HyperliquidBtcDailyDonchianTrendLS

Hyperliquid BTC Daily Donchian Breakout Trend-Follower — Ride Confirmed Trends With a Chandelier Trailing Stop, Vol-Targeted Sizing, on the UNDER-REPRESENTED Hyperliquid Venue at the DAILY Timeframe (Full History, Long-Short, Asymmetric Let-Winners-Run, 3-Parameter)

A LONG-SHORT, single-instrument, ASYMMETRIC-PAYOFF trend-follower on BTCUSD.HYPERLIQUID (HL perpetual) at the DAILY timeframe. It is anchored to the only pattern that has ever survived this factory's gates: directional trend with a trailing stop that lets winners run (the sole promoted strategy in 1163 experiments). Every symmetric mean-reversion / contrarian-fade / market-neutral / pairs sleeve this session died fee_edge or no_edge — that door is closed. This sleeve instead makes two DELIBERATE diversification moves that keep it out of the L56 'same mechanism, new ticker' spray trap and out of the pending Binance trend sleeves: (1) VENUE — it runs on HYPERLIQUID, the most under-represented actionable venue (6.7% vs a ≥20% target; options, the other gap, are currently infra-broken at Layer 3), directly testing whether the BTC trend edge generalizes to HL's different taker/liquidity microstructure; (2) TIMEFRAME — it uses DAILY bars, which per the HL data limits carry FULL history back to 2020 (unlike HL sub-hourly, which is history-capped and gets abandoned for insufficient data), so there is no measurability problem, and daily is a distinct horizon from every pending 4H sleeve. Mechanism: a simple Donchian channel breakout entry with a chandelier trailing-stop exit and volatility-targeted sizing — pure OHLCV, computationally TRIVIAL (rolling max/min + ATR, a few scalars per bar, no full-history rescans, no sorted()/min() in the hot path). It avoids every graveyard: NOT symmetric reversion (dead), NOT funding carry/differential (L57), NOT pairs/cross-sectional baskets (die sub-DSR), NOT options (infra-broken), NOT a liquidation feed (L58). Exactly 3 tunable parameters: Donchian lookback, chandelier ATR multiple, and the annualized volatility target.

ITERATION 2 - fixes the single reported failure (Layer 4 QA: unmeasurable trade count) with the smallest change that can fix it, and touches nothing that already passed. Layers 1-3 passed on this code, so signal/entry/exit/sizing logic is byte-identical; the diff is (a) the bar interval in the config, 1-DAY -> 4-HOUR, and (b) two structural constants rescaled so they keep their meaning in TIME rather than in bars - vol_lookback 30 -> 180 bars and min_vol_obs 20 -> 60, i.e. still ~30 days / ~10 days of returns behind the volatility estimate (the vol_lookback clamp was widened to [10,400] and _param_bounds updated to match, so no silent clamp). No import, class-structure or signal change; no banned imports. QA was right that daily is unfixable within the 3 tunables, so the timeframe is the only lever: on the same HL BTCUSD catalog, 4-HOUR carries 5,574 bars from 2024-01-22 to 2026-08-08 and is CURRENT (12H/8H are stale to 2026-05 and only ~20/yr; 1-HOUR and finer start 2025-10, too short) - 4H is the coarsest HL interval that is both measurable and inside the 5000-candle cap. Re-running the submitted code bar-by-bar on that real HL series with HL fees (0.045%/side): 142 trades (~56/yr) vs 56 (~9/yr) before - past the ~100-trade floor - with ~47 trades in each of three equal walk-forward windows, 60 in the last 365 days and 2 in the final 15 days, so the holdout is populated instead of empty. Economics: +0.31% mean NET per trade (above the 0.15% HL floor, though materially below the daily version's +4.30% - shorter holds mean smaller per-trade moves), 36.6% win rate, best +26.4% vs worst -8.2% (the right-tail asymmetry survives), compounded equity +26% with Sharpe 0.69 and max drawdown 16.3%. The second QA note also resolves itself: at 4H the SHORT leg is no longer the weak half - long +0.31%/trade (74 trades) vs short +0.31% (68 trades) - so the long-short construction is now genuinely two-sided rather than a long-only edge in disguise. Honest caveats: the 4H sample is 2.5 years, not 6, so the 2020-2021 trend era is gone and the remaining sample is 2024 strong (+1.42%/trade), 2025 (-0.22%) and 2026 (-0.23%) weak per-trade with the equity path still up - the recent-regime walk-forward window is the main risk, and the per-trade cushion over fees is now thin enough that slippage assumptions matter. Parameter defaults were deliberately left UNCHANGED (20 / 3.0 / 0.20) rather than re-tuned to the new timeframe, so this is a clean timeframe test and not a refit; the surrounding 4H plateau (lookback 15-40, chandelier 2.5-3.0) is positive throughout, Sharpe 0.63-0.82.

