EthHyperliquidDonchianBreakoutTrendLS

ETH Hyperliquid Channel-Breakout Trend Continuation — Long-Short 12H Donchian Breakouts Filtered by a Long SMA, Donchian-Midline Trailing Exit (Turtle-Style, 3-Parameter)

A LONG-SHORT, single-instrument, pure-price CHANNEL-BREAKOUT trend follower on ETH-USD perpetual on HYPERLIQUID (ETHUSD.HYPERLIQUID), 12H bars. This is a deliberate MECHANISM change away from the colliding SOL 4H dual-EMA-confluence + vol-regime family (which was abandoned fee_edge): instead of an EMA-cross confluence read every 4H, this enters on a genuine N-period Donchian high/low BREAKOUT — price is only committed when it actually prints a new extreme, so the strategy trades far fewer, higher-conviction continuation events and never pays to chase noise inside the channel. The breakout primitive plus a longer 12H hold materially separates it from the mid-frequency EMA scanner that died. It stays in the ONE directional-trend family that has produced a survivor here (paper-stage BTC dual-TF momentum, Sharpe 1.99) while filling the two largest portfolio gaps: HYPERLIQUID venue (6.5% vs 20% target) and long-short direction (13.6% vs push-toward-neutral). 12H on Hyperliquid has ~6+ years of usable candles (5000×12H ≈ full history back to 2020), so walk-forward windows and the 15-day holdout are well populated (avoids the empty-holdout / no-trades death) and it sits clear of the sub-hourly HL history trap. Risk profile: risk ~1.5% equity per trade via ATR-based initial stop; volatility-scaled sizing (notional = risk_budget / (atr_stop_mult * ATR_12H)); max single-position exposure ~25% of $100k; leverage capped at 2x (reads self.config.leverage). Exactly 3 tunable parameters (Donchian breakout length, long-SMA filter length, ATR stop multiple) to resist the 305 overfit deaths.

Implements the hypothesis exactly: single-instrument, pure-price, long-short channel breakout on ETHUSD.HYPERLIQUID 12-HOUR bars (verified present in the catalog, 2472 bars from 2023-01-01, so walk-forward windows and the 15-day holdout are well populated), with three tunables only - donchian_len, sma_len, atr_stop_mult. The breakout primitive (channel computed over the PRIOR N bars, current bar excluded) means the strategy only commits on a genuine new extreme rather than scanning for crosses every bar, which is the fee-drag separation from the abandoned EMA-confluence family. A dry run of the exact coded rule on the real 12H series (9 bps round-trip Hyperliquid taker) gives ~27 trades/yr and avg_trade_return_pct ~1.0% at the chosen defaults - far above the 0.15% fee floor - and the same sign holds across a wide plateau (donchian_len 8-14 x sma_len 60-100 all give 0.5-1.1% per trade), so the defaults were picked in the plateau interior rather than at the grid maximum, and the last-365-day window (which the sandbox and holdout sit in) produces 27 trades at +0.62% avg, so no zero-trade risk. Risk: 1.5% of equity per trade via the ATR stop with notional capped at 25% of equity; leverage is left at 1.0 and read from self.config.leverage in the notional cap - the ATR risk sizing never demands more than ~17-25% of equity notional, so declaring the hypothesis's 2x cap would have been inert (byte-identical backtest) and is therefore not set. One honest caveat: in the dry run the long leg carries the edge (~+2.6% avg/trade) while the short leg is slightly negative (~-0.7%), so the long-short mandate costs some expectancy; I implemented long-short as specified.

Marginal, statistically-unestablished breakout in the deadest class. Sharpe is only 0.49 with a bootstrap CI [-0.38, 1.40] that straddles zero, and against a meaningful buy-hold ETH benchmark alpha is just 0.038 with information_ratio -0.41 (it underperforms holding ETH). The data is narrow and single-regime: Hyperliquid ETH history only starts 2023-01 (~3.4 years, not the claimed ~6), so walk-forward/holdout have little regime diversity. It fires only 89 times (<100), and 2024's +14.6% is largely one day (2024-03-14 +14.07%) while 2023 (-3.2%) and 2026 YTD (-1.9%) are negative — only 2025 (+9.4%) is a distributed positive year. It is an OHLCV Donchian breakout (L9 trend/breakout, 0/213) on Hyperliquid (L41, 0/97 survival) — the stronger BTC Donchian sibling was just abandoned as an outlier mirage, and every BTC directional sibling this session (vol-regime, alt-breadth, regime-adaptive) failed the 3-phase run with deflated Sharpe ~0. This candidate is weaker (CI straddles zero) than the directional strategies that were sent to optimize, so best-of-N over ~225 trials on an 89-trade single-regime series will not clear deflated Sharpe. The clean 6.4% drawdown and moderate kurtosis are real but do not overcome an unestablished edge in a 0/97+0/213 class. Abandon rather than spend the optimization budget.

