Hyperliquid Daily Volume-Confirmed Down-Shock Absorption Panel: buy a Hyperliquid perp at the close of a <=-2.5-sigma, >=1-sigma-volume UTC day, fixed 2-day clock, 22 point-in-time names, max 8 concurrent at 8% each (HYPERLIQUID 1-DAY, long-only event trade, flat most of the time)

Hypothesis

Implementation

Long-only event panel on 22 Hyperliquid perps (1-DAY bars, pure OHLCV). On each UTC daily close, for every listed name with at least `lookback`+1 prior non-zero-volume daily bars (zero-volume pre-listing filler ignored), compute shock_z = (today's log return - mean of the prior `lookback` log returns) / their std, and volume_z = (today's volume - mean of the prior `lookback` volumes) / their std. If shock_z <= -shock_z_min (2.5) and volume_z >= volume_z_min (1.0), buy at that close with 8% of mark-to-market equity, at most 8 concurrent legs (64% gross, 1x); when more names fire than free slots, the most negative shock_z names are filled first. Each position exits on a fixed clock after hold_days (2) daily bars; no stop, no re-entry on a name while it is held, and a name that exits on a bar cannot re-enter on that same bar (its freed slot is available to other names). Execution substitution (next_bar_open_fill = use_equivalent): orders are market orders at the signal bar's close, which on a 24/7 venue is the next bar's open.

Implements the Research Lead's volume-confirmed down-shock absorption panel exactly: volatility-scaled return shock plus volume z-score trigger from past bars only, point-in-time eligibility via is_listed() plus own non-zero-volume history, fixed 2-day time exit, fixed 8% legs capped at 8 concurrent (risk route (b): time exit plus capped sizing, gross 64% of equity at leverage 1; per-leg vol contribution about 8% x ~80% = ~6% annualised, far inside the 120% profile cap). The universe is all 22 named HL perps (the 20 declared instruments plus TIA and SEI from the description); ETHUSD is the long-lived primary anchor and also trades. Extra legs contribute only when their bar is stamped exactly at the primary decision bar. Interpretation choices the hypothesis left open: returns are log returns, the shock z subtracts the rolling mean, volume z uses raw volume, and crash-day overflow beyond the 8 slots is ranked by most extreme shock_z. Note the hypothesis itself flags 2026 YTD as flat to slightly negative, so the edge may be decaying.

Verification Results

Compare with a small tolerance (e.g. >= hold_days*DAY_NS - DAY_NS//2), or record the entry decision bar's ts per instrument and count bars from it.

The time exit `_due` measures time from position.ts_opened. In the backtest a bar-close market order fills at the decision bar's timestamp, so the exit lands exactly hold_days bars later. In paper/live, a fill stamped even slightly after the bar close makes decision_ns - ts_opened fall just short of hold_days*DAY_NS at the hold_days-th close, and the exit slips one extra day.

No change needed for backtest correctness; mentioned only as a live-robustness note.

Positions whose instrument still has a working order (`busy`) are counted as active and are not exited that bar. In the backtest, market orders fill within the bar, so this path is not reachable. In live, a stuck order delays the clock exit by a bar.

The hypothesis's own pre-study shows the edge shrinking: the capped 2-day portfolio is about -1.4% in 2026 YTD, and the BTC-hedged control cut the mean to +0.9-1.4%. Most of the premium is the market-wide crash rebound, so the drawdown sits in sustained bear legs, where up to 8 correlated longs (64% gross) all lose together. The analyst should check per-year returns and returns in the stressed tercile, especially the 2026 holdout. Per-event means of +2-3% are far above the ~0.09% Hyperliquid round trip, so fees are not the worry.

Backtest Review

Sharpe

0.88

Total return

44.32%

Max drawdown

8.03%

Trades

191

Win rate

62.8%

Profit factor

2.55

The trades match the hypothesis: 117 long-only trades, every one held exactly 2 days, 8% of equity per leg, spread across all 22 names (BTC 12, LTC 11, DOGE 9 ... APT 1). The 8-leg cap binds on crash days (8 entries on 2024-03-20, 2024-08-06, 2024-12-10 and 2025-03-04), which is the market-wide crash-rebound design the hypothesis describes.

Economics are well above costs: average trade return is 3.04% of notional against the 0.15% Hyperliquid floor (about 20x), profit factor is 3.67, win rate is 67.5% with a payoff ratio of about 1.76, and fees total $993 against about $32k of profit.