Marginal, impact-crippled, toy-capacity trend sleeve far below the bar. The daily full-history premise was abandoned; the 4H fallback covers only 2.5 years (HL's 5000-candle cap), and on that sample the edge is not measurable: Sharpe 0.164 with sharpe_ci_low -0.98 (CI straddles zero), profit_factor 1.048 (fee-fragile breakeven band, L18), avg_trade_return_pct 0.155% grazing the 0.15% HL floor. Decisively, impact_cost_pct is 33.4% with capacity only $896K — market impact consumes a third of gross and the edge exists only at toy scale, a standalone promotion blocker that would go negative on real fills. 2025 is -8.2% with rolling Sharpe to -6.5. This is a far weaker version of the Binance breakout siblings (Sharpe 0.16 vs 0.62/0.65, PF 1.05 vs 1.52/1.36, capacity $896K vs $258M/$77M), one of which (ETH OI breakout, base Sharpe 0.62) just failed the full 3-phase run catastrophically (holdout -1.05, DSR 0.0). A 0.16 base Sharpe cannot clear the deflated-Sharpe gate after best-of-225 selection, and there is no path from a $896K-capacity, 33%-impact, breakeven-PF result to a promotable one. Abandon at BACKTEST_REVIEW rather than spend the optimization budget. Failure pattern: hl_trend_toy_capacity_fee_fragile.

Long-short trend-follower on BTCUSD.HYPERLIQUID (HL perpetual), now on 4-HOUR bars (iteration 2 moved it off 1-DAY purely for measurability). Each bar it computes a continuous range-position signal (close - mid) / (0.5 * (upper - lower)) from the Donchian channel of the PRIOR donchian_lookback bars (current bar excluded, no look-ahead): ~0 mid-channel, +-1 exactly at the channel edge, beyond +-1 on a breakout. Entry is a channel breakout - BUY when signal >= break_thresh, SELL when signal <= -break_thresh. There is NO take-profit: the sole exit is a CHANDELIER trailing stop (highest high since entry minus chandelier_atr x ATR for longs, mirrored for shorts), so losers are cut near chandelier_atr ATR while winners run. Sizing is VOLATILITY-TARGETED: notional = equity * min(vol_target / realized_annualized_vol, max_notional_frac), where realized vol is the stdev of the last vol_lookback bar log returns annualized by the bar interval (bars/year derived from the bar type, 2190 for 4H). Notional is hard-capped at 1x equity so leverage stays 1.0 (measured 0.37x average, 0.65x max). Pure OHLCV - no supplementary feed, so no data-availability failure mode; all state is incremental (Wilder ATR, running return sums, bounded deques) and a full 5,574-bar run takes 0.06s.

Verification failed (Layer 4 — QA review): - UNMEASURABLE trade count (L16). The developer's own full-history dry run is 56 trades over ~6 years (~9/yr), and the sandbox produced only 9 trades in the year (total_return -0.79%, Sharpe 0.033, CI [-1.55, 1.53] — indistinguishable from zero). 56 trades over the whole backtest span is well below the ~100-trade measurability floor; per L16 a sub-100-trade result cannot distinguish edge from noise no matter how large the per-trade return looks (+4.30% full-history / +0.27% sandbox). Critically, the optimization pipeline cannot be populated at this frequency: the deflated-Sharpe gate (DSR>=0.95) is effectively unachievable on 56 wide-CI trades (the developer concedes 'deflated-Sharpe will be demanding'), each of the 3 walk-forward OOS windows gets ~18 trades, and the 15-day holdout gets ~0.4 trades — untestable. Daily bars on a single asset structurally produce this, and it is NOT fixable within the 3 tunables (a longer Donchian lookback only lowers the count further). - Short-leg weakness (performance flag, not a code defect). The developer reports the short leg is flat-to-negative on the 2020-2026 sample (-0.76%/trade over 26 trades) against a strongly positive long leg (+8.69% over 30) — expected on a structurally long-biased BTC sample. The code faithfully implements the stated long-short hypothesis, so this is not a mismatch, but on top of the tiny sample the short leg is essentially untested.

At BACKTEST_REVIEW, look specifically at whether ANY edge exists outside 2024. If the 2.5-year sample is 2024-only with 2025-2026 negative (as the developer reports), ABANDON at review rather than spend the optimization budget — a single-year-concentrated edge on a 2.5-year sample will not survive walk-forward/PBO/DSR.

Edge durability is the decisive open question, and it is weak — a strong BACKTEST_REVIEW concern, borderline for a block. The iteration-2 timeframe fix genuinely resolved the trade-count problem (142 full-history trades vs 56, past the ~100 floor; holdout now populated). But the full HL 4H catalog is only ~2.5 years (2024-01 to 2026-08 — HL's 5000-candle cap), and the developer's own per-year breakdown shows the positive average is ENTIRELY 2024 (+1.42%/trade), with 2025 (-0.22%) and 2026 (-0.23%) both NEGATIVE. So ~60% of the sample and all recent data is negative, and the full-history +0.31%/trade is a one-year artifact rather than a durable edge. The sandbox (recent year) confirms it: avg_trade_return_pct -0.17% (below floor; the +3.78% dollar return and PF 1.08 come only from winners carrying larger notional). The optimizer's walk-forward OOS windows and 15-day holdout all sit in the negative 2025-2026 regime, so OOS>=0.5 / holdout / deflated-Sharpe are very likely to fail.

