Strategies

Hyperliquid BTC 1-Hour Momentum Trend-Continuation, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Flip-Via-Flat, Sound Wide-Target/Tight-Stop Geometry, Fixed-Fractional Sizing — The Highest-Holdout-Density Non-Sub-Hourly Cadence [24 bars/day], Hard Pre-Registered DATA-DEPTH Kill-Switch to One-Shot Test Whether Deep 1H HL History Exists, Fills Long-Short + HL Venue, Tail-Safe, Low-Parameter)

Outcome: Abandoned

BtcHyperliquid1hMomentumTrendContinuationLongShort

Outcome Summary

BtcHyperliquid1hMomentumTrendContinuationLongShort was the carefully reasoned culmination of the HL cadence search: 1-hour is the highest-holdout-density timeframe outside the banned sub-hourly band, so it was run as a one-shot, kill-switch-gated test of whether deep 1H HL history exists. The construction proved sound — both the sizing gate (99.95%, balanced legs) and the corrected geometry gate (avg_win 1.78x avg_loss) passed — but the binding unknown resolved unfavorably: 1H HL BTC data goes back only ~7 months, far short of the ~2 years a walk-forward needs, even though the 4H sibling confirmed 4H history reaches 2024-01. The reviewer abandoned and escalated to the data engineer rather than iterating, framing it as a fixable backfill gap rather than a fundamental limit, and noting the strategy was also a net loser over the short window but that the verdict stands on data depth alone. The preserved finding is that deep-history HL work must use 4H or daily until 1H is backfilled, at which point this exact clean-sizing, sound-geometry 1H momentum-LS should be re-queued.

Strategy report

Hypothesis

A LONG-SHORT, SINGLE-INSTRUMENT, 1-HOUR momentum / trend-continuation strategy on BTCUSD.HYPERLIQUID. It is derived directly from the fully-mapped validation geometry of every prior HL failure: the engine needs BOTH deep history (3-window walk-forward) AND enough trades in the FIXED 15-day holdout, and no single tested cadence has had both — HL daily is deep (~3yr) but a 15-day holdout is only ~15 bars (the daily breakout got 0-1 OOS trades; even high-turnover daily tops out ~7), while HL 15-min has plenty of bars/day but only ~2 months total (single regime; 5 independent runs dead on it; sub-hourly generation is banned until backfill). The ONLY cadence that can satisfy both is intraday-but-NOT-sub-hourly (1H/4H) IF its history is deep. The 4H momentum-LS probe is already pending; this is the 1-HOUR complement, which is the STRONGEST holdout-density candidate: at 24 bars/day a 15-day holdout spans ~360 bars (vs ~90 at 4H, ~15 daily), so the holdout and each WF OOS window are richly populated, and if 1H history is deep (~3yr ~26k bars) the walk-forward is well-powered. 1H is explicitly OUTSIDE the sub-hourly (<1H) ban. The BINDING UNKNOWN is whether HL 1H data is deep (~2023+) or shallow (~2 months like 15-min); rather than thrash, this is structured as a ONE-SHOT test with a hard pre-registered kill-switch. Mechanism is the geometrically-SOUND momentum-LS already demonstrated clean on HL BTC (the 15-min momentum-LS showed avg_win 710 > avg_loss 327, 2.17x, and avg_position_pct 84.95% — only data length, never sizing or geometry, killed it), on the PROVEN-clean HL BTC long-short path (TS-mom LS sized 97%). It fills the dominant direction gap (long_only 85.9% vs <=55%) and the under-weight HL venue (7.2% vs >=20%). Bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, leverage 1.0, dual flat guard, flip-via-flat. PRE-REGISTERED ACCEPTANCE (hard kill-switches, evaluated FIRST, to avoid thrash): (1) DATA-DEPTH — verify 1H HL BTC data spans >= ~2 years AND the 15-day holdout contains >= ~10 trades; if NOT, the sub-daily HL data wall extends to 1H -> ABANDON & ESCALATE to the data engineer (do NOT iterate, do NOT optimize). (2) SIZING — avg_position_pct ~<=100% and no physically impossible 1H/daily equity jump; if violated, abandon & escalate as a reproducer. (3) GEOMETRY — avg_win > avg_loss must hold (wide trail > tight stop net of ~0.09% HL RT). Low parameter count (momentum lookback, EMA confirm, ATR stop/trail, time stop).

