BtcHyperliquidDailyTSMomentumLongShort

Hyperliquid BTC Daily Time-Series Momentum, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Flip-Via-Flat, Fixed-Fractional Sizing, Validated TS-Momentum Mechanism on the One Confirmed Clean-Sizing Venue + the One HL Asset With Demonstrated Trend Edge, Fills the Dominant Long-Only Over-Representation, Pre-Registered LS-Sizing Probe, Tail-Safe, Low-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, DAILY time-series (absolute) momentum strategy on BTCUSD.HYPERLIQUID. Three convergent reasons make this the best remaining move: (1) DIRECTION is the dominant quota miss — long_only is 85.8% vs the <=55% target (30.8pp over), far larger than any other dimension's gap; the factory needs long_short, and the ONLY venue where long_short sizes correctly is HYPERLIQUID (BINANCE USD-M perp shorts/flips are engine-corrupted — proven over-exposure defect; spot-CASH MTM, multi-instrument aggregation, and options settlement are all also defective; census signals liq/OI/taker/account-ratio are data-walled, so signals are restricted to OHLCV + funding). (2) MECHANISM is separately VALIDATED — the time-series-momentum long-short family reached paper_stage on the BINANCE siblings (Sol/Eth/Btc TS-momentum), but has NEVER been run on the clean HL venue; this is the correct-sizing expression of an already-working edge, not a new untested idea. (3) ASSET-SELECTIVITY (the key recent lesson: HL daily breakout worked on BTC PF 1.76 but FAILED on ETH PF 0.70) says validate the trend edge per instrument and prefer the asset with demonstrated HL trend — BTC. Rather than clone the long/flat Donchian breakout a fourth time (HL BTC positive, HL ETH dead, HL SOL pending, CM BTC/ETH pending), this varies the MECHANISM (continuous TS-momentum sign vs discrete channel breakout) AND adds the short side on the proven-trend asset. Construction is deliberately simple and bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry (NO vol-target re-scaling, NO leverage stacking), leverage 1.0, and flips routed THROUGH FLAT (on a sign change, exit fully to flat on one bar, then open the opposite side) rather than direct reversal — so the engine never has to net opposing fills, the cleanest long-short construction. PRE-REGISTERED ACCEPTANCE / SIZING PROBE (doubles as the HL long-short sizing test, since all clean-HL evidence so far is long/flat): avg_position_pct ~<=100% and no physically impossible single-day equity return (BTC at 1x is ~+/-15-20%). If violated, the over-exposure defect extends to HL on the short/flip path -> abandon and escalate as the HL-LS reproducer. If clean, this is the first correctly-sized long-short trend book in the portfolio and establishes the HL LS path. Low parameter count (one lookback, one trend filter, one ATR stop).

Implements the hypothesis as the first correctly-sized long-short trend book in the portfolio, on the only venue that sizes long/short cleanly (HYPERLIQUID) and the one HL asset with a demonstrated trend edge (BTC, HL daily breakout PF 1.76; ETH was dead at 0.70). It addresses the dominant long_only over-representation (85.8% vs <=55% target) and the under-weight HL venue simultaneously. The mechanism is the TS-momentum long-short family already promoted to paper_stage on the BINANCE siblings, here given its correct-sizing expression on the clean venue rather than a new untested idea; it varies the mechanism from the existing HL Donchian breakout (continuous vol-normalized momentum sign vs discrete channel) and adds the short side on the proven-trend asset. Construction is deliberately defect-avoiding: (1) single-instrument, pure OHLCV (no funding subscription, no extra legs) so multi-instrument aggregation and funding defects cannot apply; (2) flip-via-flat is structural — should_exit closes to flat on a reversal and the base engine evaluates an entry only when position is None, so the opposite side opens on a later bar and opposing fills are never netted; (3) ONE fixed-fractional size (notional = 0.95 * equity, qty = notional/price) with no vol-target rescaling or leverage stacking, leverage left at 1.0 so the engine margin model and in-strategy notional agree. The vol-normalized momentum score keeps the signal continuous and conviction-bearing and makes the single entry_threshold behave consistently across volatility regimes. Daily bars match the long-term momentum timeframe and keep HL fee drag (~0.09% RT) negligible given few trades/year. Pre-registered sizing probe: avg_position_pct should be <=~100% and single-day equity returns bounded near BTC's daily move at 1x (~+/-15-20%); a violation would indicate the over-exposure defect reaches the HL short/flip path and should be abandoned/escalated as the HL-LS reproducer.

