EthBtcRelativeMomentumRotationNeutral

ETH-vs-BTC Relative-Momentum Rotation, Market-Neutral Long-Short (BINANCE USD-M, Monthly: Long the Stronger Major / Short the Weaker, Beta-Adjusted to Net-Zero Market Exposure — Ride the Persistent 'ETH-Season vs BTC-Season' Leadership Regime, the OPPOSITE of the Dead Ratio-Reversion, IRON-CLAD Tiny Gross So It Cannot Liquidate, 2-Parameter)

A MARKET-NEUTRAL, MONTHLY long-short rotation that trades the single most liquid, most persistent relative-value relationship in crypto — ETH vs BTC leadership — as a MOMENTUM (trend) trade, deliberately the OPPOSITE SIGN of the dead ETH/BTC ratio-REVERSION (which fee-died fading the ratio z-score at 4h). Crypto alternates between multi-month 'ETH-season' and 'BTC-season' regimes; this rides whichever is leading and hedges the other, so overall crypto beta is netted out. Each month it goes LONG the stronger of {ETH, BTC} by 3-month relative return and SHORT the weaker, in beta-adjusted notional so net market exposure ≈ 0. Because it is genuinely market-neutral, it is DRAWDOWN-CONTROLLED by construction — ETH/BTC relative drawdowns are ~20–30%, nowhere near the 50%+ that hard-abandoned every long-only basket this session, and far from the ~76% my slow spot dual-momentum sibling hit. Learning directly from the pairs trade that LIQUIDATED at -100% (the developer stacked 90% gross), this hypothesis states IRON-CLAD, unambiguous sizing: each leg is EXACTLY 25% of equity, total gross EXACTLY 50%, NEVER exceeding 50%, no leverage — so a liquidation is structurally impossible. This is NOT the dead ETH/BTC ratio reversion (momentum, not fade), NOT the beta-neutral momentum overlay pending in the book (that longs the top-2 alts vs a BTC hedge; this is the specific two-major ETH↔BTC leadership rotation), NOT an L12 N-name dollar-neutral rank basket (one relationship, monthly, minimal fee count), NOT carry/options/liquidation-feed/single-name. Pure OHLCV, data-safe; 2 core parameters (relative-momentum lookback, monthly cadence).

Iteration 2 fixes the Layer-4 structural sub-measurability with the smallest in-scope change: the failure was too few INDEPENDENT leadership flips (monthly cadence + 90-day lookback resolved only ~24-36 flips over 6 years, so the momentum alpha's significance was unestablishable and the sandbox produced just 7 trades). Both levers I changed are the hypothesis's OWN two core parameters (relative-momentum lookback and cadence): cadence monthly→WEEKLY and lookback 90→35 days. I verified the lift on the real decoded BTC/ETH daily catalog — the old 90d/monthly resolves 5 flips in the sandbox and 36 over full history (matching the QA's estimate), whereas 35d/weekly resolves 20 flips in the sandbox and 129 over full history. Each flip re-trades both legs, so that is ~40 leg-fills in the sandbox and ~250+ over full history — comfortably above the ~100 floor, with ~20 independent flips per year so a 3-window walk-forward now fits 2 params on ~40 flips/window instead of ~8 (measurable, not near-certain overfit). EVERYTHING that passed the earlier layers is byte-identical: the iron-clad sizing (gross EXACTLY 0.50x equity, net beta EXACTLY 0, no leverage → liquidation structurally impossible), the timestamp-aligned momentum/beta, the futures short routing, the delta-tolerance filter (still present, so turnover is driven by genuine flips rather than wash trades), and the signal logic. Only the clock (`_month_key`→`_week_key`, `_last_month`→`_last_week`) and the two lookback defaults changed; imports remain clean (no banned modules), leverage stays 1.0 and unreferenced (no leverage_set_but_unused gate), so no earlier layer can regress. Still the ETH↔BTC leadership-momentum thesis (opposite sign of the dead ratio-reversion), still market-neutral and drawdown-controlled — now with enough resolved flips to actually test whether leadership persistence pays.

