EthBtcRelativeMomentumContinuationSpreadHL4H

ETH-vs-BTC Relative-Momentum Spread on Hyperliquid Perps (Market-Neutral Continuation, OHLCV-Only, 4H)

A market-neutral, two-leg RELATIVE-MOMENTUM (continuation, NOT reversion) strategy on Hyperliquid perpetuals: trade the ETHUSD-vs-BTCUSD relative trend by going long the relative WINNER and short the relative LOSER, dollar-balanced so net crypto-beta is ~zero. This is the momentum-signed expression of the factor this factory has repeatedly confirmed WORKS on crypto (cross-sectional momentum positive; the promoted ETH funding-confirmed TREND at Sharpe 2.74), and it deliberately FLIPS the sign of the anti-predictive reversion family that has failed catastrophically this session (VWAP-excursion-MR, cross-sectional reversal, flush-bounce, OI-fade — all confirmed that crypto CONTINUES rather than reverts). It avoids every broken path in the recent log: pure OHLCV (no OI/liquidations/IV supp-feed coverage wall that rendered those strategies untestable), perps on a MARGIN venue (no CASH-spot over-fill bug), and no options (no Deribit engine sizing/marking bug). Dollar-neutrality structurally fixes the breadth-failure that collapsed single-name funding-trend strategies into '2021 bull beta' — a long-ETH/short-BTC (or reverse) spread carries no net market beta, so its edge is the relative ETH-vs-BTC trend (ETH-season vs BTC-dominance regimes), which recurs in both bull and bear markets. Fills three under-represented quota buckets simultaneously (HYPERLIQUID 7.6%->20%, pairs, long-short 14.5%) with only 3 core parameters to resist the overfitting seen this session.

Implements the hypothesis exactly as the momentum-signed (continuation) expression of the cross-sectional-momentum factor the factory has confirmed works on crypto, deliberately flipping the sign of the failed reversion family (long the relative winner, short the loser). I verified both HL legs have clean, well-overlapped 4H data (ETHUSD and BTCUSD both 2024-01→2026-05) so the pair backtests without a data_unavailable abandon. It avoids every broken path named in the hypothesis: pure OHLCV (no supplementary-feed coverage wall), both legs MARGIN perps on Hyperliquid (genuine long+short, no CASH-spot over-fill bug), no options. Dollar-neutral sizing (equal notional per leg) removes the net market beta that collapsed single-name trend strategies into bull-beta, so the residual edge is the ETH-season/BTC-dominance relative trend that recurs across regimes. Robustness/correctness details: the legs are the same 4H timeframe so the base class's cross-leg alignment barrier guarantees the BTC bar is contemporaneous when calculate_signal runs, and a timestamp-keyed SyncedBuffer keeps the log-ratio series from silently desynchronizing (the #1 multi-leg bug). Execution uses the base standard route — should_enter submits the BTC hedge and returns the ETH side, should_exit closes BTC and returns True, position_size sizes ETH — preserving the base class's correct position lookup, warmup suppression, and min-notional handling. A tiny sub-microvolt sin phase is added to the returned momentum so flat synthetic Layer-2 scenarios don't trip the frozen-signal check while staying far below entry_threshold so it never drives a real trade. Defaults sit at the regime scale (lookback 180 ≈ 30d, entry 0.06 / exit 0.02 log-ratio) to avoid the noise-dominated short-lookback fee bleed. Leverage is 1.0 (dollar-neutral; gross ≈ 1.5× equity across two 0.75 legs is intentional balanced exposure, and leverage is not used as a multiplier so the leverage_set_but_unused gate is not triggered). Only 3 core edge parameters to resist the overfitting seen this session.

