EthBtcRatioDualTimeframeMomentumConfluenceLS

ETH/BTC Ratio Dual-Timeframe Momentum Confluence — Dollar-Neutral Pairs Spread, Long-Short, 4H Primary + 1D Confirm, Flat on Disagreement (ETHUSDT.BINANCE + BTCUSDT.BINANCE, 3-Parameter)

A DOLLAR-NEUTRAL PAIRS-SPREAD strategy that applies the factory's ONE proven survivor mechanism — dual-timeframe momentum confluence (BTC Dual-Timeframe Momentum Confluence, Sharpe 1.99, paper_stage) — to the ETH/BTC RATIO instead of a single directional instrument. This is a deliberate structural fix for the #1 killer in the failure log: every single-name OHLCV trend follower (ETH vol-targeted, BNB pullback, XRP swing, DOGE fat-tail, LINK efficiency-ratio) dies because its headline return is just 2020-2021 crypto BULL BETA that goes negative every year after, so the walk-forward OOS and 15-day holdout in the recent regime lose. Trading the ETH/BTC RATIO (long one leg, short the other, equal notional) REMOVES crypto-wide beta by construction — the PnL cannot be a bull-market artifact because a dollar-neutral spread is ~flat to the market. What remains is the persistent, well-documented ETH-vs-BTC relative-strength cycle (capital rotates between the two majors in multi-week regimes). The proven confluence gate (require a FAST 4H and SLOW 1D momentum read on the ratio to AGREE, stand flat on disagreement) is precisely what separated the survivor from the outlier-driven single-signal trend followers. This is momentum/CONTINUATION on the ratio — explicitly NOT ratio mean-reversion (which died in L53) and NOT a cross-sectional rank rotation over many names (which died in L52); it is a single, dedicated, beta-neutral pair with the one gate that has actually worked here. Fills the under-weight long_short direction bucket and keeps parameters to 3 to resist overfitting.

Iteration 2 — minimal fix for the reported Layer-2 failure only. The six synthetic scenarios all died with `_ts_of() takes 1 positional argument but 2 were given`: the Layer-2 harness (`strategy_verifier._StrategyProxy`) copies every callable attribute onto a plain wrapper object via `types.MethodType(val.__func__ or val, wrapper)`, which re-binds a @staticmethod as an instance method and injects `self` as its first argument. I therefore converted the two @staticmethod helpers — `_ts_of` and `_round_qty` (the latter would have failed the same way at Layer 3) — into ordinary instance methods; every call site already used `self.`, so no other line changed. Signal logic, entry/exit gates, sizing, imports and config are byte-identical to the version that passed Layer 1, so no earlier check can regress. The underlying evidence is unchanged from iteration 1: dry-run on the real 4H catalog data for both legs (14,460 co-timestamped bars, 2019-12 to 2026-08) gives 59 trades/year, +0.47% average per round trip net of the 0.20% two-leg taker cost, 4.44x on per-leg notional, bar-level Sharpe ~0.72 and ~30% max drawdown at 1.0x per-leg notional (hence the 0.8x sizing), with a parameter plateau rather than a spike around fast 18 / slow 120 / entry_z 0.4.

Dollar-neutral ETH/BTC ratio pairs-momentum with no robust, scalable edge — not worth 2 hours of optimization (same result as this run's prior review of the identical backtest). Sharpe is 0.444 with bootstrap CI [-0.29, 1.23] straddling zero and profit_factor 1.057 (below the 1.2 meaningful-edge line). It is capacity-blocked: impact_cost_pct 26.7% (impact eats a quarter of gross) with capacity_usd only $1.4M, so the edge is real only at toy scale — a promotion blocker on its own. The hypothesis leans on L82's proven dual-TF confluence, but L82's survivable variant is a single-name directional Binance-USD-M major with Sharpe >~1.3 / PF >2.0 / DD <20%; this is a pairs spread (a different object, the L45 market-neutral 0/84 family) at Sharpe 0.44 / PF 1.06 / DD 29.8% (CI to 92.8%), failing every threshold. The edge is also decaying: the +171% headline is front-loaded into 2020-2021 while the recent regime is negative (2026 -11.9%, rolling Sharpe to ~-8), so the last-20% holdout sits in a negative window and would likely fail the OOS/holdout gate. benchmark_meaningful is correctly false, and the absolute risk-adjusted metrics are marginal-to-negative. Base Sharpe 0.44 is the ceiling; the 3-phase deflation only lowers it. Failure pattern: no_edge/capacity market-neutral pairs (L45) below the L82 survivable dual-TF bar.