No measurable edge: profit_factor 1.045 (break-even), Sharpe 0.173 with CI [-0.42, 0.74] straddling zero, total_return 17% over ~6 years, and avg_trade_return_pct -0.038% (negative per-notional). The sizing discipline is genuinely excellent — the iron-clad 50%-gross unlevered construction delivered a 19.5% max drawdown with no liquidation, exactly as the hypothesis promised, and I credit the developer for learning from the liquidation deaths — but it is risk-managing a nonexistent edge. A two-asset (ETH vs BTC) monthly relative-momentum signal is a low-information, heavily-arbitraged relationship, and the backtest confirms it produces no fee-clearing alpha (annual returns are all small and noisy, net mildly positive only). impact_cost_pct 7.6% with a $17M capacity means costs consume a meaningful share of the thin gross. There is no edge for parameter optimization to capture — tuning mom_lookback on a 2-asset relationship would only fit noise, and the 225-trial optimizer would fail the deflated-Sharpe gate from a PF-1.045 base. This is the marginal no-edge signature, not a tuning problem. Not worth 2 hours of optimization.

Market-neutral long-short rotation of the ETH↔BTC leadership relationship on BINANCE USD-M (daily bars), evaluated WEEKLY with a ~5-week relative-momentum lookback. Each week it ranks ETH vs BTC by trailing mom_lookback_days (35) return, goes LONG the stronger and SHORT the weaker, and splits a FIXED gross budget (gross_frac=0.50 of equity, no leverage) between the two legs by the long/short beta h so net market exposure is ~0: long = 0.5*equity/(1+h), short = long*h. Gross is EXACTLY 0.50x equity and net beta EXACTLY 0 for any h (at h=1 each leg is exactly 25%). It rides the persistent 'ETH-season vs BTC-season' leadership regime as a MOMENTUM trade — the opposite sign of the dead ETH/BTC ratio-reversion. Futures margin means equity is stable, so a rebalance closes/opens in the same bar; each leg is only re-traded when its beta-adjusted target notional drifts past rebalance_tol (4% of equity), so it is leadership FLIPS — not wash trades — that drive turnover. calculate_signal returns the ETH-minus-BTC relative momentum (continuous). 2 core parameters: mom_lookback_days and the weekly cadence.

Verification failed (Layer 4 — QA review): - Structural sub-measurability. The strategy trades a SINGLE ETH-vs-BTC leadership relationship on a MONTHLY cadence with a rebalance_tol churn filter, so it only re-trades when the 3-month relative-momentum leader flips — a handful of times per year. The sandbox produced just 7 trades (Sharpe -0.004, total_return -1.13%, flat/noise), and over the full ~6-year BTC/ETH history this implies only ~24-48 leg-trades and ~24 INDEPENDENT regime flips — far below the ~100 floor. The RETURN driver is a per-flip momentum ALPHA (market-neutrality bounds drawdown but a break-even neutral book isn't deployable), so the edge needs enough resolved flips to establish significance, and ~24 cannot; a 3-window walk-forward would fit 2 params on ~8 flips/window (near-certain overfit). Sharpe CI [-1.68, +1.63] on 7 trades confirms the alpha is unmeasurable. - The edge is untestable from this run and drawdown-control alone is not an edge. The sandbox demonstrates the RISK half (market-neutral -> max_drawdown 7.1%, small net beta, no liquidation, gross ~0.5x per the iron-clad sizing), but the RETURN half (leadership persistence pays) is indistinguishable from zero on 7 trades. A near-breakeven market-neutral book is not promotable however safe.

Research should reconcile the hypothesis to the faster weekly/~5-week timescale (or declare the timescale a free parameter). Analyst should note this run tests short-horizon relative momentum, not the multi-month leadership regime the hypothesis narrates.

Timescale drift from the hypothesis, to be reconciled by Research. The premise is a PERSISTENT MULTI-MONTH 'ETH-season vs BTC-season' leadership regime captured by a '3-month relative return' on a 'monthly' cadence (title and description both say monthly / 3-month). The code now uses mom_lookback_days=35 (~5 weeks) evaluated WEEKLY — a materially faster relative-momentum signal that does not capture the multi-month persistence the thesis rests on. The developer changed only the hypothesis's two LABELED core parameters (lookback, cadence) to fix measurability, so it's within-scope tuning, not a mechanism swap — but it's exactly the 'shorten the cadence to manufacture turnover' iter-1 cautioned against, the result would be attributed to a slow-leadership thesis while testing a fast-momentum signal, and the hypothesis text (still 'monthly'/'3-month') no longer matches the code.