Gate fee-viability on the per-trade %, not the equity curve, which overstates the per-trade edge here.

Metric ambiguity the analyst should note: sandbox avg_trade_return_pct is -0.17% (equal-weighted per-trade, the fee-viability metric) while avg_trade_pnl_usd is +65.83 (dollar-weighted). The gap is the vol-targeting sizing — larger positions on lower-vol (tighter-stop) entries — so the dollar equity path can be up while the average trade is a fee-losing trade.

Clean asymmetric-trend construction in the one living family; genuinely uncorrelated (beta 0.007, alpha 0.039); 143 trades is an adequate count

Tests the under-represented Hyperliquid venue

Sharpe 0.164 with sharpe_ci_low -0.98 — CI massively straddles zero, no measurable edge; cannot clear the deflated-Sharpe gate after best-of-225 selection

profit_factor 1.048 (fee-fragile breakeven band, L18); avg_trade_return_pct 0.155% only grazes the 0.15% HL floor

impact_cost_pct 33.4% with capacity only $896K — impact eats a third of gross and the edge is realizable only at toy scale (standalone promotion blocker)

Only 2.5 years of data (HL 4H cap) after the daily full-history premise was abandoned; 2025 negative (-8.2%), rolling Sharpe to -6.5 — far weaker than the Binance breakout siblings (Sharpe 0.16 vs 0.62/0.65, PF 1.05 vs 1.52/1.36, capacity $896K vs $258M/$77M)

HyperliquidBtcDailyDonchianTrendLS tried to extend the factory's one living pattern — asymmetric trend with a chandelier trailing stop — to the under-represented Hyperliquid venue, first at daily (dropped after failing the ~100-trade floor) then at 4H, testing whether the BTC trend edge generalizes to HL's microstructure. It did not: constrained to 2.5 years by HL's candle cap, it managed only +11.2% over 143 trades with Sharpe 0.164, a CI floor at -0.98, breakeven profit factor 1.048, and — decisively — 33.4% impact cost on just $896K capacity, meaning the sliver of edge is realizable only at toy scale and negative on real fills. The analyst abandoned it at backtest review as a far weaker sibling of the Binance breakouts (whose own OI-confirmed version had just failed the full run at holdout -1.05, DSR 0.0), with no path from a $896K-capacity, breakeven-PF, near-zero-Sharpe result to a promotable one. It never reached optimization, analysis, or risk review.

Porting the surviving trend construction to the under-represented Hyperliquid venue does not generalize: the 4H history cap yields too short a sample to measure edge, and the venue's thin liquidity produces 33% impact cost on $896K capacity — a toy-scale, fee-fragile result that confirms venue/timeframe diversification cannot rescue a breakout edge that is already marginal on the deeper Binance data.

The analyst abandoned it at backtest review as a marginal, impact-crippled, toy-capacity trend sleeve far below the bar: with only 2.5 years of data the edge is not measurable (Sharpe 0.164, CI straddling zero, breakeven PF 1.048), and decisively market impact consumes a third of gross with $896K capacity so the edge exists only at toy scale and would go negative on real fills — a far weaker version of the Binance breakout siblings (Sharpe 0.16 vs 0.62/0.65, capacity $896K vs $258M/$77M), one of which had just failed the full 3-phase run catastrophically, so a 0.16 base Sharpe cannot clear deflated Sharpe after best-of-N.

A long-short, single-instrument, asymmetric-payoff trend-follower on BTCUSD.HYPERLIQUID perp — a Donchian channel breakout entry with a chandelier trailing-stop exit (no fixed take-profit, letting winners run) and volatility-targeted sizing — anchored to the factory's only surviving pattern (directional trend with a trailing stop), and deliberately diversifying by venue (the under-represented Hyperliquid) and timeframe (originally daily for full history, dropped to 4H after the daily version failed the ~100-trade measurability floor), with 3 tunable parameters.

The 4H backtest (5576 bars, only 2.5 years from 2024-01 due to Hyperliquid's 5000-candle cap) was marginal: +11.2% over 143 trades, Sharpe just 0.164 with sharpe_ci_low -0.98 (CI massively straddles zero), profit factor 1.048 (fee-fragile breakeven), avg_trade_return_pct 0.155% (grazing the 0.15% HL floor), max drawdown 20.6%, uncorrelated (beta 0.007), but impact_cost_pct 33.4% with capacity only $896K, and 2025 negative (-8.2%, rolling Sharpe to -6.5).