gross); correct inverse handling

Negative expectancy: PF 0.953, avg_trade_return_pct ~0.0005%, expectancy -$28.3/trade, total_return -18.6%, Sharpe -0.26 (CI [-1.08, 0.56]), PSR 0.268 — negative in every year 2020-2026

Traded spread is a few bps on the same underlying; the 0.18% gate rarely fires (74 trades/5.8y) and a persistent >0.18% CM-vs-USDM divergence is more likely a bar-alignment artifact than real premium

max_drawdown 34.8% on a market-neutral book; normal-vol regime -21.3%

L57/L44 cross-venue basis/spread family (zero survivors); benchmark_meaningful correctly false

Verification failed (Layer 4 — QA review) [class=premise_falsified]: - The hypothesis's core premise is empirically false: the exploitable cross-venue divergence does not exist. The thesis claims the COIN-M and USD-M BTC perps diverge by 15-60 bps, but the developer measured the real spread and found std 7.1 bps, p99 19.5 bps, and |spread| reaches the stated 18 bps on only 1.8% of bars — the two Binance BTC perps are held within ~7 bps by arbitrage and essentially never reach 30-60 bps. Against a two-leg round-trip of ~0.20% (CM ~0.10% + USD-M ~0.10%; the hypothesis's '0.16%' understates CM), the max capturable divergence (~18-20 bps) barely covers costs, so the mechanism is fee-negative at any supportable threshold (developer states this outright). Sandbox: total_return -18.62%, PF 0.963 (<1), avg_trade_return_pct 0.00052% (~zero), Sharpe -0.216, end_unrealized -17.8% (large open pair loss, aggravated by inverse-vs-linear convexity mismatch), 74 trades, net loser. Re-coding cannot make the perps diverge more than arb allows. - Secondary: the hedge matches USD notional at entry, but the COIN-M leg is INVERSE (convex) and USD-M is LINEAR, so the pair's delta drifts as BTC moves over the hold, leaving residual directional exposure (the -17.8% open-pair loss and benchmark_correlation -0.76 hint at leakage). Not the primary failure (the premise is dead regardless), but a true neutral book would re-hedge the inverse leg's dollar delta.