Skip to content
All strategies

Strategies

SUI Hyperliquid 4H Donchian-Breakout Trend, Long-Short, Regime-Filtered ATR-Trailing (Single-Instrument, 3-Parameter)

Outcome: Abandoned

SuiHyperliquidDonchianBreakoutTrendLS

Outcome Summary

This strategy expressed the factory's one paper-reaching mechanism family — slow directional trend on a liquid perp — as a low-parameter Donchian breakout on the SUI Hyperliquid 4H perp, deliberately single-instrument to dodge the HL-basket crash and long-short to attack the direction and venue quota gaps. The initial backtest passed the pre-optimization gate (verdict: optimize) on a healthy 140-trade sample with a per-trade edge well above the fee floor, though the analyst flagged that its return was concentrated in the stressed vol tercile and a few crisis windows. Three-phase optimization then confirmed the concern rather than resolving it: it curve-fit entry_thresh up to 1.96, collapsing 140 trades into 11, and failed the walk-forward overfit, PBO, and holdout HARD gates while showing a deflated Sharpe of 0.0 and no significance. With every prior Hyperliquid Donchian-breakout port already dead (0/N) and the mechanism too slow to validate on HL's short history, the analyst abandoned it as an overfit/no-generalization death after one iteration.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument directional TREND strategy on the SUI Hyperliquid perpetual (SUIUSD.HYPERLIQUID), 4H bars. It deliberately targets the two most extreme quota gaps at once: direction (portfolio is 85.9% long_only vs a <=55% target) and venue (HYPERLIQUID is 7.2% vs a >=20% target, and live trading is shifting there). It stays inside the ONE mechanism family that has ever reached paper-stage in this factory -- slow directional trend on a liquid perp -- but expresses it with a DISTINCT, robust, low-parameter Donchian channel-breakout signal rather than the saturated EMA dual-timeframe momentum-confluence template that has repeatedly overfit when ported to new names (LTC/DOT/BNB/LINK/SOL all died overfit). It is deliberately SINGLE-INSTRUMENT to dodge L88 (HL multi-instrument baskets crash backtest_failed) and uses 4H bars so it lives inside HL's ~2.3y of usable 4H history (L: sub-hourly HL dies on insufficient span). SUI is a liquid HL perp with minimal presence in the BTC/ETH/SOL-saturated corpus, spending the data estate rather than re-mining one series.

Implements the hypothesis directly: a slow directional trend mechanism on a liquid perp, expressed as a Donchian channel breakout rather than the saturated EMA dual-timeframe confluence template, on SUI's Hyperliquid perp with 4H bars (catalog holds 5,732 bars from 2024-01-25 to present -- comfortably inside HL's usable 4H span, and single-instrument so there is no multi-leg HL basket failure mode). It is long AND short, which addresses the direction quota gap and is why the futures venue is required; HYPERLIQUID……Show moreShow less

Implements the hypothesis directly: a slow directional trend mechanism on a liquid perp, expressed as a Donchian channel breakout rather than the saturated EMA dual-timeframe confluence template, on SUI's Hyperliquid perp with 4H bars (catalog holds 5,732 bars from 2024-01-25 to present -- comfortably inside HL's usable 4H span, and single-instrument so there is no multi-leg HL basket failure mode). It is long AND short, which addresses the direction quota gap and is why the futures venue is required; HYPERLIQUID addresses the venue quota gap. Parameter surface is deliberately small: three economically meaningful knobs (donchian_period, atr_trail_mult, er_min) plus standard risk/ATR settings, all declared in _param_bounds so the optimizer searches exactly the ranges the code clamps to. The signal is returned raw and continuous every bar (never gated to a constant), with the entry decision thresholded in should_enter. The regime filter is a single mild condition (efficiency ratio >= 0.18, median ER on this series is 0.207) rather than a stacked conjunction: an offline check on the real 4H series shows 470 entry-eligible bars out of 5,732 (231 long / 239 short), so trade frequency is healthy in both directions and zero-trade failure is not a risk. Leverage stays at 1.0 deliberately -- ATR-risk sizing with a median ATR of 3.1% of price and a 3x trail implies ~9.3% stop distance, so a 2% risk budget deploys only ~22% of equity in notional and never needs margin; the leverage field is read in position_size only as the notional cap multiplier, so nothing is set-but-unused. Per-trade edge: a 9.3%-wide ATR trail means realised moves are two orders of magnitude above Hyperliquid's ~0.09% round-trip taker cost.

