Frozen-Rule Breadth Transfer of the Absorbed-Flush Rule to Hyperliquid: buy a Hyperliquid alt perp at the close of a down day whose range blew out (>=2x its 60-day return std) but that closed in the upper 45% of its range, fixed 3-day clock, 10% per leg, max 6 concurrent, on 15 HL perps disjoint from the Bybit panel (1-DAY, long-only)

Hypothesis

Venue: HYPERLIQUID is 14% of the portfolio against a >=20% target. All 15 names are disjoint from the Bybit panel, the factory core and the L250 list. Mechanism: the factory's most productive family, long-side shock absorption, run as a frozen-rule breadth test of rule (a) as L258 asked. Short mirrors are excluded per L230/L256. Range_mult and close_loc are allowed only narrow ranges around the Bybit values. Hold, sizing and lookback stay fixed at the promoted panel's values. Evidence counts are pre-registered (96 distinct event dates).

Implementation

Long-only 15-leg Hyperliquid 1-DAY alt-perp panel (ARB primary + OP, SUI, LDO, CRV, SNX, RUNE, DYDX, GMX, STX, WLD, BLUR, kPEPE, SEI, TIA). On each aligned daily close, a leg signals when its high-low range over the prior close is >= range_mult (2.0) x the sample std of its previous 60 daily log returns (current bar excluded), it closed below the prior close, and its close location (C-L)/(H-L) > close_loc (0.55). Eligible legs are listed now (universe_at), have >= 70 bars of their own history, are flat, and fewer than 6 positions are open; excess signals are ranked by range/sd. Market BUY at the signal bar's close (execution substitution: on this 24/7 venue that close is the next bar's open), notional = 10% of account equity at 1x. Fixed time exit: market SELL at the close of the 3rd daily bar after entry (from position.ts_opened). No stop or target.

Frozen-rule breadth transfer of the Bybit absorbed-flush rule (a) to Hyperliquid, implemented exactly as the hypothesis states. Extra legs contribute only with the bar closing at the primary's decision instant (no look-ahead). Risk control route (b): fixed 3-day time exit plus capped unlevered sizing (10% per leg, max 6 concurrent = 60% gross, leverage 1.0). point_in_time_universe=true so the window is anchored on the primary and later-listed legs join when listed and seasoned (>=70 bars). Catalog symbol for PEPE is kPEPEUSD.HYPERLIQUID (case as stored). Note: the primary ARB's HL history starts 2023-03, so the run window starts there and excludes the pre-2023 catalog bars of CRV/SNX/RUNE/DYDX/OP/LDO that the pre-study used (~34 of 200 events); the hypothesis names ARB as primary, so this is kept. min_notional (10 USD, HL minimum) is a venue constraint, not a tunable. No _param_bounds since no parameter is clamped.

Verification Results

None required. The analyst should compare event counts with the post-2023 subset of the pre-study.

The run window starts with the primary ARB in 2023-03 (point_in_time_universe anchors on the primary). Every leg then needs 70 buffered bars of its own before _features returns anything, so the first ~70 days of the window cannot trade, and the pre-2023 pre-study events are out of scope. The developer discloses this in config.rationale. It is consistent with the hypothesis, which names ARB as primary.

None required.

Positions that are due to exit on the current bar do not count toward max_concurrent, so new entries can take their slots on the same close. For one bar, gross can briefly exceed 60% (at most 6 new legs plus the closing legs, each 10% at 1x), which stays within 1x margin. This matches 'fewer than 6 positions open' after the same-close exits, so it is a faithful reading.

Pre-study 2026 events are flat (-0.02% on 29 events) and the 2025 mean is driven by a few large rebounds. Funding is not deducted in the pre-study. The run window is anchored on ARB (HL history from 2023-03), so roughly 34 of the 200 pre-registered events (the pre-2023 CRV/SNX/RUNE/DYDX/OP/LDO bars) cannot occur. Compare the measured event count with about 166 rather than 200. Five of the 15 names had negative means in the pre-study.

Backtest Review

Sharpe

0.73

Total return

32.89%

Max drawdown

9.09%

Trades

151

Win rate

53.0%

Profit factor

1.70

The trades implement the hypothesis: 128 closed positions, all long, every one held exactly 3 days, 10% of equity per leg at 1x (largest leverage 0.1). Entries fire on 58 distinct dates across all 15 legs, and the run starts at 2023-06-02, after the 70-bar point-in-time warmup.

Economics after costs, funding included, are well above the HYPERLIQUID floor: 2.14% mean per trade against the 0.15% floor and 0.09% round trip, profit factor 1.70, total return +28.8%, Sharpe 0.76, drawdown 9.1% (the contract limit is 22%). Funding cost -$2.6k against +$28.8k net P&L, so it is small.