Abandon — optimizing this re-runs a known-failing experiment and will hit the documented daily-sparsity holdout wall. (1) HOLDOUT SPARSITY: 66 trades total (~40 real post-2023) with 22d 18h average holds means the FIXED 15-day holdout contains ~0-1 trades and auto-rejects — the identical wall that killed every HL daily-trend book (the hypothesis itself names it), and a 22-day-hold momentum book is even sparser than the breakouts that already failed it. (2) NO STABLE EDGE: per-trade expectancy is NEGATIVE (-$1742), win_rate 22.7%, tail_ratio 0.023, skew 7.8 / kurtosis 79.9 — the +124.5% headline is carried by 2-3 outlier 2024-bull trades (2024-01 +51.8%, 2024-12 +53.2%; annual 2024 +83% vs 2023 +4.3%, 2026 -4.3%), i.e. bull beta concentrated in a few multi-week winners, exactly what the post-opt DSR/holdout gates deflate. (3) FAMILY ALREADY DEAD POST-OPT: the same TS-momentum-LS mechanism on the SOL-HL sibling was optimized and abandoned this session (DSR 0.0, walk-forward OOS [0,0,-1.5], holdout -0.68, PBO 0.62), and BTC's own daily trend edge in its other expression (HL daily Donchian) also failed post-opt (DSR 0.0, all-zero OOS, negative holdout) — so both the mechanism (TS-mom LS) and the asset's daily-trend premise (BTC) have already collapsed out-of-sample; this combination would just re-collapse. Not iterate: there is no code-level bug — the construction is clean and the sizing probe passes; the failure is the structural daily-sparsity holdout wall plus a fat-tail bull-beta 'edge', neither fixable by parameter changes. PRESERVED FINDINGS: (a) the HL long-short short/flip path sizes CLEANLY (avg_position_pct 58.6%, balanced legs, metrics_reliable=true) — the LS sizing question is now answered affirmatively on both BTC and SOL, so future LS work need not re-probe it; (b) low-frequency daily-trend/momentum on HL is structurally unvalidatable via the 15-day holdout regardless of asset or whether the mechanism is Donchian or TS-momentum — the research lead should stop queuing daily-cadence HL trend books (they die on holdout sparsity) and pursue edges with much higher trade frequency, OR await the 1H/4H backfill for a deep intraday trend cadence; (c) DATA FLAG to the data engineer: BTCUSD.HYPERLIQUID daily data starts 2020-08-19 with ~26 zero-PnL pre-2023 trades before HL launched — trim/verify these backfilled bars as they pollute pre-2023 walk-forward windows.

Long/short daily time-series (absolute) momentum strategy on BTCUSD.HYPERLIQUID, pure OHLCV. The signal is the 30-day return vol-normalized into a Sharpe-like t-stat score: goes long when score > +0.5 and close > 50-day SMA, short when score < -0.5 and close < 50-day SMA. Exits on momentum reversal (score crossing the exit band or close crossing the SMA against the position), a 3x-ATR hard stop frozen at entry, or a 15% percentage stop. Flips are routed through flat (exit fully, then re-enter the opposite side on a later flat bar) so the engine never nets opposing fills. One fixed-fractional position (95% of equity notional) per entry, no vol-target rescaling, leverage 1.0.

Pre-registered SIZING PROBE PASSED: avg_position_pct 58.6%, metrics_reliable=true, balanced 33 long / 33 short, no physically impossible single-day equity jump (the +51%/+53% figures are per-trade returns over ~22-day holds in the 2024 bull, not daily equity jumps) — the HL long-short short/flip path sizes cleanly, confirming the HL LS path alongside the SOL-HL TS-mom probe.

