ETHHyperliquidDailyDonchianBreakoutLongFlat

Hyperliquid ETH Daily Donchian Breakout, LONG/FLAT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Simple Fixed Single-Entry Sizing, No Flips, Clean-Sizing Venue 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 ETHUSD.HYPERLIQUID. Venue choice is deliberate and evidence-driven: the factory backtest engine has a CONFIRMED over-exposure defect SPECIFIC to BINANCE USD-M perp (proven immune to hard caps + dual flat-guards across SOL/ETH vol-targeted, SOL/BTC funding-confirmed, BTC MTF — all corrupted to 100-345% exposure / near-ruin), plus a BINANCE spot-CASH MTM defect, a multi-instrument aggregation defect, and a forced-flow/positioning data wall (liquidations/OI/account-ratio/taker). The ONLY clean evidence point this session is HYPERLIQUID: the prior HL run sized correctly (~20-29% capital, metrics_reliable=true) and failed only on short data, NOT over-exposure — localizing the bug to the BINANCE USD-M path and identifying HL as the one venue currently producing trustworthy sizing/metrics. This is the ETH sibling of the pending HL BTC long/flat Donchian, extending the factory's VALIDATED daily trend edge across the two deepest HL majors on the clean-sizing venue. Construction is the simplest, most-evaluable form: single-instrument, OHLCV-ONLY (deep wall-free feed), LONG/FLAT (no shorts -> no flip-stacking), ONE fixed-fractional position size set at entry with NO dynamic resizing, leverage 1.0, close-to-flat before re-entry. ETH is the second-deepest HL market (more history/liquidity than the alts) and a validated core trender; HL is the most UNDER-REPRESENTED venue (6.3% vs 20% target). PRE-REGISTERED ACCEPTANCE / SIZING CHECK: avg_position_pct ~<=100% and single-day returns within ~+/-20% (ETH at 1x) — prior HL sizing evidence (~20-29%) suggests it will pass; if it over-exposes, the defect is engine-wide (escalate). Low parameter count.

Implements the hypothesis exactly: the simplest, most-evaluable form of the factory's validated daily-trend edge, extended to the second-deepest Hyperliquid major (ETH) on the clean-sizing venue. (1) Venue HYPERLIQUID per the hypothesis — the only venue this session producing trustworthy sizing/metrics; long-only-no-leverage would normally route to spot but HL perp is chosen deliberately to avoid the BINANCE USD-M over-exposure defect. (2) Single-instrument, OHLCV-only — no supplementary data, no extra legs, avoiding the data-wall and multi-instrument aggregation defects. (3) LONG/FLAT only — should_enter returns BUY or None, never SELL, so no flip-stacking; base class only enters while flat -> single entry, close-to-flat before re-entry. (4) Fixed-fractional single-entry sizing at 95% of equity, leverage 1.0, no dynamic resizing -> avg_position_pct ~<=95%, satisfying the pre-registered sizing/acceptance check (within ~+/-20% single-day returns at 1x). (5) Continuous Donchian channel-position signal varies every bar with the threshold (>1 = breakout above upper band) in should_enter, per the continuous-signal requirement. (6) Low parameter count (entry/exit lookbacks, threshold, position_pct, stop). Daily bars match the validated daily-trend timeframe; the 20-day breakout entry / 10-day breakdown exit is the canonical Donchian trend configuration, with per-trade moves (multi-percent daily breakouts) far exceeding the ~0.09% HL round-trip taker cost.