Overfit / no-generalization death (NOT fee-edge — per-trade return clears the HYPERLIQUID floor with wide margin). Three unwaivable HARD gates fail: walk-forward is_overfitted=TRUE (avg IS Sharpe 4.589 -> avg OOS 0.282, windows -5.26/+6.11/0.0), PBO 0.5251 > 0.5, and the holdout FAILED with Sharpe -1.247 <= 0 on only 2 trades. Optimization made the strategy WORSE, not better: it pushed entry_thresh to 1.96 and collapsed 140 default-param trades into 11 over 2.6 years — a curve-fit to a handful of favorable……Show moreShow less

Overfit / no-generalization death (NOT fee-edge — per-trade return clears the HYPERLIQUID floor with wide margin). Three unwaivable HARD gates fail: walk-forward is_overfitted=TRUE (avg IS Sharpe 4.589 -> avg OOS 0.282, windows -5.26/+6.11/0.0), PBO 0.5251 > 0.5, and the holdout FAILED with Sharpe -1.247 <= 0 on only 2 trades. Optimization made the strategy WORSE, not better: it pushed entry_thresh to 1.96 and collapsed 140 default-param trades into 11 over 2.6 years — a curve-fit to a handful of favorable breakouts, the textbook sign the base backtest was riding a lucky path rather than a real edge. Robustness confirms best-of-N noise: deflated_sharpe 0.0 << 0.95 (expected-max ~5.09 over 225 trials), is_significant=false, Sharpe CI [-0.66, 1.69] straddles zero, and it does not survive programme-level FDR (keeps 4 of 248, candidate_p 1.0). The oos_vault is itself flagged UNDERPOWERED (~26 expected trades vs 100 needed), so this mechanism is too slow to validate on HL's ~2.6y of 4H history. Per the overfit-vs-dead-premise rule, high-IS/OOS-collapse is ABANDON not revise: every Hyperliquid Donchian-breakout port has already died (SOL HL 4H HTF-confirm overfit, SOL HL daily overfit, SOL HL 4H bidirectional risk_reject at 44% DD, ETH HL 12H, BTC HL daily fee_edge) — 0/N — so re-pointing the same recipe at SUI just overfits again. Sensitivity plateaus at ~0.4-0.6 Sharpe everywhere with no untapped robust region, so there is nothing concrete for iteration (attempt 1 of 2) to tune toward. Failure pattern: overfit / no-generalization (Hyperliquid Donchian-breakout port).

Implementation

Long-short 4H Donchian channel-breakout trend strategy on the SUI Hyperliquid perpetual (SUIUSD.HYPERLIQUID). Every bar it computes a continuous channel-position score, (2*close - dc_high - dc_low)/(dc_high - dc_low), from the highest high and lowest low of the prior `donchian_period` bars (current bar excluded): the value sits in roughly [-1, +1] inside the channel, rises above +1 on an upside breakout and falls below -1 on a downside breakout. Entries are taken when that score crosses +/- entry_thresh AND a……Show moreShow less

Long-short 4H Donchian channel-breakout trend strategy on the SUI Hyperliquid perpetual (SUIUSD.HYPERLIQUID). Every bar it computes a continuous channel-position score, (2*close - dc_high - dc_low)/(dc_high - dc_low), from the highest high and lowest low of the prior `donchian_period` bars (current bar excluded): the value sits in roughly [-1, +1] inside the channel, rises above +1 on an upside breakout and falls below -1 on a downside breakout. Entries are taken when that score crosses +/- entry_thresh AND a single regime filter passes -- Kaufman's efficiency ratio over the same lookback must be >= er_min, so breakouts are only taken while price is travelling directionally rather than chopping. Positions exit on a chandelier ATR trailing stop (atr_trail_mult * Wilder ATR measured from the highest high / lowest low reached since entry) or on an opposite-side channel breakout. Size is risk-based: risk_pct of equity divided by the ATR stop distance, capped at max_notional_frac * equity * leverage. All state is incremental (bounded deques, Wilder ATR recursion, O(donchian_period) per bar) so the full ~5.7k-bar history runs far inside the smoke-test budget.

Verification Results

No action required; confirm the override is intended (it is consistent with the strategy's lookback needs).

min_bars_required() overrides a base-template method. Static analysis flagged this as a base_shadow. It appears deliberate (returns max(donchian_period, atr_period)+5) and the signal path self-guards by returning 0.0 until the deques fill, so warm-up is safe regardless. The sandbox processed 5,732 bars and produced 140 trades, confirming warm-up gating works.

Sandbox Sharpe is modest (0.34) with a CI that straddles zero [-0.80, 1.60], and the return is concentrated in the stressed vol tercile (+19.3% vs +0.67% calm / +2.23% normal), so the edge leans on a few high-vol windows (yen-carry Aug-2024 Sharpe 7.98, tariff-shock Apr-2025 Sharpe 4.69). Single-name trend/momentum ports to HYPERLIQUID have repeatedly failed to generalize in this factory (SOL/ETH HL confluence, HL Donchian/vol-gate all died overfit or fee_edge). This is advisory only — the code is a faithful implementation; the analyst should judge robustness on full-history walk-forward, watching PBO/deflated-Sharpe and whether the edge survives outside the stressed tercile.

Backtest Review

Sharpe
0.34
Total return
31.28%
Max drawdown
18.96%
Trades
140
Win rate
36.4%
Profit factor
1.21

Per-trade edge clears costs with margin: avg_trade_return_pct 0.531% >> the HYPERLIQUID 0.15% fee floor; commissions only 2.18% of gross, so this is NOT a fee-fragile signal.