Dollar-neutral ETH/BTC pairs-spread momentum on Binance USD-M 4-HOUR bars. Each bar the strategy commits a log(ETH/BTC) sample only for timestamps where BOTH legs have a bar (single synced series — the two legs can never drift out of index alignment), and takes two momentum reads on that spread: a FAST 18-bar (~3 day, the 4H read) and a SLOW 120-bar (~20 day, the daily read). Each is divided by the spread's own realized volatility scaled by sqrt(lookback), producing scale-invariant t-statistics rather than fixed absolute thresholds. The continuous signal returned every bar is the mean of the two z-scores. Entry requires CONFLUENCE: both reads must share a sign and both must exceed entry_z = 0.4 — positive means LONG ETHUSDT / SHORT BTCUSDT at equal USD notional, negative means the reverse. The pair is flattened as soon as either read turns against the position (flat on disagreement), or if the ratio moves 6% against the entry (spread stop). Each leg is sized at 40% of equity times 2x leverage (~0.8x equity notional per leg, ~1.6x gross), so crypto-wide beta cancels between the legs and the P&L is ETH's relative strength versus BTC.

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.

Genuinely dollar-neutral by construction (equal-notional opposite legs, synced timestamps) — the +171% is not bull beta, positive in several non-bull years

Healthy two-sided trade count (782); clean multi-timeframe confluence implementation

No significant edge: Sharpe 0.444 with CI [-0.29, 1.23] straddling zero, profit_factor 1.057 (below the 1.2 meaningful-edge line)

Capacity wall: impact_cost_pct 26.7% and capacity_usd only $1.4M — edge exists only at toy scale, a promotion blocker on its own

Fails the L82 survivable dual-TF bar (needs single-name major, Sharpe >~1.3 / PF >2.0 / DD <20%); this is a pairs spread at Sharpe 0.44 / PF 1.06 / DD 29.8% (CI to 92.8%) — the L45 market-neutral 0/84 prior applies

Decaying edge: +171% front-loaded into 2020/2021, recent regime negative (2026 -11.9%, rolling Sharpe to ~-8), so the last-20% holdout is negative and would likely fail the OOS/holdout gate

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['entry_z', 'fast_bars', 'slow_bars', 'vol_window', 'min_notional', 'per_leg_frac', 'stop_spread_pct'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: EthBtcRatioDualTimeframeMomentumConfluenceLS._ts_of() takes 1 positional argument but 2 were given (bar timestamp: 1735689600000) - steady_downtrend: TypeError: EthBtcRatioDualTimeframeMomentumConfluenceLS._ts_of() takes 1 positional argument but 2 were given (bar timestamp: 1735689600000) - flat_ranging: TypeError: EthBtcRatioDualTimeframeMomentumConfluenceLS._ts_of() takes 1 positional argument but 2 were given (bar timestamp: 1735689600000) - volatility_spike: TypeError: EthBtcRatioDualTimeframeMomentumConfluenceLS._ts_of() takes 1 positional argument but 2 were given (bar timestamp: 1735689600000) - zero_volume: TypeError: EthBtcRatioDualTimeframeMomentumConfluenceLS._ts_of() takes 1 positional argument but 2 were given (bar timestamp: 1735689600000) - price_gap: TypeError: EthBtcRatioDualTimeframeMomentumConfluenceLS._ts_of() takes 1 positional argument but 2 were given (bar timestamp: 1735689600000)