Skip to content
All strategies

Strategies

ETH/BTC Relative-Strength Regime Momentum, Dollar-Neutral Long-Short (ETHUSDT.BINANCE vs BTCUSDT.BINANCE USD-M, DAILY bars, slow 21-day ratio momentum with deadband, low-turnover, 3-parameter)

Outcome: Abandoned

EthBtcRegimeMomentumSlowDollarNeutralLS

Outcome Summary

EthBtcRegimeMomentumSlowDollarNeutralLS was a careful redesign that traded the ETH/BTC ratio regime dollar-neutral on daily bars, targeting the exact impact/capacity failure that killed an earlier SOL/ETH ratio attempt. On that front it succeeded cleanly — impact cost dropped from 66.8% to 6.4%, capacity rose from ~$224k to ~$24M, and trade count fell from 1612 to 150 while per-trade return cleared the fee floor. But the surviving edge was thin and lopsided: profit factor 1.049, Sharpe 0.272 with a CI straddling zero, a 38% drawdown, and returns that were negative in the dominant 'normal' volatility regime and driven only by a few violent rotation months. The analyst abandoned it at the pre-optimization backtest-review gate, judging the 'persistent signable regime' premise falsified and concluding that optimizing three parameters on a right-tail artifact in a 0/84 strategy class could not reach significance; no optimization, analyst-verdict, or risk-review stages were run.

Strategy report

Hypothesis

A DOLLAR-NEUTRAL, TWO-INSTRUMENT long-short strategy that trades the persistent multi-week REGIME of the ETH/BTC ratio, on DAILY bars, holding the strong leg long and the weak leg short. This is NOT the dual-timeframe confluence recipe and NOT a symbol-swap of the promoted winner: it is a deliberate, evidence-driven re-engineering of the ONE dimension that killed the prior SOL/ETH ratio attempt. That death (profit_factor 1.0108, impact_cost_pct 66.78%, capacity only ~$224k) was a MARKET-IMPACT / capacity failure……Show moreShow less

A DOLLAR-NEUTRAL, TWO-INSTRUMENT long-short strategy that trades the persistent multi-week REGIME of the ETH/BTC ratio, on DAILY bars, holding the strong leg long and the weak leg short. This is NOT the dual-timeframe confluence recipe and NOT a symbol-swap of the promoted winner: it is a deliberate, evidence-driven re-engineering of the ONE dimension that killed the prior SOL/ETH ratio attempt. That death (profit_factor 1.0108, impact_cost_pct 66.78%, capacity only ~$224k) was a MARKET-IMPACT / capacity failure from over-trading a thin leg (SOL) on 4H+1D confluence (1612 trades over 6y). This design attacks that root cause three ways: (1) both legs are the two DEEPEST instruments in all of crypto (ETH and BTC USD-M) so per-fill impact is an order of magnitude smaller than SOL; (2) a SLOW 21-day daily momentum signal with a deadband cuts turnover to ~1 rotation every ~2 weeks (target ~120-150 round-trips over 6y, not 1600), so fee+impact drag per unit of capture collapses; (3) the traded object is the ETH/BTC RATIO regime, a real structural phenomenon (capital rotates between BTC-dominance and alt-risk-on in persistent multi-week cycles), not a generic price trend gated by a filter (L92) nor a re-skin of the volume-confirmed momentum winner (L91). Because it is dollar-neutral it also removes crypto-beta, so it is not another long-only trend follower (long_only is 85.7% of the corpus and 1% survival).

Implements the hypothesis exactly: the traded object is the ETH/BTC RATIO regime, on DAILY bars, with a slow 21-day momentum read plus a dead band, held long-strong / short-weak in two equal-USD legs on the two deepest perps in crypto. The three stated tunables are mom_lookback (21), entry_threshold (dead band) and exit_band_frac (hysteresis), with clamps mirrored in _param_bounds. The three impact/capacity fixes the hypothesis names are structural in the code, not incidental: (1) both legs are ETH and BTC USD-M,……Show moreShow less

