SolHyperliquidDonchianBreakoutDailyTrendConfirmedLS

SOL Hyperliquid Higher-Timeframe-Confirmed Donchian Breakout, Long-Short (Single-Instrument SOLUSD.HYPERLIQUID Perp, 4H Primary + 1D Trend Confirm, With-Trend Breakouts Only / Flat on Disagreement, ATR-Trailing, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV trend-continuation strategy on SOLUSD.HYPERLIQUID perpetual futures. It deliberately ports the ONLY archetype that has survived optimization in this factory (the BTC dual-timeframe momentum-confluence long-short, Sharpe ~1.99, now in paper) onto the most under-represented venue in the portfolio (Hyperliquid at 5.8% vs a 20% quota target) and the under-represented long-short direction bucket (13.7% vs target). The mechanism is a channel breakout on the 4H bar that only fires WHEN the daily trend agrees, so it trades WITH the higher-timeframe trend and sits flat when the two timeframes disagree — the exact 'flat-on-disagreement' filter that let the BTC sibling avoid the chop that kills every naked breakout system (L9, zero breakout survivors). It is materially differentiated from the BTC success by venue (HL, not Binance), asset (SOL, not the exhausted 2%-survival BTC family), and mechanism (Donchian channel break + daily-SMA trend gate, not EMA momentum confluence). Uses 4H bars on Hyperliquid (~2.3 years of usable candle history) and 1D bars (full history back to 2020) — both comfortably above the HL sub-hourly history wall, so this will NOT be abandoned for insufficient validation data. Pure OHLCV, no supplementary feed (best-surviving tag family at 4%), no options-expiry cadence problem, no cross-venue spot-lookup timing bug, no funding-carry fee-floor problem, no liquidation-feed 90-day wall — it structurally sidesteps every recent verification_loop / data_unavailable death.

Implements the hypothesis directly: 4H Donchian break as the trigger, 1-DAY SMA agreement as the with-trend gate, flat on disagreement, ATR trailing exit, long-short on Hyperliquid, pure OHLCV, three tunable parameters. Three implementation decisions worth flagging. (1) I verified the data and the compound gate on real catalog data BEFORE submitting, because over-selective conjunctions are the top verification-loop killer: SOLUSD.HYPERLIQUID has 4-HOUR bars from 2024-01-25 (5,494 bars) and full 1-DAY history, and with the defaults the FULL compound gate (breakout AND daily-trend agreement) fires 121 times in the 365-day Layer-3 sandbox window (46 long, 75 short) and 388 times over the whole 4H history — so trade count is comfortably above every floor, not a single-digit sample. (2) The daily leg is the exchange's own 1-DAY bar subscribed as an extra bar type, and SOLUSD.HYPERLIQUID is listed in BOTH extra_instruments and extra_bar_types on purpose: InstrumentResolver.resolve_extras iterates over extra_instruments and positionally pairs the bar types, so an empty instrument list would silently DROP the daily feed, leave the trend gate at 0 forever and produce zero trades. Because the daily leg is a slower timeframe than the primary it is excluded from the base class's same-timeframe alignment barrier, so it cannot deadlock the 4H signal; I also dedupe daily bars by timestamp since the catalog holds repeated bars for some timestamps (I checked — deriving the daily series from the 4H stream instead would have been corrupted by those duplicates and by bar-stamping ambiguity, which is why I used the real feed). If the daily feed is ever missing the gate stays 0 and the strategy trades nothing rather than degrading into a naked unconfirmed breakout. (3) calculate_signal returns a continuous channel position every bar — +1 at the channel top, -1 at the bottom, beyond ±1 on a break — so the signal varies smoothly and carries conviction while thresholding at ±1.0 is exactly equivalent to 'closed beyond the prior 20-bar extreme'; no gated constant-default pattern. Risk controls follow the compounding-blowup lesson: 1% of equity risked per trade across the 3-ATR stop, gross notional hard-capped at 1x equity, and no leverage is set or referenced. Venue is Hyperliquid futures (not spot) because the strategy takes shorts, and its ~0.09% round-trip cost is comfortably below the multi-percent moves a 3-ATR trailing trend trade targets.

