BtcPremiumIndexExtremeMeanReversionLS1H

BTC Premium Index Extreme Mean-Reversion (Single-Asset BINANCE 1H, Mark-Index Basis Dislocation Long-Short)

Single-asset long-short strategy using the BINANCE PREMIUM INDEX (mark_price - index_price spread, published live by Binance via supplementary_data['premium_index']) as a fast-mean-reverting dislocation signal. This is a GENUINELY NEW mechanism class never tested in the portfolio. The analyst has explicitly: (a) PARKED cross-venue basis class (single representative promoted as BTC×BYBIT, no more variants allowed), (b) PARKED multi-instrument portfolio strategies on Binance USD-M due to backtest engine accounting bug producing impossible -499% returns, (c) CONFIRMED that single-instrument and pair strategies on Binance USD-M produce sane metrics. This proposal complies with all three constraints: single-asset, single-venue, no cross-venue, no multi-instrument portfolio. Mechanism: the premium index = (mark_price - index_price) measures intra-venue perp-vs-spot dislocation. Mark price is determined by perp trading activity; index price is a multi-venue weighted spot average. When premium spikes positive (perp trades above index), it triggers eventual funding payment to shorts AND attracts cash-and-carry arbitrageurs (long spot + short perp) who mechanically push the spread back. The MEAN REVERSION horizon is faster than funding-rate horizon (1-8 hours vs 8h funding cycle) because spot-perp arbitrage operates continuously while funding pays only every 8h. Strategy: compute rolling 168-hour z-score of premium_index, enter contrarian (short perp when z > 2σ, long perp when z < -2σ), exit on convergence (|z| < 0.3) or 12-bar (12h) time-stop. NOT a cross-venue basis (single-venue dislocation only), NOT a funding-extreme strategy (premium is faster than funding and leads it), NOT an L/S contrarian (different signal source), NOT OI dynamics, NOT taker flow, NOT a chart pattern. Uses supplementary_data['premium_index'] which is in the documented supp_data list. Single asset BTC, single venue Binance USD-M, 1H bars (safe data, multi-year history, mid_30m_2h horizon at 12.7% near target), long-short (under-represented at 10.2%), 3 parameters total.

Implements a genuinely new mechanism class — intra-venue perp-vs-spot dislocation via the premium index — while complying with all three of the analyst's hard constraints: it is single-asset, single-venue (no cross-venue basis), and a single directional position at a time (no multi-instrument portfolio, so it avoids the parked Binance USD-M accounting bug that produced impossible returns). The edge is distinct from funding-extreme strategies because the premium reverts faster than the 8h funding cycle (continuous cash-and-carry arbitrage compresses it intra-hour), and distinct from cross-venue basis because it uses only Binance's own mark-vs-index spread. Venue is BINANCE USD-M futures (MARGIN) because the strategy goes both long and short; position_fraction keeps notional at 0.4x equity so leverage stays a true 1.0 with margin buffer and no unused-leverage gate. It fits the base template natively (should_enter returns BUY or SELL, should_exit covers), so no multi-leg machinery is needed. The premium is read from the documented premium_index supp key, parsed once into sorted arrays and looked up at-or-before each bar with a 2h staleness guard so any data gap yields no signal rather than a stale trade. Verified locally: Layer 1 static (clean), Layer 2 synthetic (all 6 scenarios pass), and a real-data check on the actual BTCUSDT premiumIndexKlines parquet (299,006 rows, continuous 2020-01 to 2026-03, no coverage gap) confirming the 168-bar rolling z-score produces 2,259 distinct |z|>=2 entry crossings over ~6 years (5.9% of bars) — hundreds of trades per walk-forward window, far above the 5-trade floor — and a z-range of +/-12.8 that justifies the stop_z regime-break guard. Three core parameters (lookback, entry_z, exit_z), 1H bars with multi-year history.

The BTC premium-index extreme mean-reversion LS is decisively unprofitable over a thorough 1819-trade sample: Sharpe -0.87 (CI [-1.68, -0.01], entirely negative), profit factor 0.92, negative expectancy (-$16/trade), Sortino -1.14, PSR 0.025, losing in 6 of 7 years (only 2023 +1.2% positive). It is also fee-dominated: commission is 19.5% of gross from over-churning (1819 trades at ~5h average hold). Funding is credited by the engine, so this is a fair test, and the mechanism is falsified: on a 1H horizon, perp premium-index extremes mark active directional flow (perp rich during pumps, cheap during dumps), so fading them is dominated by continuation — identical to the SOL premium-index MR abandoned earlier this session. Optimizing the 3 parameters over 225 trials cannot flip a negative-Sharpe, sub-1.0-PF, negative-expectancy mechanism whose edge is further consumed by fees — best-of-N over a losing, fee-eaten signal would fail holdout. This is the 'results so poor no optimization can help' case. Not revise_hypothesis: the premium-index MR mechanism has no proven sibling (its SOL variant also failed) and the failure is absence of edge plus fee fragility, not a structurally-dead target for a working mechanism. FAILURE PATTERN: intra-instrument perp premium-index extreme mean-reversion on a 1H horizon has negative edge across assets (BTC and SOL) — premium extremes mark directional flow so fading them loses to continuation, and the high-frequency churn (1819 trades, ~5h holds) lets fees consume ~20% of gross.