Implements the hypothesis exactly: the traded object is the ETH/BTC RATIO regime, on DAILY bars, with a slow 21-day momentum read plus a dead band, held long-strong / short-weak in two equal-USD legs on the two deepest perps in crypto. The three stated tunables are mom_lookback (21), entry_threshold (dead band) and exit_band_frac (hysteresis), with clamps mirrored in _param_bounds. The three impact/capacity fixes the hypothesis names are structural in the code, not incidental: (1) both legs are ETH and BTC USD-M, the deepest books available, so per-fill impact is an order of magnitude below the SOL leg that killed the prior attempt; (2) a 21-day daily signal with an entry dead band AND a far-side exit band means a full round trip requires the ratio z to swing from +0.70 to -0.21 (or the mirror), which on daily data is a multi-week event — target ~100-150 round-trips over six years rather than 1600; (3) dollar neutrality is enforced by construction (one notional N, both legs sized to it, both-or-neither submission, hedge-broken flatten), which is also the guard against the -100% failure mode of mis-sized neutral books. Venue is futures because the strategy shorts one leg every trade; leverage=2.0 is genuinely consumed — the vol-targeting sizer asks for equity * 0.008 / ratio_vol per leg, which exceeds 1x equity whenever the ETH/BTC daily ratio vol falls below 0.8%, and the cap that grants it is max_notional_frac * equity * leverage. Position sizing is volatility-targeted rather than fixed-fraction so the book's daily P&L vol stays near 0.8% of equity across regimes, capping compounding blow-ups.

Thin, right-tail/regime-dependent edge that will not survive the 3-phase deflation, not worth 2 hours of optimization. On the full 6.7-year / 150-trade sample profit_factor is 1.0495 (in the cost-drag band), Sharpe 0.272 with CI [-0.44, 1.05] straddling zero, CAGR 3.55% against a 38% drawdown (2.8yr underwater). The design's central premise — that the ETH/BTC ratio is a PERSISTENT, signable multi-week regime — is not borne out: the book is NEGATIVE in the dominant 'normal' vol tercile (-24.65%, Sharpe -0.17, 803……Show moreShow less

Thin, right-tail/regime-dependent edge that will not survive the 3-phase deflation, not worth 2 hours of optimization. On the full 6.7-year / 150-trade sample profit_factor is 1.0495 (in the cost-drag band), Sharpe 0.272 with CI [-0.44, 1.05] straddling zero, CAGR 3.55% against a 38% drawdown (2.8yr underwater). The design's central premise — that the ETH/BTC ratio is a PERSISTENT, signable multi-week regime — is not borne out: the book is NEGATIVE in the dominant 'normal' vol tercile (-24.65%, Sharpe -0.17, 803 days) and positive only in calm/stressed thirds, with returns concentrated in a few violent rotation months and losses in 2023 (-10%) and 2024 (-14%). The impact/capacity redesign genuinely fixed the prior death cause (impact 66.8%->6.4%, capacity $224k->$24M), but the residual edge is a right-tail artifact, not a persistent regime, so QA's open question resolves against the strategy. This is the dollar-neutral major-perp long-short class (L45, 0/84) whose PF~1.0 / insignificant-Sharpe base backtest is its ceiling; optimizing 3 parameters cannot lift it to significance. abandon_class = premise_falsified.

Implementation

Dollar-neutral ETH/BTC relative-strength regime momentum on Binance USD-M daily bars. Signal = volatility-normalised 21-day momentum of ln(ETHUSDT/BTCUSDT): z = (lr_t - lr_{t-21}) / (vol_daily * sqrt(21)), computed continuously every bar. Enters LONG ETH / SHORT BTC when z > +0.70 and SHORT ETH / LONG BTC when z < -0.70 (dead band), and exits only when z crosses hysteretically to the far side of zero (side*z < -0.30*entry_threshold) — the position is held through signal decay, which keeps turnover at roughly one rotation every couple of weeks. Both legs are sized to the SAME USD notional (volatility-targeted: N = equity * target_vol / ratio_vol, capped at max_notional_frac * equity * leverage), so net dollar delta is ~0 and crypto beta is removed; neither leg is submitted unless both are viable, and a broken hedge flattens the book.

Verification Results

None required; overrides are intentional and correctly chain to super().

min_bars_required, _submit_entry and _submit_exit override framework methods (static analyzer base_shadow warnings). The overrides appear deliberate and correct: _submit_entry/_submit_exit call super() and add the paired BTC leg with both-or-neither guarding; min_bars_required returns a sensible warmup. No functional defect.