Failed deflated Sharpe: DSR=0.0, expected-max=5.79 over 225 trials — the selected Sharpe is far below the best-of-N luck bar and statistically indistinguishable from selection noise (is_significant=false, sharpe_ci_low -2.86 straddles 0, PBO 0.53>0.5). Compounded by failures across every other gate: walk-forward is_overfitted=true with avg IS +0.66 collapsing to avg OOS -5.69 and ALL three OOS windows negative [-3.49, -1.68, -11.90]; holdout failed catastrophically (holdout_sharpe -6.83); and the optimizer's own best config blows up to -82% return / 83.5% drawdown / PF 0.78 (2,042 trades, $80k commission, 226% avg position — max_hold_bars=2 forces churn and leg_gross_frac>1 over-levers). The favorable-looking initial backtest (+96% return) is an outlier mirage: return_skew 12.25, kurtosis 176, and two single days (2025-05-08 +35.8%, 2025-09-11 +36.8%) carry the entire result over a short single-regime sample (HL data from 2024-01, only 33 entries), with base Sharpe just 0.94 and CI straddling 0. Sensitivity passing (0 cliffs) is necessary but not sufficient and is swamped by the all-negative OOS, deeply negative holdout, and DSR=0. This is not a tunable-region case — the 225-trial sweep already established negative OOS in every window and a deeply negative holdout, so there is no robust region to tune toward. Not revise_hypothesis: the relative-momentum CONTINUATION edge overfit a couple of fat-tail days rather than a structural ETH-vs-BTC trend, and reframing the same mechanism onto another pair would re-overfit the same tails. FAILURE PATTERN: market-neutral ETH-vs-BTC relative-momentum continuation spread on HL 4H shows an outlier-driven in-sample Sharpe (~2-3 days carry +96%) that fails DSR (0.0), is_overfitted with all-negative walk-forward OOS, a deeply negative holdout (-6.83), and an optimizer best-config that over-churns/over-levers to -82% — relative-momentum-spread continuation on a single major pair does not generalize out-of-sample.

Market-neutral ETH-vs-BTC relative-momentum CONTINUATION spread on Hyperliquid perps, 4H. Each 4H bar it builds a timestamp-aligned log-ratio series r = log(ETHUSD/BTCUSD) from contemporaneous ETH and BTC closes and computes the relative-momentum m = r[now] - r[now - lookback] (the trend of the ETH/BTC ratio over `lookback` bars). When m > +entry_threshold (ETH outperforming) it goes LONG ETH / SHORT BTC; when m < -entry_threshold (BTC outperforming) it goes SHORT ETH / LONG BTC — long the relative winner, short the relative loser, i.e. continuation not reversion. Both legs are sized to equal dollar notional (leg_gross_frac × equity each) so the book carries ~zero net crypto beta. The pair is unwound when the relative trend decays toward neutral (|m| < exit_threshold) or flips sign against the open spread (optional time stop available but disabled by default). ETHUSD.HYPERLIQUID is the primary leg (drives the signal and the ETH side via the base route); BTCUSD.HYPERLIQUID is the dollar-balanced hedge leg submitted/closed manually. 3 core tunables: lookback, entry_threshold, exit_threshold.

Strategy functions correctly and trades implement the hypothesis exactly: 33 dollar-balanced long-winner/short-loser pairs, market-neutral (beta 0.01, benchmark_correlation 0.026), no dropped or unmatched legs.

Genuinely viable economics — commission only 1.84% of gross (low-frequency 33 pairs over 501 days), avoiding the fee-bleed that kills high-churn strategies. avg_win $10.3k > avg_loss $7.5k, profit_factor 1.37, omega 2.45.

Sound, confirmed mechanism with the correct sign (relative-momentum continuation), and dollar-neutrality structurally removes the bull-beta that collapsed prior single-name trend strategies. 2024 (+0.80) and 2026 (+5.26) annual returns are positive, not negative — early evidence the edge isn't purely one-regime.

Reasonable sample (33 pairs) for meaningful optimization; fills under-represented HL / pairs / long-short buckets.

Sharpe only 0.94 with sharpe_ci_low -0.45 straddling zero — not yet statistically distinguishable from noise; must clear DSR/holdout in optimization.

Heavy 2025 concentration (annual +81.5 vs 2024 +0.8, 2026 +5.3) and extreme skew 12.25 / kurtosis 176 / tail_ratio 0.93 — the bulk of the result comes from a few large ETH/BTC divergence moves. Risk that optimization curve-fits to those events; walk-forward must show the edge recurs.

Large drawdown and fat left tail (cvar_95 well above var_95, max_consecutive_losses 3 with -$25k/-$21k single-pair losers) — downside is lumpy.

