Strategies

Frozen-Rule Hyperliquid Transfer of 4H Volume-Confirmed Capitulation Absorption: buy a Hyperliquid alt perp at the close of a <=-3-sigma, >=1.5-sigma-volume 4H bar, fixed 48h clock, 10% per leg, max 5 concurrent, flat ~90% per name (15 established HL perps, 4-HOUR, long-only event panel, pure OHLCV)

Outcome: Paper Trading

HyperliquidAltPanelCapitulationAbsorption4H

Outcome Summary

This strategy carried the promoted Bybit 4H volume-confirmed capitulation-absorption rule over unchanged to 15 Hyperliquid alt perps as a fresh test. The backtest reproduced the pre-study: +56.1% over 253 trades, average trade +1.84%, Sharpe 1.47, max drawdown 8.7%, every year positive, and the result did not depend on the two crash days. After one iteration and optimization, the analyst recommended promotion on its economics and OOS/holdout consistency, while noting that statistical significance was not established and that the per-trade edge is decaying in 2026. It has moved to the paper stage.

Strategy report

Hypothesis

Original Research Lead hypothesis. This is a frozen-rule venue transfer (lesson L228: a transfer is a fresh test, reported on its own) of the promoted 'Bybit Spot 4H Volume-Confirmed Capitulation Absorption Panel' (run dc5dc587: 405 trades, average trade +1.78%, Sharpe 1.26, max DD 15%). The source rules come over unchanged: 540-bar (90-day) z-scores, shock <= -3 sigma, log-volume >= +1.5 sigma, 6-bar per-coin cooldown, 12-bar (48h) time exit, 10% of equity per leg, at most 5 open, most negative shock filled first. The destination is HYPERLIQUID perps. That venue is 13.3% of the corpus against its >=20% target, and this mechanism has never run there at 4H. The mechanism follows L224/L230: the long side of a shock-absorption event, flat most of the time, fixed-clock exit. A forced or panic seller (on Hyperliquid, often an on-chain liquidation of a levered long) leaves an unusually large 4H drop on unusually high volume. Patient liquidity takes the other side and is paid by a partial rebound over the next 1-2 days. Why Hyperliquid, and why this should differ from the source: HL liquidations are executed on-chain by the liquidator/HLP vault against the book, transparently and at size. The flow is mechanical, with no information behind it, so the reversal premium should exist there too. But L228 shows that transfers often fail (OKX wide-range continuation, OKX upside-jump). The test is therefore registered as a fresh test with its own predictions, not as support carried over from Bybit. Universe, fixed by a data rule before looking at returns: every non-BTC/ETH alt from the promoted OKX/Bybit capitulation and NR7 universes whose HYPERLIQUID 4-HOUR history in the catalog starts on or before 2024-04-30. That gives ADA, AVAX, DOGE, DOT, LINK, LTC, NEAR, ATOM, FIL, SOL, XRP, BCH, ETC, UNI and AAVE. Names weak in the pre-study (SOL, UNI) stay in. Every leg checks is_listed() before entry. HL timeframe limit: 4H is the finest HL interval with more than 2 years of history (catalog HL 4H bars run 2024-01-22 to 2026-10-18, about 6,000 bars per name). After the 540-bar warm-up, signals start 2024-04-21, roughly 2.5 years of tradable history. PRE-STUDY (catalog HL 4H bars, z-scores from prior bars only, entry at the signal close, exactly the frozen rule including the 5-slot cap and cooldown, net of 0.09% round trip + 0.05% impact, funding NOT included). 256 trades, average net +1.99%, median +1.22%, 60% winners, PF 2.04. There were 114 distinct entry days (L238: the honest sample size; the per-day mean was about 0.8% in an uncapped variant). By year, average net per trade: 2024 +3.2% (81 trades), 2025 +2.3% (78), 2026 +0.70% (97). The edge is DECAYING in 2026, and the holdout (last ~20%, about 6 months of 2026) will sit in that weaker regime. The 5 best trades are 27% of P&L. Excluding the two best entry days, the average is +1.53%. The simple-sum equity at 10% per leg returned about +51% with about a 10% drawdown. Lesson L237 check: the overlap where all legs have bars covers the whole window, and the 114 event days are well above the ~30 minimum. Cost arithmetic: HL round trip is 0.09% taker plus about 0.05% impact. A long holding 48h pays about 0.06% funding at a typical +0.00125%/h (more in froth). Total cost is about 0.20% per trade. A conservative 50% haircut on the pre-study's +2.1% gross gives about 1.0% per trade, about 5x cost. Even the 2026-only figure (+0.70% net) is about 3.5x cost. Execution: entry is a market BUY at the signal 4H bar's close. On a 24/7 venue that is the next bar's open (registry next_bar_open_fill -> use_equivalent: market order at the signal bar's close). The exit is a market SELL after 12 closed 4H bars, on a time clock only, with no stop and no target. Risk control route (b): a time exit plus capped sizing at 10% notional per leg and at most 5 legs (50% gross, leverage setting 1x, no single position above 100% of equity). Per-leg volatility contribution is about 10% x ~90% alt vol = ~9% annualized, well inside the 120% cap. Expected strategy vol is about 15-25%. Pre-registered comparisons: (1) lesson L170, the strategy's max drawdown against equal-weight buy-and-hold of the same 15 HL perps over the same window; (2) results per year and with and without the two largest crash days (2024-08-05, 2025-10-10) reported separately; (3) funding paid per trade from the HL funding feed. Known limitation, stated up front: HL history has no 2022-style sustained bear market. The OKX daily down-shock transfer (closed underpowered 2026-10-18) lost 12% in 2022, so this test cannot measure the bear-regime risk.