Risk is small: max drawdown is 6.2% against the 20% pre-registered limit, and the worst trade loses $1,188 (about 1.2% of equity). The largest position is 8% of equity, well inside the risk_control limit of 100% of equity per position and 2% loss per trade.

The development period matches the pre-study (117 trades, +32%, Sharpe 0.96 vs the pre-study's 195 trades, +30%, Sharpe 0.75 over a longer span). 2023 and 2024 each made +13.5%; 2025 to the cut made +2.4%.

Trading only starts in June 2023 because Hyperliquid has no real volume before then. The effective sample is therefore about 2 years and 117 trades, many of them correlated on the same crash days. Daily Sharpe confidence interval is 0.12 to 1.64.

Profit is concentrated in a few crash-rebound months: 2023-11 (+8.8%) and 2024-03 (+7.7%) account for about half of it. 2025 to the cut is weak (+2.4%), consistent with QA's concern and the hypothesis's own warning that the edge may be decaying. The holdout and the stressed periods need to be read carefully after optimization.

No data before 2023 means the stress windows for 2021-2022 (China ban, LUNA, FTX) show no trades and are not tested.

With four tunables (lookback, hold_days, shock_z_min, volume_z_min) on about 117 correlated events, the optimizer can easily fit noise. hold_days in particular has an economic prior (the rebound lives in days 1-3) and could reasonably be fixed at 2.

Analysis

Pooled walk-forward OOS is significant under the registered policy: annual Sharpe 1.15, HAC t 2.58 against a 2.0 hurdle, DSR 0.995, and N_oos = 1 with no earlier result looks. OOS trades are 121 across windows 2-4 (2023-05 to 2025-07); windows 0-1 had no trades because real HL trading history starts in 2023.

Holdout (2025-07-25 to 2026-10-18, 450 days, 56 trades, not underpowered) uses the final parameters and is positive: Sharpe 0.60, drawdown 8.3%. That is statistically consistent with OOS (z = -0.42), and the ratio of 0.52 is what a real edge of about Sharpe 1 looks like on a one-year sample.

Economics are strong. With the final parameters, the full-history replay averages 2.38% per trade against the 0.18% HL floor (about 13x). PF is 2.26, win rate 60% with a payoff ratio of 1.50, and impact is 3% of gross. Commission and funding are a small share of gross.

Not beta. Exposure is 31%, beta to the equal-weight basket is 0.004 and correlation 0.05. Max drawdown is 8.3%, while the HL alt basket fell far more than 50% over the same 2023-2026 window, so the drawdown is well below 0.7x buy-and-hold. The return is earned in short 2-day windows after crash days, not by holding the market.

The sensitivity sweep shows no cliffs and wide plateaus on every parameter (lookback 30-90 gives Sharpe 0.64-0.93, shock_z 1.25-3.75 gives 0.63-0.96, volume_z 0.5-1.5 gives 0.82-0.95). All of these sit inside one Sharpe s.e. (0.41) of the base.

Every year with trades is positive: 2023 +13.5%, 2024 +13.6%, 2025 +9.4%, 2026 YTD +2.3%. The top year is 32% of PnL. The pre-registered predictions were met 6 of 6 (trades 155 vs 150, PF 2.26 vs 1.3, average trade 2.38% vs 1.2%, drawdown 8.3% vs 20%).

Measured risk control: the worst realised loss is 1.30% of equity (SEI, 2026-06-03, -16.3% on an 8% leg), the largest position is 8.0% of equity, and peak margin utilisation is 0.64x. That fits the signal-exit route: a clock exit with capped sizing, well under 100% of equity per position.

Significance on the holdout alone is not established (t 0.72), so the overall evidence is low confidence. The holdout does not demonstrate the edge, but it does not contradict it either.

Parameter stability is unstable for both tunables (shock_z_min dispersion 0.91, volume_z_min 0.87). The final volume_z_min of 0.59 sits below every window's choice (0.76-1.30). The flat sensitivity plateau means this mostly reflects a flat objective rather than a cliff, but the deployed set is not a point that the walk-forward kept choosing.

Decay risk in 2026: YTD is only +2.3%, with June 2026 at -4.8%. The holdout Sharpe of 0.60 is about half the OOS figure. The Research Lead flagged this risk in advance.

The edge is a market-wide crash rebound. It is long-only and correlated across names (BTC-hedged controls cut the mean to about 1%), so losses cluster in sustained bear legs. The 8-slot cap is what keeps the drawdown down.