Net-positive headline (total_return +124.5%, Sharpe 0.94, PF 1.52, Sortino 2.26) and deep daily data for a walk-forward.

DAILY-SPARSITY HOLDOUT WALL: 66 trades total (~40 post-2023) with 22d 18h average holds → a 15-day holdout contains ~0-1 trades → auto-reject. This is the exact wall that killed every HL daily-trend book; a 22-day-hold momentum book is even sparser.

Negative per-trade expectancy (-$1742) with win_rate 22.7%, tail_ratio 0.023, skew 7.8 / kurtosis 79.9 — the +124% is carried by 2-3 outlier 2024-bull trades (2024 +83% vs 2023 +4.3%, 2026 -4.3%), i.e. bull beta, not a stable edge.

Family already failed post-opt this session: the identical TS-momentum-LS mechanism on SOL-HL was abandoned (DSR 0.0, OOS [0,0,-1.5], holdout -0.68), and BTC-HL daily Donchian also failed (DSR 0.0, negative holdout).

Data-quality flag: BTCUSD.HYPERLIQUID daily data starts 2020-08-19 with ~26 zero-PnL trades pre-2023 (before HL existed) — flat/backfilled bars that pollute pre-2023 walk-forward windows.

BtcHyperliquidDailyTSMomentumLongShort paired a promoted TS-momentum mechanism with HL's clean-sizing venue and its one proven-trend asset, BTC, aiming to establish the first correctly-sized long-short trend book and clear the dominant long-only quota gap. The sizing probe succeeded (avg_position_pct 58.6%, balanced 33 long / 33 short) and the +124.5% headline was eye-catching, but it masked a negative -$1,742 per-trade expectancy and a fat-tail bull-beta profile concentrated in a few 2024 winners. The reviewer abandoned it before optimization, recognizing the 22-day-hold book would hit the same daily-sparsity holdout wall that killed every HL daily-trend strategy, and that both the mechanism (SOL-HL TS-momentum) and the asset's daily-trend premise (BTC-HL Donchian) had already collapsed out-of-sample this session — so optimizing would only re-collapse. It preserved three findings: the HL long-short path sizes cleanly on both BTC and SOL, daily-cadence HL trend books are structurally unvalidatable on the holdout, and ~26 pre-2023 backfilled BTC bars should be trimmed by the data engineer.

Low-frequency daily-cadence HL trend/momentum is structurally unvalidatable through the 15-day holdout regardless of asset (BTC/SOL) or mechanism (Donchian/TS-momentum) — long holds make the holdout too sparse — so pursue higher-frequency edges or await the 1H/4H backfill; the HL long-short flip-via-flat path is now confirmed to size cleanly on both BTC and SOL.

The analyst abandoned it at the backtest-review gate, before optimization, because it would re-run a known-failing experiment: 66 trades with ~22-day holds means the fixed 15-day holdout contains ~0-1 trades (the daily-sparsity wall that killed every HL daily-trend book), the edge is fat-tail bull beta rather than stable, and the same TS-momentum-LS mechanism (on SOL-HL) and BTC's daily-trend premise (in Donchian form) had both already collapsed post-optimization this session.

A long/short, single-instrument, OHLCV-only daily time-series momentum strategy on BTCUSD.HYPERLIQUID (vol-normalized momentum score with SMA trend filter, flip-via-flat, fixed-fractional entries, leverage 1.0), applying a BINANCE-sibling-promoted TS-momentum mechanism to the one clean-sizing venue and the one HL asset with a demonstrated trend edge (BTC), while doubling as the HL long-short sizing probe.

The headline looked strong — +124.5% total return, Sharpe 0.94, profit factor 1.52, Sortino 2.26 over 66 trades (33 long, 33 short) — but per-trade expectancy was negative (-$1,742) with a 22.7% win rate and extreme fat-tail profile (tail_ratio 0.023, skew 7.8, kurtosis 79.9), the return carried by 2-3 outlier 2024-bull trades; the sizing probe passed cleanly (avg_position_pct 58.6%, balanced legs, metrics_reliable=true).