Implementation

Event panel of 15 Hyperliquid alt perps on 4H bars. For each coin: shock_z is the 4H log return divided by the std of the previous 540 returns, and volume_z is the z-score of log volume over the previous 540 bars. An event is shock_z <= -3 and volume_z >= +1.5, with a 6-bar cooldown per coin. Entry is a market BUY at the signal bar's close; on a 24/7 venue that price is the next bar's open, which is the registry's use_equivalent for next_bar_open_fill. Each leg is 10% of mark-to-market equity, at most 5 legs are open, and the most negative shock is filled first. Exit is a market SELL after exactly 12 closed 4H bars (48h). No stop, no target, long-only, leverage 1x.

This is a line-for-line frozen-rule transfer of the promoted Bybit spot panel code (dc5dc587) to HYPERLIQUID perps. The statistics, event rule, cooldown, ranking, slot cap, sizing and time exit are unchanged. There are three changes. (1) Venue and universe: the 15 HL perps named in the hypothesis, with ADA as primary because it has the earliest HL 4H history. (2) min_notional is 10, the HL minimum, fixed in the research plan. (3) point_in_time_universe=True, so ETC (HL 4H history from 2024-04-20) does not clamp the window; every leg still checks is_listed() before entry. Risk route (b): a fixed 48h time exit plus capped sizing (10% per leg, at most 5 legs = 50% gross, 1x). Each position contributes about 9% annualized vol, inside the conservative profile. The engine charges funding on the HL perp legs. Per the plan, only shock_z_min is exposed to optimization (2.7-4.0); all other parameters are fixed.

Verification Results

Make every leg warm up at the same time. Either return a small value from min_bars_required() (e.g. 2) and rely on the per-coin ready_ flags, while keeping _max_bars large, or feed the primary bar into _update() for the bars the base template skips (e.g. seed from self._bars on the first calculate_signal call). Cross-sectional ranking already ignores coins that are not ready.

The primary leg (ADA) warms up one full lookback later than the 14 extra legs. min_bars_required() returns vol_lookback+2 = 542. The base template skips calculate_signal while len(self._bars) < 542, and ADA's rolling windows are updated only inside calculate_signal, via _update. So ADA's return and log-volume windows start filling at primary bar 542 and are full only at about bar 1083. The extra legs are ready at about bar 541. Over the full backtest, ADA is untradable for about 90 extra days (about 2024-04 to 2024-07). If walk-forward windows restart with only the 545-bar warm-up in evaluation.warmup_bars, ADA can stay unready for roughly the first 540 bars (~90 days) of every outer window. That quietly drops 1 of the 15 frozen names from much of the OOS test. The rule itself is unchanged and ranking for the extras starts on time, so this is not a hypothesis mismatch.

No change needed if this matches the promoted Bybit source code, which the rationale says is ported line for line.