The signal-research diagnostic (IC 0.047, t 1.1) does not separate from noise on its 4-bar forecast grid. It is an execution-path diagnostic and is weaker evidence than the trade-level OOS, but I note it.

There is no stop-loss. Risk rests on the 2-day clock and 8% legs, and the worst adverse excursion was 6.0% of equity.

soft; consistent_with_oos true

advisory only

inconclusive, consistent with OOS (z -0.42)

t 0.72, Sharpe 0.60, 56 trades

t >= 2.0

signal-exit route

t 2.58 / DSR 0.995

Risk Review

Defined risk control holds on the signal-exit route: a fixed 2-day clock exit with capped sizing at 1x leverage. The largest measured position is 8.0% of equity (limit 100%). The worst realised trade loss is 0.97% of equity (AAVE, 2024-04-13, -12.1% on an 8% leg) and the worst adverse excursion is 2.63% of equity. Against the conservative profile, realized volatility is 5.5% (band 5-60%) and peak single-position volatility is 12.3% (TIA, cap 120%). There is no stop-loss and none is required.

Concentrated crash-day exposure. Up to 8 correlated long legs at 8% each means 64% gross long, all opened on the same market-wide sell-off day. The hypothesis's own BTC-hedged control shows the premium is mostly the market-wide crash rebound, so a second leg down within the 2-day hold hits every leg together. The development drawdown of 6.8% (holdout-inclusive 8.3%) is well inside the pre-registered 20%. A tail of about 64% x a further 15-20% alt drop (roughly 10-13% of equity in two days) has not been seen in the sample, because 2022 (LUNA, FTX) predates real HL trading and shows zero exposure in every stress window.

Correlation with the deployed book. At least six paper strategies trade the same crash-rebound or capitulation-absorption premium and fire on the same days: HyperliquidNr7DownsideEventPanel (same venue and similar universe), Nr7FailedDownsideBreakReversalPanel, OkxNr7FailedDownsideBreakReversalPanel, BybitSpotAltPanelCapitulationAbsorption4H, LinkBybitVolumeShockAbsorption1H and SpotLedIdioAltSelloffReboundBtcHedged. Their losses will cluster in the same sustained bear legs. No book member is in breach and there are no open execution items, but several members show diverged decision replays or paper-parity warnings.

Parameter stability is unstable for both tunables. The deployed volume_z_min of 0.59 sits below every walk-forward window's choice (0.76-1.30), so the live filter is looser than any window selected. Final parameters are not statistically accepted (low confidence: OOS significant at t 2.58 and DSR 0.995, but the holdout is inconclusive). The hypothesis flagged that 2026 YTD is flat to slightly negative, which points to possible decay.

Live robustness (from QA): `_due` compares decision_ns minus ts_opened against exactly hold_days x 86,400 s. A live fill stamped after the bar close pushes the exit back one extra day, which turns a 2-day hold into 3 days, the variant the pre-study showed decaying in 2026. A stuck order also delays the clock exit by one bar.

Paper only, under the low-confidence label 'not statistically proven'; no real-money release on this evidence.

Before any live release, fix the clock exit to count bars from the entry decision bar (or apply a half-day tolerance), so live holds match the backtested 2 days.

Portfolio manager: treat this strategy and the other crash-rebound panels (HL NR7 downside, Bybit/OKX NR7 failed-break, Bybit capitulation absorption, LINK volume-shock absorption) as one correlated risk cluster when allocating capital; do not size them as independent.

Outcome Summary

A long-only crash-rebound panel can clear OOS significance with a concentration cap that limits drawdown, but a holdout Sharpe at about half the OOS figure, weak 2026 results (+2.3% YTD) and unstable tuned parameters mean paper results should be watched for decay.

The analyst recommended promotion because pooled OOS significance was clear, economics were about 13x the fee floor, every year was positive and the drawdown was low, while flagging the holdout evidence as low confidence. The run then advanced to paper trading.

Buy up to 8 of 22 point-in-time Hyperliquid perps (8% of equity each) at the close of a UTC day with a <=-2.5-sigma return on >=1-sigma volume, then exit after a fixed 2-day clock, betting that forced and impatient selling is followed by a partial rebound.

The full-history backtest returned +44.3% over 191 long-only trades, with Sharpe 0.88, a 62.8% win rate, profit factor 2.55, an average of 2.49% per trade and an 8.0% max drawdown. Walk-forward OOS was significant (Sharpe 1.15, HAC t 2.58, DSR 0.995, 121 trades); the holdout was positive but not significant (Sharpe 0.60, t 0.72, 56 trades).