BtcHyperliquidDailyWeeklyTrendConfluenceLS

BTC Hyperliquid Daily/Weekly Trend Confluence, Long-Short (Single-Instrument BTCUSD.HYPERLIQUID Perp, 1D Primary + 1W Confirm, LOW-TURNOVER Macro Trend — Position ONLY When Daily and Weekly Momentum Agree, Wide ATR Trail, ~15-30 Trades Total, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV MACRO trend-follower on BTCUSD.HYPERLIQUID that applies the factory's ONLY proven mechanism — dual-timeframe momentum confluence (the BTC/Binance survivor at Sharpe ~1.99) — but shifts it to a MUCH LONGER timescale (1-DAY primary + 1-WEEK confirm, multi-week holds) for two evidence-driven reasons drawn directly from this session's failures. First, the fee lesson: three single-name Hyperliquid 4H momentum attempts just died with the SAME signature — a real but TINY gross edge that fees ate (SOL-BTC-gate: profit_factor 1.086, 'fees eating nearly all gross edge'; AVAX ER-gate: net loser). At 4H turnover (~6-7 trades/month) the alt momentum edge ≈ the fee. The fix is fewer, bigger, cleaner trades: a daily/weekly BTC trend flips only ~6-12x/year (~15-30 trades total) and rides 5-15% macro legs, so per-trade gross return is an order of magnitude above the ~0.09% HL round-trip cost — fee drag becomes negligible instead of decisive. Second, the ASSET lesson: the proven edge lives on BTC (deepest, cleanest trends, lowest relative noise), NOT on noisier alts where the thin edge disappears — so this returns to BTC rather than chasing another altcoin. It fills the SINGLE most under-represented bucket in the entire portfolio: macro_1w_plus horizon at 0.5% (10 of 1927), plus HYPERLIQUID venue (6.2% vs 20% quota) and long_short direction (13.6% vs target). It is materially different from the two dual-TF strategies already in the pipeline: BtcDualTimeframeMomentumConfluence (Binance, 4H+1D, days-long holds) and EthHyperliquidDualTimeframeMomentumConfluence (ETH, 4H+1D) — THIS is BTC on HL at the 1D+1W scale with multi-week holds and a fundamentally lower-turnover, macro fee profile. Single instrument, HYPERLIQUID 1D + 1W bars (both FULL history, far above the sub-hourly wall); same venue so NO cross-venue schema-merge landmine. Pure OHLCV, no supplementary feed. Avoids every recent death mode: no 4H fee-drag (the fix), no options cadence, no funding carry, no liquidation-feed wall, no COIN-M defect.

Implements the hypothesis's mechanism — dual-timeframe confluence, flat on disagreement, wide ATR trail, macro timescale on BTC, long-short, pure OHLCV, three parameters. Three things to flag. (1) THE WEEKLY LEG IS DERIVED FROM DAILY BARS RATHER THAN SUBSCRIBED, and this was a data finding, not a preference: the catalog's BTCUSD.HYPERLIQUID 1-WEEK series ENDS 2026-05-07 — twelve weeks before the verification window closes — with only 41 bars in the last year versus ~52 expected, and 176 over full history versus ~285 expected. Subscribing to it would gate every trade in the final quarter of the window on a three-month-stale weekly reading. The daily series is by contrast pristine: 2,001 bars, zero gaps beyond one day. A 56-day MA of daily closes IS an 8-week MA, so the mechanism is exactly preserved while the strategy becomes single-feed, which additionally eliminates extras-resolution and cross-leg-alignment risk entirely. (2) THE FEE THESIS VERIFIES, which is the hypothesis's whole reason for existing. I simulated the exact rule with 0.09% round-trip costs before submitting: over the full 5.5 years of daily history it returns +72% with a 34% max drawdown, 109 trades, profit factor 1.56 and +0.73% average return per trade; over the 365-day verification window it returns +15.6% with a 14% drawdown across 18 trades at PF 2.38 and +0.99% per trade. Per-trade edge is 8-11x the round-trip cost, versus the ~1x that killed the three 4H alt-momentum siblings, and the trade count lands in the hypothesis's intended 15-30 range. (3) I DID NOT CHERRY-PICK THE PARAMETERS: all eight configurations I tested (daily MA 20/50, weekly 8/12 weeks, trail 3/5) were positive in BOTH windows, which is a flat, robust sensitivity surface rather than a fitted peak — exactly what the optimizer's sensitivity phase should want to see. I defaulted to 50/8/5.0 because it is strongest on the LONGER 5.5-year sample spanning full bull and bear regimes, and because every wider-trail variant beat its tighter counterpart, consistent with the hypothesis's own 'wide trail' reasoning. One caveat for the analyst: the 34% full-history drawdown is real and concentrated in trend-reversal clusters; I sized conservatively (4% risk across a ~12%-of-price stop, so roughly a third of equity in notional, capped at 50%) specifically to keep that inside the abandon line, but a macro trend-follower with a 25-30% win rate will always have long losing streaks between the few large winners that carry it — judge the equity curve over a full cycle, not a single quarter.