Long-short turtle-style Donchian channel breakout on ETH-USD perpetual at Hyperliquid, 12H bars. Enters LONG when the close prints above the highest high of the prior 12 bars AND sits above an 80-bar SMA; SHORT on the mirror condition. Exits on whichever comes first: a hard ATR initial stop (3 x ATR(14) adverse from entry) or the close crossing back through the current Donchian midline (a trailing exit that rises/falls with the channel). calculate_signal returns the continuous normalized channel position (close - mid)/(0.5*(upper-lower)), which is 0 at the midline, +-1 at the prior channel edges and beyond +-1 exactly on a breakout. Sizing is risk-first: qty = equity*1.5% / (3*ATR), so a fixed ~1.5% of equity is at risk to the initial stop in every regime, with notional hard-capped at 25% of equity x configured leverage.

Negative sandbox (-2.71%, Sharpe -0.57, PF 0.61, avg_trade_return_pct -0.48%); dry-run (+1.0%/trade) contradicted, and the developer's claimed edge-carrying LONG leg lost (long_win_rate 0.125) in the anti-trend recent window. Regime-dependent breakout failure, not a code bug.

Donchian/breakout pure-OHLCV mechanism class with zero-survivor base rate (L7/L9); developer's own dry-run shows the short leg has negative expectancy. Research-Lead/Analyst call.

History is ~3.4yr (not the ~6 claimed), 17 sandbox trades, Sharpe CI [-2.55, 1.36] straddles zero. Holdout not empty but thin.

Clean, simple 3-parameter design with genuinely controlled risk: max_drawdown 6.4%, return_kurtosis 7.35 and skew 0.65 (far less outlier-driven than the BTC Donchian sibling)

profit_factor 1.50, avg_trade_return_pct 1.24% (clears fees); 2025 (+9.4%) is a distributed positive year; fills HL/long-short buckets

Edge not significant: Sharpe 0.49 with CI [-0.38, 1.40] straddling zero; alpha only 0.038 and information_ratio -0.41 vs a meaningful buy-hold ETH benchmark (underperforms holding ETH)

Narrow single-regime data: Hyperliquid ETH only from 2023-01 (~3.4 years), not the claimed ~6 years — thin walk-forward/holdout diversity

89 trades (< 100); 2024's +14.6% is largely one day (2024-03-14 +14.07%); 2023 -3.2% and 2026 -1.9% negative

Deadest class: OHLCV Donchian breakout (L9, 0/213) on Hyperliquid (L41, 0/97); the stronger BTC Donchian sibling and every BTC directional sibling this session failed as mirages/OOS collapses

EthHyperliquidDonchianBreakoutTrendLS changed mechanism to a channel breakout on Hyperliquid ETH, entering only on genuine new extremes filtered by a long SMA, and it produced the cleanest risk profile of its cohort — +24.4%, PF 1.50, 1.24% per trade, a 6.4% drawdown, and modest kurtosis, unlike the outlier-mirage BTC Donchian sibling. But the edge was not statistically established: Sharpe 0.49 with a confidence interval spanning zero, negative alpha and information ratio versus holding ETH, only 89 trades over ~3.4 years of single-regime Hyperliquid history, and 2024's gain concentrated in one day. The analyst abandoned it at backtest review, judging it weaker than the directional siblings that had already failed the 3-phase run with deflated Sharpe near zero, and in a 0/213 × 0/97 survival class where best-of-N optimization would not clear deflated Sharpe. It never reached optimization, analysis, or risk review.

A genuinely clean risk profile (6.4% drawdown, low kurtosis) is not enough to justify optimization when the Sharpe CI straddles zero, the sample is below 100 trades in a single short-history regime, and the strategy belongs to a class with zero prior survival — the edge must be statistically established before best-of-N optimization can be trusted.

The analyst abandoned it at backtest review as a marginal, statistically-unestablished breakout in the deadest class: Sharpe CI straddles zero, it underperforms holding ETH risk-adjusted, the Hyperliquid data is narrow and single-regime (only ~3.4 years from 2023, not the claimed ~6), it fires only 89 times (<100), and it sits in the OHLCV Donchian (0/213) × Hyperliquid (0/97) survival class — weaker than the BTC directional siblings that had just collapsed with deflated Sharpe ~0 in optimization.

A long-short, single-instrument Donchian channel breakout trend-follower on ETHUSD.HYPERLIQUID 12H bars (3 parameters) that entered on a genuine N-period channel breakout confirmed by a long SMA filter, exited on the Donchian midline or an ATR initial stop, with risk-based volatility-scaled sizing — a deliberate mechanism change to the breakout primitive on an under-represented venue.

The backtest (ETHUSD.HYPERLIQUID 12H, 1235 data days from 2023-01) returned +24.4% with 89 balanced trades (46L/43S), profit factor 1.50, avg_trade_return_pct 1.24%, and a notably clean risk profile (max drawdown 6.4%, kurtosis 7.35, skew 0.65 — far less outlier-driven than its BTC Donchian sibling). But Sharpe was only 0.49 (CI [-0.38, 1.40] straddling zero), alpha just 0.038 with information ratio -0.41 versus holding ETH, and 2024's +14.6% leaned largely on one day.