HLBTCDailyDonchianBreakoutLongFlat

Hyperliquid BTC Daily Donchian Breakout, LONG/FLAT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Simple Fixed Single-Entry Sizing, No Flips, Chosen Because HL Sizes Correctly While BINANCE USD-M Perp Is Engine-Corrupted, Tail-Safe, Low-Parameter)

A LONG-ONLY (LONG/FLAT, never short), SINGLE-INSTRUMENT, DAILY Donchian-channel breakout trend strategy on BTCUSD.HYPERLIQUID. Venue choice is the entire point: the factory's backtest engine currently has a CONFIRMED, strategy-unfixable OVER-EXPOSURE defect specific to BINANCE USD-M perp (proven immune to hard caps + dual flat-guards across SOL vol-targeted 345% fresh run, ETH vol-targeted 246%, SOL/BTC funding 252%/106%, BTC MTF 136% — all corrupted/near-ruin), PLUS a BINANCE spot-CASH held-position MTM defect (unbacktestable), a multi-instrument aggregation defect, and a forced-flow/positioning data wall (liquidations/OI/account-ratio/taker). HOWEVER, the one HYPERLIQUID run this session (the multi-instrument funding-contrarian) sized CORRECTLY at ~20-29% capital with metrics_reliable=true — it failed only on HL's short data, NOT on over-exposure. That evidence localizes the over-exposure bug to the BINANCE USD-M venue/account path and identifies HYPERLIQUID as the one perp venue that currently produces TRUSTWORTHY sizing/metrics. This hypothesis therefore moves the factory's VALIDATED daily trend edge to HL, in the simplest, cleanest, most-evaluable form: single-instrument, OHLCV-ONLY (the deep wall-free feed), LONG/FLAT (no shorts -> no flip-stacking, belt-and-suspenders), ONE fixed-fractional position size set at entry with NO dynamic resizing, leverage 1.0, close-to-flat before re-entry. It uses BTC, the cleanest core trender, on the most UNDER-REPRESENTED venue (HL 6.3% vs 20% target). PRE-REGISTERED ACCEPTANCE / SIZING CHECK: avg_position_pct ~<=100% and single-day returns within ~+/-15% (BTC at 1x) — if HL also over-exposes, the defect is engine-wide (escalate); if it sizes correctly (as prior HL evidence suggests), this is the first cleanly-evaluable trend run of the recent batch. Low parameter count.

Implements the hypothesis exactly: moves the factory's validated daily trend edge (Donchian/Turtle breakout) to HYPERLIQUID, the one perp venue whose engine path produced trustworthy sizing/metrics this session, while the BINANCE USD-M path remains over-exposure-corrupted. Kept maximally clean and evaluable: single instrument (BTCUSD.HYPERLIQUID), OHLCV-only (the deep, wall-free feed), LONG-ONLY (should_enter only ever returns BUY; SELL is used solely to exit via the base class), low parameter count, and ONE fixed-fractional 95%-equity size set at entry with no resizing and leverage=1.0 (position_size never multiplies by leverage, so sizing matches the venue margin model and honors the pre-registered avg_position_pct<=100% / +/-15% single-day sizing check). The signal is continuous and varies every bar (breakout strength in ATR units) rather than being gated behind the entry condition, so the verifier sees graded conviction. Daily bars match the long-term timeframe; the 20/10 entry/exit Donchian asymmetry plus a 3xATR stop and 60-bar time-stop give well-defined, fee-viable per-trade moves well above the ~0.09% HL round-trip taker cost. If HL also over-exposes, the defect is engine-wide and should be escalated; if it sizes correctly (as prior HL evidence suggests), this is a cleanly-evaluable trend run on the most under-represented venue.