Single-asset long-short mean-reversion on BTCUSDT.BINANCE perp (1H) using the Binance premium index (mark_price - index_price, from supplementary_data['premium_index'] / premiumIndexKlines close) as a fast-reverting dislocation signal. Each bar it samples the most recent premium at or before the bar timestamp (2h staleness guard) into a rolling buffer and computes its z-score over a 168-bar (7-day) window. When z >= +entry_z (2.0) the perp is rich to spot so it SHORTS; when z <= -entry_z it LONGS. It exits on convergence (|z| <= exit_z = 0.3), a regime-break stop (|z| >= stop_z = 4.0), an adverse price stop (1.5%), or a 12-bar (12h) time stop. Single position at a time, sized to position_fraction (40%) of equity notional at leverage 1.0. The signal returned by calculate_signal is the premium z-score (continuous, varies every bar); entry/exit thresholds live in the hooks. Single instrument, single venue, long+short, 3 core parameters, no cross-venue and no multi-instrument portfolio.

Clean single-asset two-sided design with correct premium-index data plumbing; large 1819-trade sample thoroughly tests the mechanism

Low drawdown (17.9%) from the dollar-neutral-ish small-size construction

Decisively unprofitable: Sharpe -0.87 (CI fully negative), profit factor 0.92, negative expectancy (-$16/trade), losing in 6 of 7 years

Fee-dominated: commission 19.5% of gross from over-churning (1819 trades, ~5h holds)

Falsified mechanism: 1H premium-index extremes mark directional flow, so fading them is dominated by continuation — same failure as the abandoned SOL premium-index MR

13 max consecutive losses; PSR 0.025

BtcPremiumIndexExtremeMeanReversionLS1H was a carefully constraint-compliant 'new mechanism class' — single-asset, single-venue BTC, fading premium-index (mark-vs-index) z-score extremes on the thesis that continuous spot-perp arbitrage reverts dislocations faster than funding. Across a thorough 1,819-trade sample it lost decisively: a -0.87 Sharpe with a fully-negative confidence interval, sub-1.0 profit factor, negative expectancy, losses in six of seven years, and commissions eating 19.5% of gross from ~5-hour churn. The analyst found the premise falsified — 1H premium extremes mark directional flow, so fading them loses to continuation, exactly as the earlier SOL premium-index variant did — and abandoned it before optimization. It ended after one iteration as abandoned, never advancing to optimization or risk review.

Intra-instrument perp premium-index extreme mean-reversion on a 1H horizon has negative edge across assets (it failed on both BTC and SOL): premium extremes mark directional flow rather than dislocations that revert, so fading them loses to continuation, and high-frequency churn lets fees consume a fifth of gross — a 'new' signal source is not a new edge if its only sibling already failed.

It was abandoned at the pre-optimization backtest-review gate (verdict: abandon) because the mechanism was falsified — on a 1H horizon premium-index extremes mark active directional flow, so fading them loses to continuation, and the high-frequency churn lets fees eat ~20% of gross — making it unfixable by tuning three parameters, so optimization and all later stages were never reached.

A single-asset long-short mean-reversion strategy on BTCUSDT Binance USD-M perps (1H bars) that faded extremes in the Binance premium index (mark-minus-index basis) — shorting the perp when the rolling 168-hour z-score exceeded +2σ and going long when below -2σ, exiting on convergence, a regime-break stop, a price stop, or a 12-hour time-stop — on the thesis that spot-perp arbitrage compresses premium dislocations faster than the 8h funding cycle.

Over 1,279 data-days and a large 1,819-trade sample it was decisively unprofitable: total return -11.8%, Sharpe -0.87 (CI [-1.68, -0.01], entirely negative), profit factor 0.92, negative expectancy of -$16/trade, Sortino -1.14, PSR 0.025, losing in 6 of 7 years, with commissions consuming 19.5% of gross from over-churning (~5h average holds) despite a low 17.9% max drawdown.