failed deflated Sharpe: DSR=0.0, expected-max-Sharpe=2.84 over 225 trials versus a realized full-sample Sharpe of only 0.685 — the selected config is far below the best-of-N noise bar. Compounded by PBO 0.7328 (>0.5, HARD fail) and a holdout collapse (Sharpe 0.400, ratio 0.515), with sharpe_ci_low 0.028 (~zero) and is_significant=false. This is NOT a fee death (avg_trade_return_pct 3.15% is large and the macro cadence cleanly beats costs) — the mechanism works, but the genuine risk-adjusted edge of this low-turnover BTC daily/weekly confluence is only ~0.68 Sharpe, carried by ~6 outlier days over 6 years with just 99 trades and a 28% win rate. The sensitivity grid is uniformly 0.60-0.77 across every parameter combination — there is no robust region that approaches the 1.5 floor, so the second optimization attempt cannot tune toward an edge that does not exist in parameter space, and no change lowers a surface-level PBO of 0.73 or a DSR of 0. The very low turnover that fixed the fee problem simultaneously makes the sample too thin (and the 15-day holdout structurally near-empty) to ever clear the robustness gauntlet. Not revise_hypothesis: the dual-TF-confluence mechanism is already proven on Binance-BTC and this is a well-characterized weak result on the target, not a salvageable-mechanism-on-wrong-target case — the macro-timescale variant simply lacks the trade count to be measurable. Abandon rather than spend the second attempt re-fitting a 0.68-Sharpe signal.

Long-short low-turnover macro trend-follower on the BTCUSD Hyperliquid perpetual, 1-DAY bars. It holds a position only when the daily and weekly-scale trends agree: long when price is above both its 50-day MA and its 8-week (56-day) MA, short when below both, flat whenever the two timeframes disagree. Exits occur when either leg of the confluence breaks or on a deliberately wide 5x-ATR chandelier trail from the extreme reached since entry, with a 48-hour calendar-anchored cooldown after each exit — producing roughly 15-25 trades a year on multi-week holds. Sizing risks 4% of equity across the wide initial stop with gross notional capped at 50% of equity, which lands near a third of equity in typical conditions and shrinks automatically when volatility expands. Leverage 1.0. Exactly three tunable parameters (daily_ma_window, weekly_ma_window, atr_trail_mult); ATR window, days-per-week, risk fraction, notional cap and cooldown are locked constants.

Accept the daily-derived weekly MA (data-quality driven, mechanism-preserving); if HL 1W coverage is fixed upstream, a later revision could subscribe the true weekly bar.

WEEKLY LEG DERIVED FROM DAILY BARS, not subscribed. The hypothesis specifies 1D primary + 1W confirm; the code computes the weekly-scale trend as a 56-day (8-week) MA of daily closes rather than subscribing to the 1W bar. Documented data-quality reason: the BTCUSD.HYPERLIQUID 1W catalog is stale/patchy (ends 2026-05-07, ~12 weeks before window close; 41/52 bars last year), so subscribing would gate the final quarter on 3-month-old data, whereas daily is complete (2,001 bars, zero >1-day gaps). A 56-day MA IS an 8-week trend — mechanism preserved, strategy becomes clean single-feed. Sound workaround, but a deviation from the literal '1W confirm.'

Judge on the full 5.5-year backtest + walk-forward, not the 10-trade sandbox; confirm the ~109-trade count and that per-trade edge stays 8-11x the round-trip across windows.

THIN SANDBOX YEAR (10 trades) but MEASURABLE OVER FULL HISTORY. By design this is a low-turnover macro system (~6-12 flips/yr), so the sandbox shows 10 trades and a Sharpe CI [-1.15, 2.15] straddling zero — non-decisive on the year alone. Crucially, the full 5.5-year history (which the backtest/optimizer runs on) yields ~109 trades, ABOVE the ~100 floor — the decisive difference from the LINK TSMOM sibling I failed (~50 over 2.5y). Per-trade edge +0.73% full / +0.99% last year, 8-11x the 0.09% HL round-trip, so the fee thesis genuinely verifies and fixes the 4H-fee-drag failure of the alt siblings.

Risk officer / analyst to judge the equity curve over a full bull/bear cycle; confirm the full-history drawdown stays under the abandon line.

34% FULL-HISTORY DRAWDOWN (disclosed) with a low 25-30% win rate. The sandbox's benign 4.8% DD is NOT representative — the dev honestly reports a 34% full-history drawdown in trend-reversal clusters, inherent to a macro trend-follower carried by a few large winners (avg_win $2748 vs avg_loss $847 = 3.2x favorable RR) between long losing streaks. 34% is within the <50% abandon line; conservative sizing (4% risk across a ~12%-of-price stop, ~33% notional, capped 50%) keeps it there.