Faithful implementation but this is the dollar-neutral major-perp long-short class the corpus has struggled with (L45; the prior SOL/ETH ratio attempt died). The sandbox smoke test shows profit_factor 1.062, Sharpe 0.289 (CI [-0.43, 1.08] straddles zero), and — most notably — a NEGATIVE 'normal' vol tercile (-25.3% / Sharpe -0.19) while being positive only in calm and stressed thirds. avg_trade_return_pct 0.773% clears the fee floor and the impact/capacity fixes clearly worked (impact_cost_pct fell from 66.8% to 5.1%, capacity ~$37M vs prior ~$224k, 146 trades vs 1612), so the root-cause redesign is real. The open question for the analyst is whether the ETH/BTC ratio regime persists as a signable edge on full history rather than being right-tail/regime-dependent. Advisory only.

Backtest Review

Sharpe
0.27
Total return
28.45%
Max drawdown
38.10%
Trades
150
Win rate
50.7%
Profit factor
1.05

Genuine, verified root-cause redesign of the prior SOL/ETH ratio death: impact_cost_pct fell from 66.8% to 6.4%, capacity rose from ~$224k to ~$24M, and trades fell from 1612 to 150 (75 long / 75 short) — the deepest-legs + slow-daily + deadband design worked as intended.

Faithful implementation (QA-confirmed): balanced dollar-neutral long/short legs, both submitted together, 21-day vol-normalised ratio momentum with hysteretic exit; 75 signals -> 75 submitted, benchmark_meaningful correctly false.

avg_trade_return_pct 0.653% clears the Binance USD-M 0.15% fee floor with margin; commission only 1.35% of gross.

profit_factor 1.0495 sits in the cost-drag band ([0.85,1.10]) — gross edge barely exceeds costs even though per-trade return clears the floor.

Sharpe 0.272 with CI [-0.44, 1.05] straddling zero — not distinguishable from no-skill; CAGR only 3.55% against a 38% max drawdown (recovery_factor 0.64) and a 1031-day (2.8yr) underwater stretch.

Edge is right-tail / regime-dependent, contradicting the 'persistent signable regime' premise: NEGATIVE in the dominant 'normal' vol tercile (-24.65% / Sharpe -0.17, 803 days) while positive only in calm (+50.6%) and stressed (+11.7%). Returns are carried by a handful of violent rotation months (2020-08 +18%, 2021-06 +15%, 2024-10 +16.8%, 2025-05 +13.8%, 2025-09 +23.7%).

Loses two of the last four calendar years badly (2023 -10.0%, 2024 -13.9%).

Dollar-neutral major-perp long-short class prior is 0/84 (L45); a PF~1.05 / insignificant-Sharpe base is the ceiling for this class and will not deflate to significance — optimization would curve-fit the lumpy right-tail winners.

Outcome Summary

Fixing the known cause of a prior death (here, market impact and capacity) is necessary but not sufficient — a residual edge concentrated in a handful of right-tail months rather than present across the dominant regime falsifies the strategy's core premise and will not survive deflation.

The analyst issued an 'abandon' verdict at the backtest-review gate before optimization, because the edge was thin and right-tail/regime-dependent — negative in the dominant 'normal' vol tercile (-24.65%, Sharpe -0.17 over 803 days) and carried by a few violent rotation months — which falsifies the 'persistent signable regime' premise; with PF~1.05 and a Sharpe straddling zero in the 0/84 dollar-neutral major-perp long-short class, 3-parameter optimization could not lift it to significance.

A dollar-neutral ETH-vs-BTC long-short on daily USD-M perps that traded the persistent multi-week regime of the ETH/BTC ratio via a slow 21-day volatility-normalised momentum signal with an entry deadband and far-side exit band, deliberately re-engineering the capacity/impact failure that killed a prior SOL/ETH ratio attempt.

Over the 6.7-year backtest (2020-2026, 150 trades: 75 long / 75 short) it returned 28.45% total (CAGR 3.55%) with a Sharpe of 0.272 (CI [-0.44, 1.05]), profit factor 1.049, win rate 50.7%, and a 38% max drawdown with a 1031-day underwater stretch. The impact/capacity redesign worked as intended — impact cost fell from 66.8% to 6.4%, capacity rose from ~$224k to ~$24M, and avg_trade_return_pct of 0.653% cleared the 0.15% fee floor.

Backtest and paper results are hypothetical. Trading involves risk of loss.