Failed deflated Sharpe: DSR=0.0, expected-max=9.35 over 225 trials — the optimized Sharpe (1.15) is far below the best-of-N luck bar and statistically indistinguishable from selection noise (is_significant=false, PBO=0.746>0.5, sharpe_ci_low -1.90 straddles 0). Every generalization gate fails in concert even though SENSITIVITY PASSED (1 cliff, healthy heatmap): the walk-forward out-of-sample Sharpes are [0, 0, 0] against in-sample [0, 6.54, 10.09] with is_overfitted=TRUE (the entire ~5.5 avg IS Sharpe vanishes OOS), and the holdout FAILED with a NEGATIVE holdout_sharpe -1.63 (ratio 0) — the once-tested recent window loses money. Sensitivity passing only confirms full-period in-sample stability; it does not measure generalization, and the generalization gates (OOS, holdout, DSR) all fail. Root cause (symptom, not independently verified): the BTC-HL trend edge is concentrated in the 2023-2024 bull (optimized annual +30%/+81%) and has decayed/reversed in 2025-2026 (-18.8%/-8.2%), on only ~3 years of Hyperliquid history with 32-35 trades — the OOS walk-forward windows are empty/edgeless and the recent holdout is negative, so the sample is too short and too regime-concentrated to survive forward-testing or multiple-testing deflation. The attractive optimized headline (PF 1.63, +88.9%) is an in-sample best-of-225 fit, not a validated edge. Not iterate (attempt 1 of 2): the failure is OOS non-generalization + deflation + a negative holdout, not a tunable parameter region — sensitivity already passed, so a second sweep would re-find the same IS-positive / OOS-zero / negative-holdout pattern on the same thin data. Not revise_hypothesis: the Donchian breakout is being validated per-instrument and BTC-HL was the best candidate; it fails purely because HL's ~3-year history is too short to validate a low-frequency daily-trend edge, an engine/data limitation (the deep-history venues remain blocked by the USD-M over-exposure defect), not a premise a different instrument reframes. FAILURE PATTERN: single-instrument daily Donchian breakout on Hyperliquid majors cannot be validated on HL's ~3-year history — even the strongest sibling (BTC, PF 1.36 initial, sensitivity-passing) collapses to all-zero walk-forward OOS, a negative holdout (-1.63), DSR 0.0 and PBO 0.75 because the trend edge concentrates in 2023-2024 and the thin/regime-concentrated sample defeats forward-testing and multiple-testing deflation. The clean-sizing finding (HL sizes correctly, metrics_reliable=true) is preserved; the daily-trend edge is real in-sample but unvalidatable on this venue's short data.

LONG/FLAT daily Donchian-channel breakout on BTCUSD.HYPERLIQUID (single instrument, pure OHLCV, leverage 1.0). Enters long when the daily close prints a new 20-day high (continuous breakout strength = (close - prior 20-day high)/ATR, thresholded at a small ATR buffer). Exits to flat when the close prints a new 10-day low, an ATR-based protective stop (3xATR below entry) trips, or a 60-bar time-stop fires. One fixed-fractional position (95% of equity notional) is set once at entry with no dynamic resizing; the strategy is always fully flat before any re-entry, so there is no flip-stacking and exposure stays <=100% of capital.

Pre-registered sizing check PASSES: avg_position_pct 55.3% (<100%), exposure_pct 47.7%, metrics_reliable=true, end_unrealized_pct 10.9% (return is mostly realized) — no over-exposure/position-stacking artifact. Second clean Hyperliquid data point.

Genuinely profitable default run with correct trend-following payoff asymmetry: avg_win 23,790 vs avg_loss 11,686 (~2:1), profit_factor 1.36, total_return +61%, Sortino 2.44, Calmar 1.41.

Sound, simple implementation: long/flat (no flip-stacking), Donchian bands exclude current bar (no look-ahead), ATR stop + time-stop, leverage 1.0, one fixed-fractional entry.

Clearly stronger than its ETH HL sibling (which had symmetric payoff and PF 0.65) — this is the better of the two to invest optimization time in.

Sharpe 0.92 but CI -2.17 to +2.35 straddles 0 — significance not established on the thin sample.

Only ~3 effective years / 35 trades (no Hyperliquid BTC data pre-2023; 2020-2022 annual = 0.0), so walk-forward windows + 15-day holdout will be thin.

Recent-year decay: 2023 +40.8% and 2024 +38.5% but 2025 -17.9% and 2026 -6.2% — the edge concentrated in 2023-2024; recent OOS is the real test.

