Hyperliquid Sequential Drift-Change Events: CUSUM Persistence on APT, ICP, FIL and INJ

Hypothesis

Original research hypothesis: detect sustained shifts in standardized 4-hour returns with a two-sided cumulative-sum detector, rather than an indicator crossover, single shock candle, cross-sectional ranking or price-ratio reversion. Independently trade four predeclared perpetuals, long or short, only after sequential evidence accumulates; hold for 72 hours unless the protective stop intervenes. Use all available actual 4-HOUR history; do not assume the API can reconstruct earlier history. Market orders execute at the completed signal bar close, the capability registry equivalent of next-bar-open execution on continuous venues. Stops are platform-matched intrabar protective stops. No exchange-hosted protection, hedge mode or selectable same-bar fill policy is required. Capture and event-rate figures are predictions to test, not pre-study findings.

Hyperliquid is 13.4% versus its 20% venue target, while long-only is 75.6% against a 55% ceiling. This independent four-name long-short event panel addresses both imbalances without adding to Binance USD-M concentration. Options have the largest venue deficit, but the supplied feasibility record argues for establishing a reliable option-leg path before another autonomous structure proposal. APT, ICP, FIL and INJ have only 1, 2, 4 and 4 Hyperliquid runs respectively in the queried snapshot. corpus_concentration was empty, so no recent instrument/timeframe shares can be inferred. Snapshot FTS searches returned no CUSUM prior art; variance-ratio persistence and LTC/BCH convergence were already tried and are deliberately excluded. This is sequential drift detection with mandatory multi-bar accumulation and fixed-clock exposure, not another variance-ratio gate or continuous momentum ranking. No supplied library excerpt provides this rule; this proposal is not attributed to those sources.

Measured negative expectancy over the visible development period: 124 closed positions across 80 entry dates average -1.798% net, with profit factor 0.513. Both directional arms and every instrument lose, including negative aggregate price PnL. There is no demonstrated implementation defect or evidence-supported improvement path justifying optimization. This rejects the tested operationalization; no conclusion is drawn about the withheld holdout.

Implementation

Independent two-sided CUSUM drift events on APT, ICP, FIL and INJ completed 4-hour bars. Signal-close market execution is the continuous-venue equivalent of next-bar-open execution. Frozen entry ATR intrabar protective stops, 18-bar holding periods and six-bar post-exit cooldowns.

Implements prior-return sample volatility, separate sequential evidence and run counts, Wilder ATR, prior-bar median quote-volume capacity and ordered MTM-equity risk allocation. These are independent events, not hedge legs. Evaluation boundaries and historical warmup are platform-owned; no research-date guard or synthetic history is introduced. Funding and execution costs remain platform-owned.

Backtest Review

Sharpe

-1.47

Total return

-0.34%

Max drawdown

0.44%

Trades

163

Win rate

39.3%

Profit factor

0.56

The code implements prior-volatility CUSUM accumulation and capped capital-relative sizing.

The ledger shows 124 closed positions across 80 entry dates, including 63 longs and 61 shorts. Of these, 81 lasted exactly 72 hours, consistent with the fixed holding rule.

Average net trade return is -1.798%, profit factor 0.513 and Sharpe -1.741.

Both directional arms lose: longs -$197.43 and shorts -$99.82. All four instruments lose.

Aggregate price PnL is -$276.37, with an additional -$20.87 funding contribution: losses are not merely execution-cost drag.

Closed-trade losses of $297.24 broadly reconcile with the approximately $302.19 marked-to-market account loss. Small positions limit account damage; larger sizing would magnify negative expectancy.

Outcome Summary

This tested CUSUM operationalization needs evidence of positive directional expectancy before parameter optimization is justified.

Backtest review recommended abandonment because both directions and every instrument lost, including negative aggregate price PnL, without a demonstrated implementation defect or evidence-supported improvement path. Optimization and later stages were not reached; the withheld holdout was not assessed.

Trade persistent drift long or short in APT, ICP, FIL and INJ perpetuals using CUSUM evidence from standardized 4-hour returns, holding for 72 hours unless stopped.

The supplied backtest reports -33.89% return, -1.47 Sharpe, 163 closed trades and 44.35% maximum drawdown. The abandonment review cites a different ledger of 124 closed positions averaging -1.798% net, with profit factor 0.513.