Failed deflated Sharpe: DSR=0.0, expected-max=8.858 over 225 trials — the optimized Sharpe (1.70) is far below the best-of-N luck bar and statistically indistinguishable from selection noise (is_significant=false, PBO=0.7619>0.5, sharpe_ci_low -2.53 straddles 0). Every decisive generalization gate fails in concert: the holdout FAILED (holdout_sharpe -0.85, ratio -1.27 — the once-tested recent window is negative); the walk-forward has 2 of 3 zero-trade windows (IS [0,0,5.69] / OOS [0,0,2.01]) so the entire edge sits in a single 2024 window and is_overfitted=false is an averaging artifact over empty windows; and sensitivity FAILED with 2 cliffs (entry_lookback, entry_threshold) over a heatmap that is negative Sharpe across nearly the whole entry_threshold x entry_lookback grid — there is NO robust profitable region to tune toward. The superficially attractive optimized backtest (Sharpe 1.70, PF 2.58, +28.7%, DD 13.5%) rests on only 14 trades / 81 effective days and is a best-of-225 over-selection, not a validated edge. The initial run was already a net loser (PF 0.65) with avg_win ~= avg_loss — no trend-capture asymmetry — and the sweep confirms that asymmetry exists nowhere in the parameter space for ETH-HL. Not iterate (attempt 1 of 2): with a negative holdout, DSR 0.0, PBO 0.76, two empty walk-forward windows, and an edgeless sensitivity grid, there is no region to tune toward — a second sweep would just re-select another best-of-N artifact. Not revise_hypothesis: this is overfitting + genuine no-edge on this instrument, not a proven mechanism stranded on a dead target; the Donchian breakout is being validated per-instrument and its working expression is the BTC-HL sibling (PF 1.76, advancing), so ETH-HL is simply the negative arm of that asset-selectivity test. FAILURE PATTERN: single-instrument daily Donchian breakout ported across HL majors is instrument-specific — it captures trend asymmetry on BTC-HL but on ETH-HL produces a net-loss default (PF 0.65, symmetric payoff) whose best-of-225 optimized config fails DSR (0.0), fails the holdout (negative), shows PBO 0.76, and trades in only one of three walk-forward windows. The clean-sizing finding (HL sizes correctly, metrics_reliable=true) is preserved; the trend edge is not present on this asset.

A long-only (LONG/FLAT), single-instrument daily Donchian-channel breakout trend strategy on ETHUSD.HYPERLIQUID. Goes long when the daily close breaks above the prior 20-day high (Donchian upper band) and exits to flat when the close falls below the prior 10-day low (Donchian lower band) or a 15% hard stop-loss is hit. Uses pure OHLCV data, never shorts, sizes one fixed-fractional position (95% of equity) once at entry with no dynamic resizing, and runs at leverage 1.0. The signal is a continuous channel-position score (~0 mid-channel, >1 on an upside breakout) thresholded in should_enter.

Sizing is CLEAN — the pre-registered acceptance check passed: avg_position_pct 66.6% (<100%), single-day returns within +29%/-19% (plausible for ETH at 1x), metrics_reliable=true. Confirms Hyperliquid as the trustworthy-sizing venue and shows NO position-stacking artifact.

Implementation is sound: long/flat (no flip-stacking), Donchian channel correctly excludes the current bar (no look-ahead), leverage 1.0, one fixed-fractional entry, close-to-flat before re-entry.

Belongs to a validated daily-trend family with promoted siblings; default 20/10 is canonical-but-untuned, so a parameter sweep is meaningful (not fitting to a bug).

30 trades over ~3 effective years is a workable sample for a daily breakout sweep.

Default config is a net loser: total_return -37.8%, profit_factor 0.65, Sharpe -0.71 (CI -2.9 to 1.09, straddles 0), win_rate 27%, negative in 3 of 4 active years.

Trend-capture asymmetry is ABSENT: avg_win (11,098) ~= avg_loss (11,445). A low-win-rate breakout needs winners >> losses to profit; the 15% hard stop + 10-day exit may be truncating winners.

Effective history is only ~3 years (no Hyperliquid ETH data pre-2023; 2021/2022 annual returns are 0.0), so walk-forward windows + 15-day holdout will be thin.

max_drawdown 59% with an 787-day drawdown duration at default params — high pain even if a better region exists.