When an event fires but no slot is free, it still sets last_event_ts and starts the 6-bar cooldown. This matches the source rule ('an event within cooldown of the coin's previous event is ignored'). It does mean a coin whose first event is blocked by the slot cap cannot take a deeper follow-on shock within 24h. This is noted for transparency, not as a defect.

The pre-study shows the edge decaying (+3.2% per trade in 2024, +0.70% in 2026), and the holdout falls in 2026. The 5 best trades are 27% of P&L, and two crash days (2024-08-05, 2025-10-10) are likely to dominate. HL history has no sustained bear market. The analyst should check the per-year split, the result with the two crash days excluded, and funding paid per trade on 48h longs during froth.

Backtest Review

Sharpe: 1.47
Total return: 56.09%
Max drawdown: 8.71%
Trades: 253
Win rate: 58.5%
Profit factor: 1.88

The trades match the mechanism: all 180 are LONG, every one is held exactly 2 days (12 x 4H, a time exit only), and each leg is 10% of equity (leverage 0.1 per leg). This is the frozen rule as written.

The economics clear costs several times over. Average trade is +2.29% of notional against the 0.15% Hyperliquid floor, PF is 2.10, win rate 60%, and average win/loss is $857/$612. Funding is credited at -$492 in total (about 0.03% per trade), so it is not material.

Risk is modest. Max drawdown is 6.4% against the pre-registered 20% (Calmar 3.5). The worst trade lost $1,929, about 1.9% of equity. Sharpe is 1.66 and +48.5% total over the development period (2024-04 to 2026-03). The result is close to the pre-study (avg +1.99%, PF 2.04), so the backtest reproduces the registered rule.

QA's concerns, checked: removing the two named crash days (2024-08-05, 2025-10-10) leaves 173 trades averaging +2.23%. Removing the two best entry days by P&L leaves an average of +1.61%. Funding per 48h long is negligible here. The edge is not carried by the crash days.

The edge is decaying, as the pre-study predicted. Average trade by year: 2024 +2.81% (78 trades), 2025 +2.21% (78), 2026 Q1 +0.85% (24). That is still above the floor, but the forward regime is weaker.

The honest sample is 79 distinct entry days, not 180 independent trades. Clustered entries on crash days are correlated.

There is no sustained bear market in the HL history, so bear-regime risk cannot be measured here (a stated limitation).

Only one tunable (shock_z_min, 2.7-4.0) and everything else frozen, so the optimization mainly tests robustness rather than adding much to it.

Analysis

Frozen rule transferred unchanged; pooled walk-forward OOS Sharpe 1.03 (t 1.30), 5 of 6 windows positive; 200-day holdout Sharpe 1.02-1.10 on 72 trades in the weaker 2026 regime, ratio 0.805 to OOS, z +0.04 (consistent with OOS)

Average trade +1.84% vs the 0.18% floor; PF 1.88; win rate 58.5%; funding and impact included; ledger reconciles

Max drawdown 8.7% vs the pre-registered 20%; Calmar 2.25; worst realised loss 1.40% of equity; largest position 10.4% of equity

All 6 checkable pre-registered predictions met; every year positive; top year carries 47% of profit; removing the two best entry days leaves $41.3k of the $56.1k

Fixed parameters all sit on plateaus within one standard error

Significance not established: pooled OOS t 1.30 and holdout t 0.83, both against a hurdle of 2.0

shock_z_min has a one-sided edge: Sharpe drops from 1.66 to 0.91 at 2.7; keep it frozen at 3.0

Per-trade edge is decaying (pre-study +3.2% in 2024, +0.7% in 2026); longest drawdown 170 days

Signal IC 0.089 (t 1.9) is not distinguishable from noise

History has no sustained bear market, so bear-regime risk is unmeasured

Forward returns will be correlated with the promoted Hyperliquid daily down-shock panel

LOW CONFIDENCE: not statistically proven. Labelled paper observation only, at reduced allocation, sized jointly with the Hyperliquid daily down-shock panel as one correlated sleeve. Kill criteria: re-evaluate after 60 paper trades; kill if the average net trade return is below +0.40% or PF is below 1.15; kill at any time if paper drawdown exceeds 15%; no retuning of shock_z_min during paper.