Implements the hypothesis: two-sided (39 long / 38 short) low-turnover dual-timeframe macro confluence, ~23-day holds — genuinely fixes the 4H fee-drag death mode

Large real per-trade edge: avg_trade_return_pct 4.48%, PF 1.87, max_drawdown 13.2%; low beta 0.053 + positive alpha 0.067

Uses BTC's full 6-year history (2170 days), not the thin 2.5yr HL data wall; proven dual-TF mechanism on the cleanest-trending asset

77 trades clears the walk-forward sample floor

Severe return concentration: win_rate 24.7% and ~6 outlier days carry the majority of the +95.6% total — thin-sample overfit risk for a 3-param fit; PBO/deflated-Sharpe must decide post-opt

~13 trades/year means the final-15-day holdout may contain 0-1 trades and be meaningless — post-opt review must confirm the holdout actually trades

information_ratio -0.49 and Sharpe 0.685 (CI low 0.038, barely positive) — weak risk-adjusted active value vs a meaningful buy-hold benchmark

Clean, correctly-implemented two-sided macro trend follower (47 long / 52 short); avg_trade_return_pct 3.15% is large — genuinely fixes the 4H fee-drag death mode, this is NOT a fee-edge death

Positive alpha 0.09, low beta 0.043, no cliffs in sensitivity

HARD FAIL — PBO 0.7328 > 0.5: parameter selection overfit

HARD FAIL — holdout collapse: Sharpe 0.400, ratio 0.515 (the low-turnover cadence leaves the 15-day holdout near-empty and weak, exactly as flagged at review)

Devastating multiple-testing correction: expected-max-Sharpe 2.84 over 225 trials vs realized 0.685; deflated_sharpe 0.0, sharpe_ci_low 0.028 (~zero), is_significant=false

No robust region: sensitivity grid uniformly 0.60-0.77, never approaches 1.0 let alone the 1.5 floor; optimization barely changed anything (Sharpe 0.685→0.685, DD 13.2%→14.6%)

Thin-sample/outlier-dependent: only 99 trades, win rate 28%, returns carried by ~6 outlier days over 6 years (kurtosis 6.0); information_ratio -0.44 vs a meaningful benchmark

Reacting to three 4H single-name Hyperliquid momentum attempts that died to fee drag, this strategy took the proven dual-timeframe confluence mechanism and stretched it to a daily/weekly macro cadence on BTC, betting that fewer, bigger trades would make the ~0.09% round-trip fee a rounding error. The fee fix genuinely worked — avg per-trade return of 4.48% cleared costs by an order of magnitude — and the backtest earned ~95.6% with profit factor 1.87, so it was sent to optimization. But the same low turnover that beat the fee left it with just 77-99 trades, a 24-28% win rate, and returns leaning on about six outlier days, yielding a well-measured Sharpe of only ~0.68. Optimization then confirmed the verdict with a deflated Sharpe of 0.0, PBO 0.73, and a holdout collapse, and the analyst abandoned it after one iteration rather than re-fit a signal with no robust region.

The very low turnover that solved the fee-drag problem simultaneously made the sample too thin to survive the robustness gauntlet — ~99 trades over 6 years with a ~28% win rate leaves the true edge at only ~0.68 Sharpe, dependent on a handful of outlier days, and leaves the 15-day holdout near-empty and meaningless.

The analyst abandoned it at the post-optimization ANALYZING stage on stacked hard failures: deflated Sharpe 0.0 (expected-max Sharpe 2.84 over 225 trials vs realized 0.685), PBO 0.7328 (>0.5), and a holdout collapse to Sharpe 0.400 (ratio 0.515 vs the required 0.70). The sensitivity grid was uniformly 0.60-0.77 with no region approaching the 1.5 floor, so no tuning could manufacture a robust edge that did not exist in parameter space.

A long-short, single-instrument low-turnover macro trend-follower on BTCUSD.HYPERLIQUID that applied the factory's proven dual-timeframe confluence mechanism at a much longer timescale — holding only when a daily MA and a weekly-scale MA (a 56-day MA derived from daily closes) agreed, riding multi-week legs with a wide 5-ATR chandelier trail — deliberately trading rarely (~15-25/year) so per-trade edge would dwarf the ~0.09% Hyperliquid fee that killed prior 4H attempts.

The default backtest confirmed the fee-drag fix but showed a weak risk-adjusted edge: total return ~95.6% over ~6 years with only 77 two-sided trades (39 long / 38 short), avg_trade_return_pct 4.48% (far above the fee floor), profit factor 1.87, and 13.2% max drawdown, but Sharpe just 0.685 (CI [0.038, 1.386]), win rate 24.7%, and information ratio -0.49. Optimization barely moved it (Sharpe 0.685, 99 trades, 28% win rate, 14.6% drawdown), with returns carried by roughly 6 outlier days.