Adequate, reliable sample for a 4H trend strategy: 140 trades over 955 days (metrics_reliable=true), balanced 65 long / 75 short — a genuine long-short book, not a closet long (beta 0.03, alpha +11.4%).

Faithful mechanism: 140 signaled = 140 submitted (zero size/notional/cash drops); trades implement the Donchian breakout + efficiency-ratio regime gate as hypothesized. Single-instrument design dodges the HL-basket backtest_failed trap (L88).

PF 1.209 with controlled risk: max_drawdown 18.96%, recovery_factor 1.61, capacity ~$2.6M. Positive in 2024 (+15%) and 2026 (+16.2%).

Return is concentrated in the stressed vol tercile (+19.3% vs +0.67% calm / +2.23% normal) and in a handful of windows (yen-carry Aug-2024, tariff-shock Apr-2025) — the QA edge concern. Sharpe 0.337 with CI [-0.93, 1.51] straddling zero; 2025 was -4%.

Market-impact is a meaningful drag (impact_cost_pct 19.6%) though below the 25% flag.

HL single-name trend ports have a poor generalization track record in this factory — the 3-phase run (PBO / deflated-Sharpe / holdout) must confirm the edge survives outside the stressed tercile.

Analysis

Per-trade edge clears the venue fee floor (avg_trade_return_pct 0.53% default / 5.79% optimized >> HYPERLIQUID 0.15%); commission only ~2% of gross — this is NOT a fee-edge death.

Sensitivity surface is broadly non-cliff (only atr_trail_mult flagged) with a modest ~0.4-0.6 Sharpe plateau; low max drawdown on the optimized config (5.95%).

Single-instrument design avoided the L88 HL-basket backtest crash and the default backtest produced an adequate 140-trade sample.

Walk-forward is OVERFIT (HARD gate): avg IS Sharpe 4.59 collapses to avg OOS 0.28, with a deeply negative first OOS window (-5.26); is_overfitted=true.

PBO 0.5251 > 0.5 (HARD gate) — the parameter selection is more likely than not overfitting.

Holdout FAILED (HARD gate): Sharpe -1.247 on only 2 trades over the unseen window — a negative, structurally underpowered forward result.

Optimization made the strategy WORSE, not better: it curve-fit entry_thresh up to 1.96, collapsing 140 default-param trades into 11 over 2.6 years — the classic sign the sweep fit a handful of favorable breakouts rather than a robust edge.

Robustness confirms best-of-N noise: deflated_sharpe 0.0 << 0.95, is_significant=false, Sharpe CI [-0.66, 1.69] straddles zero, and it does not survive programme-level FDR (keeps 4 of 248, candidate_p 1.0). Selected Sharpe 0.58 sits far below the expected-max luck bar of ~5.09 over 225 trials.

oos_vault is flagged UNDERPOWERED (~26 expected trades vs 100 needed) — the mechanism is too slow to validate on HL's ~2.6y of 4H history.

Class prior: every Hyperliquid Donchian-breakout port has died (SOL HL 4H HTF-confirm overfit, SOL HL daily overfit, SOL HL 4H bidirectional risk_reject 44% DD, ETH HL 12H other, BTC HL daily fee_edge) — 0/N. Re-pointing the same recipe at SUI just overfits again.

Benjamini-Hochberg at q=0.10 over 248 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

A Hyperliquid Donchian-breakout trend port on a slow 4H timeframe is too slow to validate on ~2.6y of history and keeps overfitting (0/N prior HL ports survived) — a promising default backtest riding a lucky, stressed-regime-concentrated path is not a real edge, and re-pointing the same recipe at a new name (SUI) just reproduces the failure.

The analyst abandoned it after the 3-phase optimization failed three unwaivable HARD gates: walk-forward is_overfitted=true (avg IS Sharpe 4.59 collapsing to avg OOS 0.28, first OOS window -5.26), PBO 0.5251 > 0.5, and a failed holdout (Sharpe -1.247 on only 2 trades); it also did not survive programme-level FDR and its deflated Sharpe was 0.0.

A single-instrument, long-short 4H Donchian channel-breakout trend strategy on the SUI Hyperliquid perpetual, gated by a Kaufman efficiency-ratio regime filter and exited via a chandelier ATR trailing stop, aimed at filling the factory's direction and HYPERLIQUID-venue quota gaps.

The default backtest looked viable — 140 trades over 955 days, +31.3% total return, profit factor 1.21, Sharpe 0.337 (CI [-0.93, 1.51]), max drawdown 19.0%, and avg per-trade return 0.53% clearing the 0.15% fee floor — but returns were concentrated in the stressed vol tercile (+19.3% vs +0.67% calm). Optimization made it worse: it pushed entry_thresh to 1.96 and collapsed the sample to 11 trades, with Sharpe 0.58, deflated Sharpe 0.0, and PBO 0.5251.

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