Strategies

Hyperliquid ETH Perp — Trend Continuation with a Funding-Crowding BRAKE: Long-Short 4H Momentum That Is VETOED When Funding Is at a Positioning Extreme (Don't Chase an Over-Leveraged Move), Chandelier ATR Trailing Exit (HYPERLIQUID, 4H Bars, 3-Parameter)

Outcome: Abandoned

HyperliquidEthTrendFundingCrowdingBrake

Outcome Summary

HyperliquidEthTrendFundingCrowdingBrake rode a 4H EMA trend on the Hyperliquid ETH perp with a Chandelier trailing exit, adding a crypto-native funding-crowding brake that refuses to chase a trend when leveraged positioning is already extreme in the same direction — deliberately the opposite polarity to the dead OI-confirmed breakout and distinct from funding carry. The implementation was clean and honestly reported, but the result was marginal: Sharpe 0.40 with a CI straddling zero, PF 1.21, a 0.38 win rate, and — by the developer's own ablation — a funding veto worth essentially nothing (Sharpe 0.48 off vs 0.49 on), leaving a bare EMA trend. The analyst abandoned it on the first iteration at the backtest_review gate as the single-name trend clone no-edge class (0/97 Hyperliquid, 0/213 OHLCV survival), below the OOS floor with no robust region to optimize. It never advanced to optimization, analysis, or risk review.

Strategy report

Hypothesis

A LONG-SHORT, SINGLE-INSTRUMENT directional TREND strategy on ETHUSD.HYPERLIQUID (4H bars, ~2.3yr HL history — comfortably above the WF/holdout requirement, no sub-hourly HL history trap). It leans into the factory's ONE demonstrated survivor class (directional momentum with trailing-stop winners — the BTC dual-TF book at Sharpe 1.99) but is NOT a fanout of that template: the signal structure is a single robust EMA-trend, and the genuinely new element is a crypto-native OVER-CROWDING BRAKE built from the Hyperliquid funding feed. The mechanism: take the trend, but REFUSE to enter (and exit early) when perpetual funding is at a positioning EXTREME in the same direction as the trade — i.e. do not buy an up-trend when longs are already paying rich funding (over-leveraged, reversal-prone), and do not short a down-trend when shorts are crowded (funding deeply negative). This is deliberately the OPPOSITE polarity to the dead OI-CONFIRMED breakout (which REQUIRED leveraged confirmation and overfit) and is NOT funding CARRY (no funding is harvested — funding is used only as a veto filter on a price-momentum trade, so the abandoned single-asset/basket carry economics do not apply). Fills three under-represented buckets at once: HYPERLIQUID (7.1% vs ≥20% target), long_short (14.2% vs push target), and a funding-data edge that is neither carry nor cross-sectional-rank (both of which are confirmed dead classes here).

Implements the hypothesis as specified: single-instrument HL ETH 4H directional trend, chandelier trailing exit, and the genuinely new element — a funding-CROWDING brake with veto polarity (refuse to chase a move that leveraged positioning is already paying for), which is the opposite sign to an OI/funding-CONFIRMED breakout and is not carry (no funding cash flow is earned). Four implementation points. (1) The brake is a ROLLING Z-SCORE of funding against its own 30-day distribution, never an absolute bps threshold — HL funding sits pinned at its 1.25e-5/hr interest baseline for long stretches, so a fixed level would fire constantly in one regime and never in another; measured on the real series, |z| >= 2.5 occurs on ~1.1% of bars, which is what a 'positioning extreme' should mean. It is computed in O(1) from prefix sums of the rate and its square built once in __init__ (two bisects per bar, no rescan). (2) A funding GAP means NO VETO, deliberately: the edge here is price momentum and funding is only a filter, and the HL ETH funding history has five real holes over 12h (two of about 30 days) covering ~12% of bars — blocking entries inside a data outage would hand an eighth of the sample to the collector rather than to the hypothesis. This is the opposite of the correct policy for a funding-carry strategy, and the docstring says so explicitly. (3) A netting-venue bug is fixed in code and worth flagging for other strategies: Hyperliquid runs OMS NETTING, so every new position REUSES the same position.id. Keying the chandelier's peak/trough on position.id meant each fresh position inherited the previous trade's extremes and was stopped out on its first bar — the engine produced 1800 four-hour trades and -43% before the fix; the trail is now keyed on ts_opened and the same code returns 117 multi-day trades and +23.7%. (4) Honest reporting of the brake's contribution, measured in the engine rather than claimed: defaults give 117 trades / +23.7% / Sharpe 0.49 / maxDD 20.2% / PF 1.28 / avg_trade_return_pct +0.96% over the full funding-overlap window (2024-08 to 2026-08) and 56 trades / +5.7% / Sharpe 0.30 / avg_trade +0.32% over the last 365 days, versus 110 trades / +22.9% / Sharpe 0.48 / maxDD 21.5% with the veto disabled — so the brake is worth about a point of return, a point of drawdown and two points of win rate. It improves the trade; it is not the trade. Per-trade capture (0.3-1.0% of notional) is far above the 0.09% HL round trip, and ~55 trades a year at a ~5-day mean hold keep the walk-forward and 15-day holdout windows populated.