Overfit — failed every robustness gate decisively and unwaivably. Walk-forward flags is_overfitted=TRUE with IS Sharpe 3.07 collapsing to OOS -0.577 (windows 2.50, 2.17, -6.40), OOS <= 0, PBO 0.6442 > 0.5, and the untouched holdout at Sharpe -1.67 (ratio 0.0 vs 0.70 floor). The multiple-testing correction confirms best-of-N noise: deflated_sharpe 0.013 (vs 0.95) and the optimized full-sample Sharpe 1.15 sits below expected-max 2.30 over 225 trials — so the flattering optimized backtest (Sharpe 1.15, PF 1.80, +33%) is a search artifact, not a persistent edge. Sensitivity shows 2 cliff parameters (atr_trail_mult, donchian_lookback) and no robust plateau, so there is nothing stable to tune toward. This is the Donchian-breakout family (zero survivors, L9); the daily-trend gate that was the entire differentiator did not generalize (holdout -1.67, 2026 -3.6%). Per the overfit->abandon rule, re-pointing the same breakout mechanism at another instrument tends to re-overfit rather than reveal edge — and a near-duplicate sibling (SOL HL BTC-regime-gated momentum, similarity 0.758) was already abandoned for overfit. The developer can only change parameters, and walk-forward already proved no parameter set carries an OOS edge, so no articulable improvement path exists. Abandon rather than spend the final iteration re-overfitting a dead breakout signal.

Long-short 4H Donchian channel breakout on the SOLUSD Hyperliquid perpetual, confirmed by the 1-DAY trend. Enters long when the 4H close breaks above the prior 20-bar high AND the daily close is above its 50-day SMA; enters short on the mirror condition; stays flat whenever the two timeframes disagree. Exits on a 3x-ATR chandelier trail from the extreme reached since entry, or immediately when the daily trend flips against the position. Sizing risks 1% of equity across the initial stop distance with gross notional capped at 1x equity; leverage 1.0. Exactly three tunable parameters (donchian_lookback, daily_sma_window, atr_trail_mult); ATR window, risk fraction, notional cap and the breakout threshold are locked constants.

At backtest_review, weight the zero-survivor class prior heavily: demand OOS Sharpe clearly above zero and stable across walk-forward before optimizing; treat a single favorable full-history number skeptically.

UNFAVORABLE CLASS PRIOR (L7/L9). This is a pure-OHLCV Donchian channel breakout — one of the explicitly-named abandoned mechanism classes with zero survivors across the last 50 outcomes, and no OHLCV-only strategy has ever reached risk review. The differentiator is real (a dual-timeframe 'flat-on-disagreement' confirmation filter, the same construct that let the BTC sibling reach paper at Sharpe ~1.99, ported to a new venue/asset), so it is not a NAKED breakout — but the OOS-robustness burden is high and the sandbox is already net-negative.

Judge on the full ~2.3yr 4H history (~100+ trades) plus walk-forward; a low-win-rate/large-winner trend system needs multiple trend cycles to establish expectancy.

Sandbox is net-negative but NON-DECISIVE. avg_trade_return_pct -0.276% (below fee floor), total_return -2.7%, BUT this is a textbook trend-follower profile with FAVORABLE reward:risk (avg_win $1040 vs avg_loss $716 = 1.45x, positive skew +0.86, one $4884 capture) and a 37% win rate only marginally below the ~41% break-even. With 43 trades and Sharpe CI [-2.06, 1.26] straddling zero, the year is within noise — not the decisive negative signature of a doomed mechanism.

Rely on engine trade counts, not offline gate-open tallies; confirm the full-history realizable count clears the measurability floor.

Dry-run vs engine trade-count discrepancy. The rationale claims the gate 'fires 121 times' in the sandbox window, but the engine produced 43 actual trades — expected (offline counted gate-open bars; the engine cannot re-enter while in a position), but it means the offline projections (incl. 388-over-full-history) overcount realizable trades.