Implements the hypothesis as a one-shot test of the highest-holdout-density non-sub-hourly cadence on the proven-clean HL BTC long-short path. The 1H cadence is chosen by validation geometry: at 24 bars/day a fixed 15-day holdout spans ~360 bars (vs ~90 at 4H, ~15 daily), so the holdout and each walk-forward OOS window are richly populated, and if 1H HL history is deep (~3yr ~26k bars) the 3-window walk-forward is well-powered; 1H is explicitly outside the banned sub-hourly (<1H) generation window. Mechanism is the geometrically-sound momentum-LS already shown clean on HL BTC (the 15-min sibling had avg_win > avg_loss 2.17x and avg_position_pct ~85% — only data length killed it, never sizing or geometry), so it is reused here with the binding unknown (1H data depth) handled by the pre-registered hard kill-switch rather than iteration. Payoff geometry is corrected and explicit: a tight 1.5-ATR initial stop caps realized losses while the 2.5-ATR trailing stop lets winners run, so avg_win > avg_loss net of the ~0.09% HL round-trip; the trailing stop updates a price level only and does NOT resize the position, preserving the bug-avoiding fixed-fractional construction. LONG/SHORT fills the dominant direction gap on the under-weight HL venue, with flips routed through flat structurally (the base engine evaluates an entry only when flat) so the engine never nets opposing fills. Construction stays defect-avoiding: single-instrument pure OHLCV (no funding, no extra legs, no options), ONE fixed-fractional size (notional = 0.95 * equity, qty = notional/price), leverage 1.0 so the engine margin model and in-strategy notional agree. Pre-registered kill-switches evaluated first: (1) DATA-DEPTH — if 1H HL BTC data is not >= ~2 years or the 15-day holdout has < ~10 trades, the sub-daily HL data wall extends to 1H -> abandon & escalate to the data engineer (do not iterate/optimize); (2) SIZING — if avg_position_pct exceeds ~100% or a physically impossible equity jump appears, abandon & escalate as a reproducer; (3) GEOMETRY — avg_win > avg_loss must hold.

Abandon & ESCALATE to the data engineer — the hypothesis's hard, pre-registered DATA-DEPTH kill-switch failed and it explicitly specified abandon-and-escalate (do NOT iterate, do NOT optimize) in this exact case. The 1H HL BTC data spans only 2025-10-11 to 2026-05-20 (~7 months, 5243 bars, data_days 208), far short of the required ~2 years, so the sub-daily HL data wall extends to 1H and a 3-window walk-forward cannot be powered over a single-regime 7-month sample. IMPORTANTLY this is a data-collection gap, not a fundamental limit: the sibling 4H HL BTC review confirmed 4H data goes back to 2024-01 (~2.3yr, 5057 bars), while 1H has only ~7 months — the Hyperliquid catalog is simply not backfilled at 1H even though the API provides deeper 1H history. ACTIONABLE ESCALATION: backfill BTCUSD.HYPERLIQUID (and ETHUSD.HYPERLIQUID) 1H bars to ~2023 to match the 4H depth; once backfilled, this exact 1H momentum-LS becomes validatable (its sizing and geometry gates already pass). The other two pre-registered gates PASSED: SIZING (avg_position_pct 99.95%, metrics_reliable=true, no impossible jumps, balanced 80 long / 81 short) and GEOMETRY (avg_win $1876 > avg_loss $1054, 1.78x) — so the construction is clean; only data length is the blocker, mirroring the 15-min sibling. Note the strategy is also a net loser over the 7-month window (total_return -8.36%, Sharpe -1.34, PF 0.85, avg_trade -$107, win_rate 32% below the 36% geometry breakeven), but per the pre-registration the verdict is data-depth abandon regardless of performance. PRESERVED FINDING for the research lead: of the non-sub-hourly HL cadences, ONLY 4H currently has deep (~2.3yr) data; 1H is shallow (~7 months) until the data engineer backfills it — so until then, deep-history HL work must use 4H or daily, not 1H, and 1H momentum-LS should be re-queued only after the backfill lands.