The implied deployment per trade is 28.8 / (128 x 2.14%), about 10.5% of equity. That matches leg_fraction, so the drawdown and return are at the intended scale and are not shrunk by under-sizing.

On QA's concern: the run window is anchored on ARB's Hyperliquid history, so the right comparison is roughly 166 pre-study events less the concurrency cap and the holdout. The 128 trades measured in the development period are consistent with that. The edge is not refuted at this stage.

128 trades is enough to run the trade-count-sized walk-forward (about 6 windows of 20 trades) with only 2 tunable parameters in narrow ranges.

The profit is concentrated: the 5 largest trades ($17.6k) are about 61% of net P&L. Return skew is 1.57 and kurtosis 11.8, so the edge relies on a few large rebounds, as the pre-study already warned.

Results decay by year: 2023 +$3.0k (30 trades), 2024 +$17.5k (69), 2025 +$12.4k (23), 2026 development slice -$4.2k (6). The recent weakness that QA and the Research Lead flagged shows up here too, on a thin sample.

Per-name results are mixed. BLUR (-$4.3k), LDO, RUNE, DYDX and SEI are negative, broadly the same names the pre-study flagged. SUI, TIA and CRV carry much of the gain.

The Sharpe confidence interval spans zero (-0.40 to 1.81) and the probabilistic Sharpe is 0.72, so the development-period evidence is not yet statistically strong.

Analysis

Out-of-sample evidence pooled across the four walk-forward windows is statistically sufficient under the registered test: annual Sharpe 1.38, HAC t 2.46 against a hurdle of 2.0, deflated Sharpe 0.993, N_oos 1. All 4 windows are positive (0.38, 2.42, 0.84, 1.48).

No hard or soft gate failures. Replay of the final parameters (range_mult 2.275, close_loc 0.506) over all history: 162 trades, PF 2.50, 3.69% net per trade against a 0.18% floor (about 20x the HL round trip), win rate 0.59 with payoff 1.76, max drawdown 7.7% against a pre-registered 22%, Calmar 2.33. All six pre-registered predictions met.

The baseline before optimization was already positive with the frozen Bybit-style values: 151 trades, PF 1.70, 2.02% per trade, Sharpe 0.73, drawdown 9.1%. That matches the promoted Bybit flush-absorption sibling (PF 1.72, 1.92% per trade). The transfer replicates the rule's economics on a disjoint universe.

Tuning stayed stable. Every window chose close_loc of about 0.50-0.55 and range_mult of about 2.19-2.52. The final set sits inside those choices (overall stability: moderate, no unstable parameter). The close_loc cliff sits at 0.69-0.83, well outside the tuned range of 0.50-0.62. Sharpe rises steadily with range_mult with no cliff, so stricter flush definitions do better, which fits the mechanism.

Risk control route (b) holds. Worst realised loss is 1.80% of equity (BLUR, 2024-08-02). Largest position is 10.1% of equity. Peak margin use is 0.60x, with no leverage. Funding is credited and costs -$2.6k against +$32.9k on the baseline. Impact is 4-6.6% of gross. Implied deployment is about 10% of equity per trade, so the drawdown and return figures are comparable with the floors at the intended allocation.

PnL is heavily concentrated by date. On the baseline ledger (151 trades, 73 distinct entry dates), the 3 best entry dates contribute +$29.7k of the +$32.9k total. Without them the panel earns +$3.2k. The top 5 trades contribute $17.7k. The edge is effectively 'buy absorbed flushes after market-wide liquidation cascades' (e.g. Aug 2024 yen carry, Oct 2025). These are rare, clustered events, so the independent-event count is far below 151. The regime split agrees: all profit came in the stressed-vol third (+36.5%), while the calm and normal thirds are slightly negative.

The recent edge is flat to negative. 2026 lost $75 on 29 trades in the baseline. The holdout (2026-02-09 to 2026-10-30, 263 days, 19 trades) has Sharpe -0.21 and t -0.18, so its significance is inconclusive. It is statistically consistent with the walk-forward OOS (z -0.91) and its drawdown was only 3.2%. The pre-study also showed 2026 flat, and the Bybit sibling's 2026 was negative. Decay cannot be ruled out.

The validation verdict is 'not enough trades to validate': only 2 of 4 windows reached 20 trades (windows 2 and 3 had 10 and 12).

Per-name results are mixed: LDO, RUNE, BLUR and WLD are net negative. The universe was fixed before measurement, so the names are kept, but the edge is not uniform.