single_name_trend_clone_no_edge (Sharpe CI straddles zero, Hyperliquid class prior). A 4H EMA directional trend on ETHUSD.HYPERLIQUID — the single-name OHLCV trend/momentum clone class (0/97 Hyperliquid, 0/213 OHLCV survival, L41/L9). Base Sharpe is only 0.40 with sharpe_ci_low -0.67, so the edge is statistically indistinguishable from zero and already below the 0.5 OOS floor before the ~225-trial deflation; PSR 0.78 and PF 1.21 / win_rate 0.38 confirm a marginal result. The novel funding-crowding veto is not the source of the edge — the developer's own ablation shows Sharpe 0.48 with it off vs 0.49 on — so the mechanism reduces to a bare EMA trend. Returns are regime-dependent over a short 2.5-yr sample (2024 -13.1%). Not tunable: the base Sharpe is the ceiling and it is below the floor, so optimization would curve-fit and die overfit at holdout as every sibling in this class has.

Implementation

Long-short 4H trend continuation on ETHUSD.HYPERLIQUID with a funding-crowding veto. Every bar it returns (EMA30 - EMA120) / ATR24 — the trend spread in ATRs, a continuous scale-free value. It goes long when that exceeds +0.15 and short below -0.15, but REFUSES the entry when Hyperliquid funding is at a positioning extreme in the same direction: a rolling z-score of the last 24h of hourly funding against its own 30-day mean and stdev, vetoing longs at z >= +2.5 and shorts at z <= -2.5. Positions are ridden with a Chandelier ATR trailing stop (highest high since entry minus 4 ATRs for a long, mirrored for a short) and are also closed early if the crowding extreme develops against the open position, or if the trend flips to the opposite threshold. Funding is never harvested — it is only a filter — and when the HL funding feed has a gap the brake simply cannot fire rather than blocking the price-momentum trade. Sizing risks 2% of equity across the initial 4-ATR stop, capped at 60% of equity notional; leverage 1.0.

Verification Results

Billed as '3-parameter' but _param_bounds exposes 10 tunables — meaningful overfit surface over a ~2yr / ~117-trade window. Analyst should apply deflation/PBO and not trust the full-window Sharpe 0.49.

Modest, decaying directional-momentum edge (last-year Sharpe 0.30, CI straddles zero, PF 1.19); far weaker than the BTC dual-TF survivor. Crowding brake is honestly a ~1-point contributor, not the edge. OOS matter for analyst.

Backtest Review

Sharpe: 0.40
Total return: 24.11%
Max drawdown: 20.21%
Trades: 149
Win rate: 38.3%
Profit factor: 1.21

Clean incremental implementation (O(1) funding index, netting-safe trail keyed on ts_opened), fills under-represented HYPERLIQUID/long_short buckets

avg_trade_return_pct 0.73% clears the HL fee floor; 149 trades over the window

Genuinely new funding-crowding veto idea, and honestly reported as a modest contributor

Sharpe 0.40 with sharpe_ci_low -0.67 — CI straddles zero, below the 0.5 OOS floor before optimization; PSR 0.78 (<0.95)

Single-name OHLCV directional trend clone — 0/97 Hyperliquid, 0/213 OHLCV survival (L41/L9)

The funding veto is not the edge: with it off, Sharpe 0.48 vs 0.49 — the strategy is a bare EMA trend

PF 1.21 and win_rate 0.38 in the class-fail band; returns lean on a few big winners (kurtosis 7.5)

Regime-dependent over a short 2.5-yr sample (2024 -13.1%); walk-forward/holdout windows will be thin and noisy

Outcome Summary

A genuinely new filter idea (a funding-crowding veto) does not create an edge if an ablation shows it contributes almost nothing — the strategy then reduces to a bare single-name EMA trend, which is a 0/300+ survival class whose Sharpe CI straddling zero cannot be optimized into significance; honestly measuring the filter's marginal contribution is what exposes that it was never the edge.

It was abandoned at the pre-optimization backtest_review gate on iteration 1 as the single-name OHLCV directional trend/momentum clone class (0/97 Hyperliquid, 0/213 OHLCV survival): the base Sharpe of 0.40 with a CI straddling zero is below the 0.5 OOS floor before optimization, PF and win rate sit in the class-fail band, and the novel funding-crowding veto adds essentially nothing — so the base Sharpe is the ceiling and optimization would only curve-fit and die overfit at holdout.

A long-short, single-instrument directional EMA-trend strategy on ETHUSD.HYPERLIQUID (4H bars) that rides the trend with a Chandelier ATR trailing exit but vetoes entering (and exits early) when the Hyperliquid funding z-score is at a positioning extreme in the trade's own direction — a crowding brake that refuses to chase an over-leveraged move, with funding used only as a filter (not harvested), and three parameters (fast/slow EMA, ATR multiple, veto z).

The backtest returned 24.1% total with Sharpe 0.40 (CI low -0.67, straddling zero), profit_factor 1.21, win_rate 0.38, information_ratio -0.05, and max_drawdown 20.2% over 149 trades, with avg_trade_return_pct 0.73% clearing the fee floor. Critically, the developer's own ablation showed the funding veto was not the edge — Sharpe 0.48 with it off vs 0.49 on — so the strategy reduces to a bare EMA trend, and returns leaned on a few big winners (kurtosis 7.5) and were regime-dependent (2024 -13.1%) over a short 2.5-year sample.

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