BtcHyperliquid15mBreakoutContinuationLongFlat

Hyperliquid BTC Intraday 15-Minute Breakout Continuation, LONG/FLAT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Daily-Trend-Filtered, Fee-Gated Selective Entries Holding Hours, Fixed Single-Entry Sizing, No Flips — The One Remaining Evaluable Path: Clean-Proven HL BTC at Intraday Cadence to Defeat the HL Daily Data-Sparsity Wall, Fills Under-Weight HL Venue + Under-Weight Short Horizon, Pre-Registered Sizing Probe, Tail-Safe, Low-Parameter)

A LONG-ONLY (LONG/FLAT, never short), SINGLE-INSTRUMENT, INTRADAY 15-MINUTE breakout-continuation trend strategy on BTCUSD.HYPERLIQUID. This is the single configuration that survives every binding constraint discovered this session: (1) BINANCE USD-M perp AND COIN-M are both confirmed over-exposure-broken (analyst directive: stop probing BINANCE perp venues); spot-CASH MTM, multi-instrument aggregation, options settlement are broken; census signals (liq/OI/taker/account-ratio) are data-walled -> signals restricted to OHLCV. (2) HL is the only clean-sizing venue BUT not uniformly: HL BTC (90%) and HL ETH (64.5%) sized correctly while HL SOL over-exposed to 386% and HL long-short just over-exposed — so the SAFE HL configuration is BTC, LONG/FLAT (the proven-clean instrument+direction). (3) HL DAILY cadence fails validation on a structural data-sparsity wall: ~3yr of HL history gives a ~30-trade daily strategy 0-1 OOS trades per walk-forward window (HL BTC daily Donchian was clean & positive PF 1.76 but holdout-failed for exactly this reason). The analyst's explicit prescription: if HL must be used, ONLY higher-frequency (intraday) strategies can accumulate enough OOS trades. A 15-minute strategy over ~3yr (~75k bars) generates hundreds of trades, making a 3-window walk-forward + 15-day holdout statistically valid — directly defeating the wall that killed the HL daily books. It also fills the TWO most under-represented STILL-REACHABLE buckets: HYPERLIQUID venue (6.6% vs >=20%) and SHORT horizon (1m-15m, 3.5% vs >=10%). FEE DISCIPLINE is central (HL ~0.09% RT, need >0.15% net/trade): entries are SELECTIVE breakouts gated by range/volume expansion + a higher-timeframe trend filter, holding HOURS (not minutes), targeting >0.3-0.4% gross moves — explicitly NOT tick/sub-bar scalping (which fees would eat). Construction stays maximally bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, NO flips (long/flat), leverage 1.0, dual flat guard. PRE-REGISTERED ACCEPTANCE / SIZING PROBE (retained because HL is no longer assumed uniformly clean): avg_position_pct ~<=100% and no physically impossible intraday equity jump; if violated, the over-exposure defect reaches HL BTC intraday -> abandon & escalate. Low parameter count (breakout lookback, trend-filter EMA, ATR stop, one expansion filter).

Iteration 2 fixes the Layer-3 'no trades' failure (should_enter never fired over 4818 bars) with the smallest possible loosening, leaving the LONG/FLAT, fixed-fractional, leverage-1.0 structure and all signal scaffolding from previous_code intact. The zero-trade cause was an over-strict conjunction of three filters on a much shorter-than-assumed ~5k-bar window: (1) breakout_lookback cut 24->12 so new-high breakouts are far more frequent; (2) trend_ema_period cut 192->96 (1-day) so the trend EMA tracks price closely and a fresh breakout actually clears it instead of sitting below a slow long EMA, with the comparison relaxed from '>' to '>=' ; (3) range/vol expansion multipliers cut 1.5->1.05 so the fee-discipline selectivity gate remains but is no longer the binding constraint. No imports, class structure, or sizing logic changed, so Layers 1-2 stay green; these changes only increase entry frequency to clear the >=1-trade gate while preserving the breakout-continuation edge.

