SolHlActivitySurgeExhaustionFade

SOL Hyperliquid Activity-Surge Exhaustion Fade, Single-Leg Long/Short, 4H Bars, Fixed 72h Clock Exit

Single-instrument version of the source's volume-change factor on SOLUSD.HYPERLIQUID 4H bars (about 2.6 years of history). An attention surge is when the 24h mean volume is at least 1.8x the 96h mean volume. When that happens, fade the 24h move that came with it and exit on a fixed 72h clock. The strategy is flat whenever no surge is active; expected time in market is about 30-40%. Mechanism without instrument or timeframe: after an abnormal short-window volume surge, fade the concurrent return on a multi-day fixed clock. This is the opposite claim to the ETH volume-confirmed momentum survivor, so it tests that survivor directly. Selection choices: HL for the venue quota, SOL as the most liquid non-BTC/ETH HL perp, and 4H because HL 1h only has about 7 months. The 24/96 windows are copied from the source.

This is a single-instrument version of the source's volume-change factor, turned into a fade. The surge ratio and the 24h return are updated each bar from bounded deques, so per-bar cost stays fixed. The signal is -ret24h * surge_ratio: it changes every bar and its sign gives the fade direction. Entry needs surge_ratio >= 1.8 and a move of at least 1%, so the faded move is larger than the ~0.09% Hyperliquid round-trip cost. The 72h exit compares the bar timestamp with position.ts_opened, so it follows calendar time and survives a restart. The venue is futures because the strategy goes short. The hypothesis gives no stop-loss, so none is added. Because the 24h window sits inside the 96h window, the ratio cannot go above 4.

Trades SOLUSD.HYPERLIQUID on 4H bars, long or short, one instrument. An activity surge is when the 24h mean volume (6 bars) is at least 1.8x the 96h mean volume (24 bars). During a surge where the 24h price move is at least 1%, the strategy opens against that move: long after a drop, short after a rally. Position notional is 0.9x equity at 1x leverage. It exits on a fixed 72h clock counted from the position's open time, and is flat the rest of the time.

Keep it and document it as a fee filter, or let the optimizer take it toward 0 so the pure-hypothesis variant is also measured.

The code adds a filter the hypothesis does not name: |24h return| >= min_abs_ret (1%). The hypothesis fades the concurrent 24h move whenever a surge is active. The filter is a reasonable fee guard, since a 1% move is well above the ~0.09% HL round-trip, and it does not change the mechanism. It does thin the sample slightly compared with the stated factor.

Optional: add a catastrophic stop via a bracket order (e.g. 15-20%) if the risk review asks for one. Do not add it silently, because it would change the tested hypothesis.

The only exit is the fixed 72h clock. There is no stop, and the hypothesis specifies none, so this matches the hypothesis. With 0.9x equity notional, a fade into a continuing trend can take a large single loss; the sandbox's largest loss was about -$25k, roughly 25% of starting equity. Gross exposure is bounded at 0.9 x leverage (1.0) x 1 leg = 0.9x equity. There is no unbounded scalar and the strategy is flat between trades, so this is not the compounding-blowup pattern.

None needed.

The min_bars_required override is the documented extension point ('Override as needed') and returns 26, which is enough for the 24-bar long window plus the 7-close return window. The static-analysis base_shadow warning is benign here.

The sandbox's +84.8% comes almost entirely from the stressed-vol tercile (+120.6%, Sharpe 2.46). The normal tercile lost 17.6% and the calm tercile was flat. There are only 38 trades and the Sharpe CI of [-0.45, 2.23] includes zero, so the edge may just be a few crash-rebound fades (e.g. Aug 2024). funding_folded=false with 0 funding events, so Hyperliquid funding on the 72h holds at 0.9x notional is not in PnL. The analyst should check that funding drag and whether the result survives outside high-vol windows. The ledger reconciles: 2.38% x 38 = ~90% implied vs 84.8% actual, and PF 1.66 agrees in sign with the return.

Economics are well clear of the HL fee floor: avg_trade_return_pct 2.38% vs a 0.15% floor, PF 1.70, win rate 0.61, and commission is only 2.0% of gross. Impact is 4.4%.

Funding IS credited in this full-history run (funding_folded=true, 2684 funding events; the first trade alone carries +$1.28k funding). QA's worry that funding drag on 72h holds at 0.9x notional was left out came from the sandbox and does not apply to these numbers.

The trades match the mechanism. Every one is a fixed ~72h hold that fades the 24h move on a surge: shorts after rallies, longs into selloffs. Split is 26 long / 12 short. entry_diag shows 38 signalled and 38 submitted with no drops.

Positive in both full years (2024 +36.7%, 2025 +37.1%). Beta is 0.05, so the return is not market exposure. end_unrealized_pct is 0.07, so the result is not an open-position artifact.

Sharpe 0.95 and PF 1.70 sit at the base-run floor used as advisory context (Sharpe ~1.0 / PF 1.5). A 2.4%-per-trade edge is not the fee-starved shape.

QA's regime concern is partly borne out. The stressed-vol tercile makes +111% (Sharpe 2.40), the normal tercile loses 14.9% and the calm tercile is roughly flat. The edge looks like crash/spike exhaustion, not a general surge fade.

38 trades is thin. The Sharpe CI [-0.30, 2.33] straddles zero and PSR is 0.80. The intraday sample floor at promotion is 100 trades, which this configuration cannot reach over 2.2 years.

Max DD is 19.4% (CI up to 57%) and the longest drawdown lasted 177 days. Position is 0.9x equity per trade, and the worst single trade was -20.5%.

The daily_returns series shows paired, equal-and-opposite blips on flat days (e.g. -0.3182 / +0.3192) about every 8 days. They net to about zero. It looks like a metrics artifact, but I have not verified the mechanism; flag it for engine review.