Clean sizing confirmed (avg_position_pct 66.6% initial / 29.9% optimized, metrics_reliable=true, single-day returns within plausible ETH range) — the hypothesis's sizing-probe purpose succeeded and reinforces Hyperliquid as the trustworthy-sizing venue.

Implementation is sound (no look-ahead, long/flat, one fixed-fractional entry, leverage 1.0).

Failed deflated Sharpe: DSR 0.0 (<<0.95), is_significant=false; optimized Sharpe 1.70 is far below the expected-max luck bar 8.858 over 225 trials — indistinguishable from best-of-N selection noise.

PBO 0.76 (>0.5): parameter selection is more likely than not overfit; sharpe_ci_low -2.53 straddles 0.

Holdout FAILED: holdout_sharpe -0.85, ratio -1.27 — the once-tested recent window is negative; no forward generalization.

Walk-forward: 2 of 3 windows are zero-trade (IS [0,0,5.69], OOS [0,0,2.01]); the entire edge sits in one 2024 window. is_overfitted=false is an averaging artifact over empty windows.

Sensitivity FAILED: 2 cliffs (entry_lookback, entry_threshold); the entry_threshold x entry_lookback heatmap is negative Sharpe across nearly the whole grid — no robust profitable region exists.

Optimized result rests on only 14 trades / 81 effective days; the initial run was a net loser (PF 0.65) with avg_win ~= avg_loss (no trend-capture asymmetry).

Confirms instrument-specificity: Donchian breakout works on BTC-HL (PF 1.76) but not ETH-HL — the edge is BTC-specific and already captured by the sibling.

ETHHyperliquidDailyDonchianBreakoutLongFlat ported a validated BTC-Hyperliquid trend strategy onto ETH while doubling as a sizing probe to confirm Hyperliquid's backtest engine sizes correctly. The sizing probe succeeded — avg_position_pct 66.6%, metrics_reliable=true, plausible single-day returns — but the trading edge did not: the default config lost 37.8% with profit factor 0.65 and symmetric win/loss sizes, so the reviewer sent it to optimization mainly on the strength of its family and clean sizing. Optimization produced a superficially attractive Sharpe-1.70 config that proved to be best-of-225 selection noise, failing the deflated Sharpe (0.0), the holdout (-0.85), and sensitivity (2 cliffs), with the entire apparent edge concentrated in a single 2024 walk-forward window. The analyst abandoned it as a genuine no-edge result on this instrument, logging that the Donchian breakout is asset-selective — alive on BTC-HL, absent on ETH-HL — while preserving the useful finding that Hyperliquid remains the trustworthy-sizing venue.

A daily Donchian breakout is instrument-specific, not portable across majors — it captures trend asymmetry on BTC-HL (PF 1.76) but on ETH-HL the payoff is symmetric (avg_win ≈ avg_loss) with no robust profitable region to tune toward.

After optimization the analyst abandoned it (iteration 1 of 2) because every generalization gate failed together: deflated Sharpe 0.0 vs an expected-max-luck bar of 8.86, PBO 0.76, sharpe_ci_low -2.53, a failed negative holdout (-0.85, ratio -1.27), two of three zero-trade walk-forward windows, and a sensitivity grid with 2 cliffs that was negative across nearly the whole entry-threshold/lookback space.

A long/flat, single-instrument, OHLCV-only daily Donchian-channel breakout on ETHUSD.HYPERLIQUID (leverage 1.0, one fixed-fractional entry, no shorts), extending a validated daily-trend family to ETH on Hyperliquid — the venue identified as producing clean position sizing.

The default backtest was a net loser: -37.8% total return, Sharpe -0.71, profit factor 0.65, 27% win rate over 30 trades, 59% max drawdown, with avg_win (~$11.1k) roughly equal to avg_loss (~$11.4k); the best-of-225 optimized config looked better (Sharpe 1.70, PF 2.58, +28.7%, 13.5% DD) but rested on just 14 trades over 81 effective days.