Implementation

Long/short 1-hour momentum / trend-continuation strategy on BTCUSD.HYPERLIQUID, pure OHLCV. The signal is the 48-bar return vol-normalized into a Sharpe-like t-stat score; entries require momentum AND a fast/slow EMA (24/72) confirmation to agree: long when score > +0.5 and EMA-fast > EMA-slow, short when score < -0.5 and EMA-fast < EMA-slow. Exits use wide-target/tight-stop geometry: a TIGHT 1.5-ATR initial stop that ratchets into a 2.5-ATR TRAILING stop (lets winners run, caps losers), plus an EMA-cross-against trend-over exit, a 6% percentage backstop, and a 72-bar (3-day) time stop. Flips are routed through flat (the base engine only enters when flat). One fixed-fractional position (95% of equity notional) per entry with no dynamic resizing — the trailing stop is a price risk level, not a size change — leverage 1.0.

Backtest Review

Sharpe: -1.34
Total return: -17.65%
Max drawdown: 29.39%
Trades: 161
Win rate: 32.3%
Profit factor: 0.85

Pre-registered SIZING gate PASSED: avg_position_pct 99.95%, metrics_reliable=true, no impossible 1H/daily equity jumps, balanced long (80) / short (81) legs — the HL BTC 1H long-short flip-via-flat path sizes cleanly.

Pre-registered GEOMETRY gate PASSED: avg_win $1876 > avg_loss $1054 (1.78x), tail_ratio 1.33 — the wide-trail/tight-stop construction works as designed.

Holdout density would be fine if the data were deep (24 bars/day) — the design correctly identified 1H as the highest-density non-sub-hourly cadence; only the data length defeats it.

DATA-DEPTH KILL-SWITCH FAILED: 1H HL BTC data spans only 2025-10-11 to 2026-05-20 (~7 months, 5243 bars) vs the required ~2 years. A single-regime 7-month window cannot power a 3-window walk-forward, so the strategy is unvalidatable at this cadence.

Net loser over the available window: total_return -8.36%, Sharpe -1.34, Sortino -2.23, profit_factor 0.85, avg_trade_return -$107 (negative) — win_rate 32% is below the 36% geometry breakeven.

Fee drag meaningful (commission_pct_of_gross 14.0%), 14 max consecutive losses.

Outcome Summary

Among non-sub-hourly HL cadences only 4H currently has deep (~2.3yr) data while 1H is shallow (~7 months) until backfilled — so deep-history HL work must use 4H or daily for now; the 1H momentum-LS construction itself is clean (sizing and geometry both pass) and becomes validatable once the data engineer backfills HL 1H bars to ~2023.

The analyst abandoned and escalated at the backtest-review gate because the hard data-depth kill-switch failed: 1H HL BTC data spans only ~7 months (5,243 bars, 2025-10-11 to 2026-05-20) versus the required ~2 years, so a 3-window walk-forward cannot be powered on a single-regime sample; per the pre-registration this is a data-depth abandon regardless of performance.

A long/short, single-instrument, OHLCV-only 1-hour momentum trend-continuation strategy on BTCUSD.HYPERLIQUID (EMA-confirmed entries, tight ATR initial stop with a wider ATR trailing stop, flip-via-flat, leverage 1.0), chosen as the highest-holdout-density non-sub-hourly cadence (24 bars/day) and structured as a one-shot test, gated first by a hard pre-registered data-depth kill-switch, of whether deep 1H HL history exists.

Two of three pre-registered gates passed — sizing (avg_position_pct 99.95%, metrics_reliable=true, balanced 80 long / 81 short) and geometry (avg_win $1,876 > avg_loss $1,054, 1.78x) — but over the available ~7-month window it was a net loser: -8.36% total return, Sharpe -1.34, profit factor 0.85, 32.3% win rate over 161 trades (below the 36% geometry breakeven), avg_trade_return -$107, with fees at 14% of gross.

Backtest and paper results are hypothetical. Trading involves risk of loss.