Favorable trend-follower geometry: avg_win $1,139 > avg_loss $721, profit_factor 1.26, expectancy +$104/trade, positive skew 0.80 — not the inverted-RR breakout death

avg_trade_return_pct 1.04% (~7x HL fee floor), max_drawdown 8.9% (controlled), 106 trades matching the hypothesis (49L/57S)

Trend-gated (flat-on-disagreement) variant — the one breakout archetype that has survived here (BTC dual-TF sibling Sharpe ~1.99, paper); profits long and short across 2024-2026

Fills under-represented HL venue + long-short buckets; clean implementation, ATR-trailing exit, no fee/data/verification structural traps

Sharpe only 0.493 with CI [-0.56, 1.47] straddling zero — edge not statistically established

Short HL history (~2.5yr/916 days, 106 trades) — thin walk-forward OOS windows, limited regime coverage, high overfitting risk (a same-class LINK sibling just failed at PBO 0.89)

Recent deterioration: 2026 YTD -4.4%, rolling Sharpe negative through mid-2026; returns lean on 2025 (+14.3%)

Clean execution and favorable trend-follower geometry in-sample (optimized avg_win $1,333 > avg_loss $629, PF 1.80, avg_trade_return_pct 2.14%)

Adequate trade count (111) and controlled optimized drawdown (7.3%)

Walk-forward overfitted: IS Sharpe 3.07 -> OOS -0.577, with the latest OOS window at -6.40

Holdout failed catastrophically (Sharpe -1.67, ratio 0.0 vs 0.70 floor) — the edge does not exist on untouched data

PBO 0.6442 (> 0.5) and deflated_sharpe 0.013 (far below 0.95): selected Sharpe indistinguishable from best-of-225-trials noise; optimized Sharpe 1.15 < expected-max 2.30

2 cliff parameters (atr_trail_mult, donchian_lookback) — no robust parameter plateau to tune toward

Donchian-breakout family with zero survivors; near-duplicate sibling (SOL HL BTC-regime-gated momentum, similarity 0.758) already abandoned for overfit

This strategy tried to replicate the factory's one surviving breakout archetype — trend-gated dual-timeframe momentum — by rebuilding it as a 4H Donchian breakout on SOL/Hyperliquid confirmed by the daily SMA trend, targeting the under-filled Hyperliquid venue and long-short direction buckets. The initial backtest was clean with attractive trend-follower geometry (PF 1.26, 1.04% per trade, 8.9% drawdown) but a statistically weak Sharpe of 0.49, so the analyst sent it to optimization rather than abandoning outright. Optimization then exposed it as a search artifact: in-sample Sharpe of 3.07 collapsed to -0.577 out-of-sample, the holdout came in at -1.67, PBO was 0.644 and deflated Sharpe just 0.013 across 225 trials, with two cliff parameters and no stable plateau. The analyst abandoned it as decisively overfit, noting a near-duplicate SOL/Hyperliquid sibling had already died the same way and that no parameter set carried a genuine out-of-sample edge.

Porting a surviving archetype onto a new asset/venue does not transfer the edge: with only ~2.5 years of Hyperliquid history and 106 trades, the Donchian-breakout family re-overfit under multiple-testing, and the daily-trend gate that was the sole differentiator failed to generalize out-of-sample.

The analyst abandoned it at the post-optimization ANALYZING stage after it failed every robustness gate: walk-forward flagged overfit (IS Sharpe 3.07 → OOS -0.577, worst window -6.40), the untouched holdout collapsed to Sharpe -1.67 (ratio 0.0 vs 0.70 floor), PBO 0.644 (>0.5), deflated Sharpe 0.013, and two cliff parameters with no robust plateau. Risk-review and promotion were never reached.

A long-short, single-instrument pure-OHLCV trend-continuation strategy on SOLUSD.HYPERLIQUID perps that took 4H Donchian channel breakouts only when the 1-day SMA trend agreed (flat on disagreement), exiting via an ATR chandelier trail — a deliberate port of the factory's surviving BTC dual-timeframe archetype onto the under-represented Hyperliquid venue and long-short bucket.

The initial backtest showed favorable trend-follower geometry (106 trades, avg_win $1,139 > avg_loss $721, profit factor 1.26, avg_trade_return_pct 1.04%, 8.9% max drawdown) but a weak Sharpe of 0.49 with a CI [-0.56, 1.47] straddling zero. Optimization produced a flattering full-sample result (Sharpe 1.15, PF 1.80, +33% return) that did not hold up out-of-sample.