Data-wall + no edge. The hypothesis's defining bet was that 15-minute cadence over ~3yr (~75k bars) would generate hundreds of trades and defeat the HL daily data-sparsity wall. The actual HL BTCUSD 15m data covers only ~44 days / 5020 bars (2026-03-16..2026-06-26) and 74 trades — the wall is intact, just relocated to the intraday series, and a valid 3-window walk-forward + 15-day holdout is impossible on 6 weeks of data. The pre-registered sizing probe PASSES (avg_position_pct 98.78%, no impossible jumps), and trades match the mechanism, but within the only available window the strategy is decisively negative (Sharpe -7.69 with fully-negative CI, PF 0.49, win_rate 0.257, expectancy -$239/trade, -17.7% total). Optimizing 74 losing trades over 44 days would fit noise, and the developer cannot create HL history that does not exist. Same HL data-sparsity wall that killed the daily Donchian books, now confirmed at 15m — abandon and escalate per the hypothesis's own pre-registered acceptance criteria.

LONG/FLAT 15-minute breakout-continuation on BTCUSD.HYPERLIQUID, pure OHLCV, leverage 1.0. Enters when the bar close breaks above the prior 12-bar Donchian high while at/above a 1-day (96-bar) trend EMA and the bar shows range OR volume expansion (>=1.05x its prior-24-bar baseline). Exits on a trailing 12-bar Donchian low break, a frozen 2.5-ATR hard stop, or a 4% stop-loss. One fixed-fractional position per entry (95% of equity notional), never short, never flips.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: BTCUSD.HYPERLIQUID-15-MINUTE-LAST-EXTERNAL, Bars processed: 5011 Diagnostics: should_enter() returned a side 0 times over 4818 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Clean construction: pure OHLCV, long/flat, fixed-fractional sizing, leverage 1.0

Sizing probe PASSES — avg_position_pct 98.78% (<=100%), end_unrealized_pct 0.07%, no impossible equity jump, so the over-exposure defect did NOT reach HL BTC intraday

Trades match the hypothesized mechanism (74/74 long, breakout-direction, ~6h holds) — no mechanism mismatch

Central premise FALSE: claimed ~3yr / ~75k bars / hundreds of trades, but only bars_processed=5020, data_days=44 (2026-03-16 to 2026-06-26) — the HL data-sparsity wall was NOT defeated, it relocated to the 15m series

Only 74 trades over 44 days — a 3-window walk-forward + 15-day holdout would not be statistically valid (the exact failure the hypothesis tried to avoid)

Decisively losing in the only available window: Sharpe -7.69 (CI fully negative), profit_factor 0.49, win_rate 0.257, expectancy -$239/trade, total_return -17.7%, max 10 consecutive losses

Breakout-continuation has no edge here: 74% of entries lose; winners don't cover the 3x-more-frequent losers

This strategy was the session's last evaluable path: a clean, low-parameter long/flat 15-minute breakout-continuation book on the proven-clean HL BTC instrument, designed specifically to accumulate enough intraday trades to pass walk-forward and holdout validation. The construction held up — sizing was correct (avg position 98.78%, no impossible jumps) and the 74 long trades matched the breakout mechanism — but the central bet collapsed: Hyperliquid BTC 15m data spanned only ~44 days and 5,020 bars, not the assumed ~3 years and ~75k bars, so the data-sparsity wall simply relocated to the intraday series. In the single available window the strategy was decisively unprofitable (Sharpe -7.69, profit factor 0.49, win rate 25.7%, -17.7% total return), and optimizing 74 losing trades over six weeks would only fit noise. The analyst abandoned it at the backtest-review gate per the hypothesis's own pre-registered acceptance criteria, confirming the same HL data wall that killed the earlier daily Donchian books now persists at 15m, after two iterations.

Moving to a higher (15m) frequency does not defeat the Hyperliquid data-sparsity wall when the underlying instrument's available history is only ~6 weeks; verify actual data coverage before betting a hypothesis on assumed bar/trade counts.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate: the strategy's defining premise (~3yr / ~75k bars / hundreds of trades) was false — Hyperliquid BTC 15m data covers only ~44 days / 5,020 bars, so a valid 3-window walk-forward plus 15-day holdout was impossible, and within the available window the edge was clearly negative.

A long/flat, single-instrument 15-minute breakout-continuation strategy on BTCUSD.HYPERLIQUID using pure OHLCV, gated by a higher-timeframe trend EMA and range/volume-expansion filters, betting that intraday cadence would generate enough trades to defeat the Hyperliquid daily data-sparsity wall.

Over the only available window of 44 days / 5,020 bars it took 74 long trades and lost decisively: Sharpe -7.69 (CI fully negative), profit factor 0.49, win rate 25.7%, expectancy -$239/trade, and -17.7% total return. The pre-registered sizing probe passed (avg_position_pct 98.78%, no impossible equity jumps) and trades matched the intended breakout mechanism, but 74% of entries lost.