Risk Review

Correlation with the deployed book: this is the same frozen rule as BybitSpotAltPanelCapitulationAbsorption4H (paper), on largely the same alt names. It also fires on the same crash days as HyperliquidDailyDownShockAbsorptionPanel, HyperliquidNr7DownsideEventPanel, Nr7FailedDownsideBreakReversalPanel and OkxNr7FailedDownsideBreakReversalPanel. On a market-wide alt liquidation such as 2024-08-05 or 2025-10-10, these long-only shock-absorption panels will all be full at once. Each panel's 50% gross cap does not bound the combined book's long-alt exposure on those days.

Losses on single trades run well past the 2-day clock. The worst trade (SOL, entered 2024-08-02) lost 13.97% of notional, which is 1.40% of equity, and the worst adverse excursion reached 6.68% of equity across open legs. With no stop, a crash that continues takes all 5 slots (50% gross) at once. The worst development-period days were -4.18% (2024-12-10) and -3.97% (2024-11-26). These are acceptable against the 20% pre-registered drawdown, but there is no per-trade loss bound beyond the 10% leg size.

The edge is shrinking. Development-period annual returns were +23.3% in 2024, +18.1% in 2025 and +2.0% in Jan-Mar 2026, and the hypothesis's own pre-study shows the average trade falling from +3.2% to +0.70%. The holdout result (2026-04-01 to 2026-10-18) is inconclusive, and statistical significance is not established. The strategy is admitted only as low-confidence paper.

QA-noted warm-up lag: the primary leg (ADA) builds its rolling windows only after the base template's 542-bar gate, so it becomes tradable about 90 days after the other 14 legs and after each walk-forward reset. Before live, fix it so all 15 names warm up together. Risk is not increased by this, because the lag only removes trades.

Market impact: each leg is 10% of equity (largest measured 10.4%, ETC). That is small relative to Hyperliquid books for SOL, XRP, DOGE and LINK. ETC, ATOM, FIL, BCH and DOT are thinner on Hyperliquid, and entries are market orders placed in exactly the moments when the book is being swept. The capacity estimate is withheld, so impact at a larger allocation has not been verified.

Hyperliquid history (2024-2026) has no sustained bear market. Drawdown in a 2022-style regime is unmeasured; the hypothesis states this limitation itself.

Paper only, under the 'low confidence: not statistically proven' label. Real-money release needs significant forward or holdout evidence, or the owner's informed consent.

For capital allocation, treat this strategy and the other long-only shock-absorption and NR7-downside panels (BybitSpotAltPanelCapitulationAbsorption4H, HyperliquidDailyDownShockAbsorptionPanel, HyperliquidNr7DownsideEventPanel, the NR7 failed-downside panels) as one correlated sleeve, so their combined long-alt gross on a crash day is bounded at the portfolio level.

Keep all parameters frozen as tested (position_fraction 0.1, max_concurrent 5, hold_bars 12, leverage 1x). Do not raise the leg size or slot count without a new review.

Outcome Summary

The transferred edge held up on Hyperliquid but is decaying: per-trade average fell from about +2.8-3.2% in 2024 to about +0.7-0.85% in 2026, and the signal IC (t 1.9) was not distinguishable from noise. Paper trading should watch whether the 2026 per-trade edge stays above costs, and keep shock_z_min frozen at 3.0, where Sharpe drops to 0.91 at 2.7.

Backtest review sent it to optimization. The analyst then recommended promotion: the economics were well above the 0.18% floor, drawdown was well inside the pre-registered 20%, all checkable pre-registered predictions were met and the holdout agreed with OOS. However, significance was not established (OOS t 1.30, holdout t 0.83, against a hurdle of 2.0). The run is now in paper trading.

A frozen-rule transfer of the promoted Bybit 4H capitulation-absorption panel to 15 Hyperliquid alt perps: buy at the close of a 4H bar with a return of -3 sigma or worse on volume of +1.5 sigma or more, hold exactly 48h, 10% of equity per leg, at most 5 legs, long-only, on the idea that liquidity providers are paid for absorbing forced selling.

The full backtest (2024-04 to 2026-10) returned +56.1% over 253 long trades, with Sharpe 1.47, average trade +1.84% of notional, PF 1.88, win rate 58.5% and max drawdown 8.7%, with funding and impact included. Pooled walk-forward OOS Sharpe was 1.03 (t 1.30, 5 of 6 windows positive) and the 200-day holdout Sharpe was 1.02-1.10 on 72 trades.

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