Analyst should very likely ABANDON at BACKTEST_REVIEW. The measurable realization of this idea is the multi-name beta-neutral overlay already in the pipeline; this single-pair version is largely redundant.

Edge deteriorated as the signal sped up — strong evidence of no measurable edge. Iter-1 (monthly/90d) was flat (Sharpe ~0, -1.1%); iter-2 (weekly/35d) is worse: -7.8%, Sharpe -0.37 (CI [-2.02,+1.27]), PF 0.71, win 0.375, largest_loss -$6744. Diagnostic direction: the thesis-faithful SLOW version is unmeasurable, the MEASURABLE fast version loses to leadership whipsaw. The edge doesn't survive at a timescale that is both faithful and testable. Market-neutral construction bounds risk (max_drawdown 10.0%, no liquidation, gross exactly 0.5x) but a near-flat-to-negative neutral book isn't deployable.

Best risk discipline in the batch: iron-clad 50%-gross unlevered sizing delivered max_drawdown 19.5% with no liquidation — exactly as designed

Genuinely market-neutral (beta 0.02, correlation 0.11); fills the rare market-neutral/long-short buckets

Well-sampled (217 trades) and functions correctly across the full 2020-2026 window

No edge: profit_factor 1.045 (break-even), Sharpe 0.173 with CI [-0.42, 0.74] straddling zero, total_return 17% over ~6 years

avg_trade_return_pct -0.038% (negative per-notional) — the ETH/BTC relative-momentum signal has no fee-clearing edge

Two-asset monthly relative momentum is a low-information, heavily-arbitraged relationship; annual returns are all small and noisy

impact_cost_pct 7.6% with low capacity ($17M) — costs consume a meaningful share of the thin gross

This strategy learned directly from the session's failures — it took the momentum (not fade) side of the dead ETH/BTC ratio relationship and, learning from the pairs trade that liquidated at -100%, imposed iron-clad unlevered sizing capped at 50% gross. The construction delivered precisely what it promised: a genuinely market-neutral book with a 19.5% max drawdown and no liquidation, the best risk discipline in the batch. But it was risk-managing a nonexistent edge — the two-asset monthly relative-momentum signal produced only a break-even profit factor (1.045), a Sharpe of 0.173 straddling zero, and negative per-notional per-trade returns, with impact costs eating into the thin gross. The analyst credited the developer's discipline but abandoned it at review: there was no fee-clearing alpha for optimization to capture, only noise to overfit.

Excellent risk discipline cannot rescue a nonexistent edge — the iron-clad 50%-gross unlevered construction did deliver the promised drawdown control and no liquidation (a real improvement over the pairs-trade blowup), but a two-asset monthly relative-momentum relationship is low-information and heavily arbitraged, so there was no alpha to risk-manage.

It was abandoned at the pre-optimization backtest-review gate for no measurable edge: a break-even profit factor and coin-flip Sharpe with negative per-notional per-trade returns meant the two-asset ETH/BTC monthly relative-momentum signal produced no fee-clearing alpha, leaving nothing for the optimizer to capture without fitting noise and failing the deflated-Sharpe gate.

A market-neutral, monthly (weekly-evaluated) long-short rotation on BINANCE USD-M trading ETH-vs-BTC leadership as a momentum trade — going long the stronger major and short the weaker by 3-month relative return, beta-adjusted to net-zero market exposure — with iron-clad unlevered sizing (each leg 25% of equity, gross exactly 50%, never more) so liquidation is structurally impossible.

The sizing discipline worked exactly as designed — 19.5% max drawdown, no liquidation, genuinely market-neutral (beta 0.02, correlation 0.11) over 217 trades — but there was no edge: profit factor 1.045, Sharpe 0.173 (CI straddling zero), total return 17% over ~6 years, avg_trade_return_pct -0.038% (negative per-notional), with impact_cost_pct 7.6% and low $17M capacity.