The full-history replay of the final parameters (PF 2.50, Sharpe 1.28) is partly in-sample. The untuned baseline (Sharpe 0.73) is the more honest level to expect. In-sample Sharpe below 1.5 is advisory only.

inconclusive, consistent_with_oos true (z -0.91)

positive & significant

advisory only

unevaluated in backtest metrics; pooled OOS DSR is 0.993

Risk Review

The book already holds several long-side shock and absorption panels: BybitFlushAbsorptionPanel (this rule's parent), HyperliquidDailyDownShockAbsorptionPanel, HyperliquidAltPanelCapitulationAbsorption4H, HlIdioDownShockReboundBtcHedged, HyperliquidNr7DownsideEventPanel, BybitSpotAltPanelCapitulationAbsorption4H and others. The universe is disjoint by name, but the trigger is the same market-wide liquidation cascade. The analyst's ledger shows most of the PnL comes from a handful of cascade dates (Aug 2024, Oct 2025), and every profit came in the stressed-vol third. So these sleeves will enter together and win or lose together. The tail correlation with the existing book is high, even if correlation on ordinary days is low.

Worst adverse excursion was 8.1% of equity against a worst realised single-trade loss of 1.8% (BLUR, entered 2024-08-02, -18% on a 10% leg). Gross exposure reaches about 60% of equity across 6 alt legs, each with 5-7% daily vol. On a cascade day like 10 Oct 2025, the open-position drawdown could approach or trip the 5%-per-hour strategy circuit breaker. That would close the legs near the lows, which is exactly where this edge gets paid. The backtest does not model that breaker, so paper PnL on cascade days may come in below the backtest.

The recent edge is weak. 2026 is -4.3% in the development period (Jan-Feb 2026: -3.1% and -1.2%). The pre-study's 2026 was flat (-0.02% on 29 events), and the holdout was inconclusive with a slightly negative Sharpe, though statistically consistent with the walk-forward results. Validation outcome: not enough trades to validate (2 of 4 windows reached 20 trades). This is low-confidence paper, not real-money evidence.

Risk control route (b) holds: a fixed 3-day time exit plus capped unlevered sizing. Leverage is 1.0, each leg is 10% of equity (largest measured position 10.1%), there are at most 6 concurrent legs, and peak margin use is 0.60. The conservative profile holds: realized vol 15.0% (band 5-60%), peak position vol 17.3% (cap 120%), worst trade 1.8% of equity (cap 15%). Max drawdown 7.7% is well inside the pre-registered 22%. No soft floors are in play, so no waivers are needed. The one advisory, in-sample Sharpe below its floor, is non-blocking during the soak and needs no waiver. deflated_sharpe is unevaluated in the backtest metrics, so it is unknown, not passed. The pooled walk-forward deflated Sharpe is 0.993.

Live book: no members in breach and no open execution items. Many paper members show a decision replay of 'diverged' or 'no_data', and several carry paper-parity warnings. That is not a blocker for a paper admission, but it means the book's paper evidence is still thin.

Admit to paper with the low-confidence label ('low confidence: not statistically proven'). This is not real-money evidence.

Count this sleeve in the same correlation and risk bucket as the existing absorption and capitulation panels, BybitFlushAbsorptionPanel above all. Cap the bucket's combined gross exposure so that one market-wide liquidation cascade cannot stack every long-shock sleeve at full size.

Keep leg_fraction at 0.10, max_concurrent at 6 and leverage at 1.0. Any increase needs a new risk review.

Outcome Summary

The edge rests on a few market-wide liquidation-cascade dates (the 3 best entry dates gave +$29.7k of the +$32.9k baseline P&L, all of it in the stressed-vol third), and 2026 was flat to negative, so paper results should be judged on independent events and checked for decay, not on raw trade count.

Backtest review chose optimize, and the analyst then recommended promote: pooled OOS cleared the significance test, there were no hard or soft gate failures, all six pre-registered predictions were met and risk control route (b) held (worst loss 1.8% of equity). The run reached paper_stage after 1 iteration.

A breadth transfer of the promoted Bybit absorbed-flush rule to 15 disjoint Hyperliquid alt perps: buy at the close of a down day whose range was at least 2x the 60-day return std but which closed in the upper part of its range, hold 3 days, 10% of equity per leg, at most 6 positions open, long-only and unlevered.

The untuned baseline made +32.9% over 151 trades (PF 1.70, 2.02% net per trade, Sharpe 0.73, max drawdown 9.1%). Pooled walk-forward OOS had Sharpe 1.38 (HAC t 2.46, deflated Sharpe 0.993), but the holdout (19 trades, 2026-02-09 to 2026-10-30) had Sharpe -0.21, which is inconclusive and consistent with OOS (z -0.91).