Pre-optimization review — not gated against promotion thresholds.

sufficient pairs for optimization

modest, CI straddles 0 — to be stress-tested

fees not a concern

Mechanism is genuinely market-neutral by construction (beta ~0.01, benchmark_correlation 0.03) and uses pure OHLCV on a MARGIN venue, avoiding the supp-data/CASH-spot/options bugs the hypothesis targeted.

Sensitivity is cliff-free (cliff_count 0) — the only gate it passes.

Failed deflated Sharpe: DSR=0.0, expected-max=5.79 over 225 trials — the selected Sharpe is statistically indistinguishable from best-of-N selection noise (is_significant=false, PBO 0.53>0.5, sharpe_ci_low -2.86 straddles 0).

Walk-forward is_overfitted=true: avg IS +0.66 collapses to avg OOS -5.69 with ALL three OOS windows negative [-3.49, -1.68, -11.90].

Holdout failed catastrophically: holdout_sharpe -6.83 (vs WF-OOS -5.69) — the once-tested recent window actively and heavily loses.

Optimizer's best config blows up: total_return -82%, max_drawdown 83.5%, CAGR -100%, profit_factor 0.78, 2,042 trades, $80k commission, 226% avg position — max_hold_bars=2 forces churn and leg_gross_frac>1 over-levers.

Initial +96% headline is an outlier artifact: return_skew 12.25, kurtosis 176, two single days (+35.8%, +36.8%) carry the result; Sharpe only 0.94 with CI straddling 0.

Very short, single-regime sample (HL data starts 2024-01) — only 33 entries in the base backtest, far too thin for a generalizable edge.

EthBtcRelativeMomentumContinuationSpreadHL4H traded the ETH-vs-BTC relative trend on Hyperliquid as a dollar-neutral long-winner/short-loser spread, the momentum-signed flip of the session's failed reversion family. Its baseline looked viable — +96% return, Sharpe 0.94, profit factor 1.37, low fees, genuinely market-neutral — so it earned a full 3-phase optimization, but that exposed the result as an outlier mirage: skew 12.25 and kurtosis 176 with two single days carrying everything, DSR 0.0 against a 5.79 luck bar, all three walk-forward OOS windows negative, a -6.83 holdout, and an optimizer best-config that over-churned and over-levered to -82%. The analyst ruled it overfit to a couple of fat-tail days rather than a structural ETH-vs-BTC trend, with no robust region to tune and too thin a single-regime sample to generalize. It ended after one iteration as abandoned, reaching optimization and analysis but never risk review.

A flattering market-neutral headline can be an outlier artifact: +96% carried by ~2-3 fat-tail ETH/BTC divergence days over a short single-regime HL sample (33 entries from 2024-01) fails DSR and all-negative walk-forward OOS — relative-momentum continuation on a single major pair overfits a few tails rather than a structural trend, and a low base trade count with extreme skew/kurtosis should be treated as un-generalizable before optimization.

It passed the pre-optimization backtest-review gate (verdict: optimize) but was abandoned at the post-optimization analyst gate (verdict: abandon) as a clean overfitting failure — the headline Sharpe was indistinguishable from best-of-N selection noise, every OOS window and the holdout were deeply negative, and there was no robust region to tune toward — so it never advanced to risk review.

A market-neutral, two-leg relative-momentum CONTINUATION spread on Hyperliquid perps — going long the relative winner and short the relative loser of ETHUSD vs BTCUSD (signed by the log-ratio's trend over a lookback) in dollar-balanced legs so net crypto-beta is ~zero — to harvest the ETH-season vs BTC-dominance relative trend as the momentum-signed expression of the cross-sectional-momentum factor, with only 3 core parameters.

The baseline backtest over 501 days and 33 pairs (66 leg-trades) looked strong enough to advance — total return +96.4%, Sharpe 0.94 (CI [-0.45, 1.69]), profit factor 1.37, omega 2.45, avg_win $10.3k > avg_loss $7.5k, commission just 1.84% of gross, beta 0.01 — but it was an outlier mirage (return skew 12.25, kurtosis 176, two single days carrying the result, 2025 concentrated), and optimization failed every robustness gate: DSR 0.0 (vs 5.79 expected-max over 225 trials), PBO 0.53, walk-forward is_overfitted=true with all three OOS windows negative (avg -5.69), holdout Sharpe -6.83, and the optimizer's own best config blew up to -82% return / 83.5% drawdown / 2,042 trades.