Minor internal metric inconsistencies (expectancy -3577 vs avg_trade_return +1431; avg_win/avg_loss x trade counts don't cleanly reconcile with PF 1.36) — worth verifying, though headline metrics agree on 'profitable' and metrics_reliable=true.

max_drawdown 33% with a 518-day drawdown duration — meaningful pain even at the profitable default config.

Clean sizing confirmed across both runs (avg_position_pct 55%/52%, exposure ~48-52%, metrics_reliable=true) — reinforces Hyperliquid as the trustworthy-sizing venue while BINANCE USD-M is corrupted.

Sensitivity PASSED (1 cliff; heatmap a stable 0.6-1.33) — the full-period parameters are not fragile.

Initial and optimized full-period backtests look attractive in isolation (PF 1.36->1.63, ~2:1 win/loss asymmetry, Sharpe 0.92->1.15) — genuine trend behavior in 2023-2024.

Failed deflated Sharpe: DSR 0.0 (<<0.95), is_significant=false; optimized Sharpe 1.15 is far below the expected-max luck bar 9.35 over 225 trials.

PBO 0.746 (>0.5): parameter selection more likely than not overfit; sharpe_ci_low -1.90 straddles 0.

Holdout FAILED: holdout_sharpe -1.63, ratio 0 — the once-tested recent window is NEGATIVE.

Walk-forward OOS = [0, 0, 0] with IS [0, 6.54, 10.09] and is_overfitted=TRUE — the entire in-sample Sharpe vanishes out-of-sample.

Edge concentrated in 2023-2024 (+30%/+81%) and decayed/reversed in 2025-2026 (-18.8%/-8.2%); only ~3 years of HL history and 32-35 trades — too thin and regime-concentrated to validate a low-frequency daily-trend edge.

HLBTCDailyDonchianBreakoutLongFlat was the strongest of the daily-trend ports — a genuinely profitable BTC breakout (+61%, PF 1.36, clean ~2:1 win/loss asymmetry) whose sizing check passed, reinforcing Hyperliquid as the trustworthy-sizing venue while Binance USD-M remains corrupted. The reviewer judged it the best candidate to optimize, and optimization both improved the headline (PF 1.63, +88.9%) and passed sensitivity. But the generalization gates failed in concert: deflated Sharpe 0.0, PBO 0.75, an all-zero walk-forward OOS, and a negative -1.63 holdout, because the edge was concentrated in 2023–2024 and decayed in 2025–2026 over a sample of only ~3 years and 32–35 trades. The analyst abandoned it as a real in-sample edge that is simply unvalidatable on this venue's short history — an engine/data limitation rather than a design flaw — while preserving the clean-sizing finding.

Passing sensitivity proves in-sample parameter stability, not generalization — a low-frequency daily-trend edge cannot be validated on Hyperliquid's ~3-year history when the edge concentrates in the 2023–2024 bull and the holdout and walk-forward OOS windows turn negative or empty.

Even with sensitivity passing, the analyst abandoned it (iteration 1 of 2) because every generalization gate failed: deflated Sharpe 0.0 vs an expected-max-luck bar of 9.35, PBO 0.746, sharpe_ci_low -1.90, walk-forward OOS of [0,0,0] against in-sample [0,6.54,10.09] with is_overfitted=TRUE, and a negative holdout (-1.63).

A long/flat, single-instrument, OHLCV-only daily Donchian/Turtle-style breakout on BTCUSD.HYPERLIQUID (ATR stop, time-stop, leverage 1.0, one fixed-fractional entry), run on Hyperliquid specifically because it sizes cleanly while the Binance USD-M perp path has a confirmed over-exposure defect.

The default backtest was genuinely profitable with correct trend asymmetry: +61% total return, Sharpe 0.92, profit factor 1.36, avg_win ~$23.8k vs avg_loss ~$11.7k (~2:1) over 35 trades at 23% win rate and 33% max drawdown; optimization lifted it to Sharpe 1.15, PF 1.63, +88.9% and passed the sensitivity gate (1 cliff, stable heatmap), and the pre-registered sizing check passed (avg_position_pct 55%/52%, metrics_reliable=true).