Hyperliquid Correlation-Contagion Episodes: Six-Alt Common-Factor Continuation

Hypothesis

Derived by the Research Lead from Rich C and the answers to Equity Risk Model Using PCA (https://quant.stackexchange.com/questions/1020/equity-risk-model-using-pca): operationalize the dominant correlation factor as a detector of market-wide synchronization, then test directional continuation for three days after synchronization rises. The source discusses risk estimation, not this trading rule or its profitability. Use synchronized 4-HOUR bars on six Hyperliquid perpetuals. 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, atomic baskets, options execution or queue advantage is required. Fit only six-dimensional, bounded-window matrices. This is an event-entry common-factor continuation test, not lowest-variance PCA spread reversion or continuously rebalanced cross-sectional momentum.

The largest directional imbalance is long-only at 75.6% against a ceiling of 55%; a symmetric directional event basket addresses it while also filling Hyperliquid at 13.4% against a 20% target and reducing single-symbol concentration. Options have a larger venue shortfall, but the supplied feasibility record identifies unresolved option-leg execution failures; this proposal uses supported linear perps. Corpus concentration was supplied as an empty object, so instrument-specific shares cannot be verified. Prior-art searches found failed Hyperliquid LTC/BCH/ETC PCA reversion and failed APT/ICP/FIL/INJ CUSUM drift events. The present test instead requires a measurable increase in cross-asset correlation concentration, positive common loadings and broad directional agreement; it does not assume residual convergence or transfer those results. The six names form an explicitly fixed research panel, not a claim about the historical top or representative crypto universe.

underpowered: closed under the inconclusive policy at analyzing; not a verdict on the idea. Closed as underpowered. This is not a verdict on the idea. It closes this six-name Hyperliquid 4H operationalization, which produced 4 episodes in 2.6 years and none in the holdout. Any retest needs a new hypothesis. Options: (a) test the ungated factor-move plus agreement trigger as its own hypothesis; (b) use a wider point-in-time universe or a venue with longer history, such as Binance USD-M from 2020; (c) test whether entering on the capitulation or rebound bar is the failure mode, for example by requiring the signal to persist one more bar.

Implementation

Six-alt Hyperliquid common-factor continuation basket with synchronized 4-hour signals, fixed intrabar protective stops and an 18-interval holding limit. Market entries fill at the completed signal-bar close, the continuous-market equivalent of next-bar-open execution.

Only min_bars_required changes to 1 so aligned warmup bars initialize the existing rolling history immediately. Internal readiness checks remain unchanged: one initial price bar, 252 prior returns, and seven concentration observations make the full condition evaluable on aligned bar 260 (176 at the default lookback). The frozen 260-bar warmup, trading rules and previous config are preserved. Pipeline verification remains responsible for execution checks.

Backtest Review

Sharpe

0.13

Total return

0.17%

Max drawdown

0.68%

Trades

18

Win rate

38.9%

Profit factor

1.34

The ledger shows three synchronized six-leg baskets: one short and two long, with equal entry notionals within each episode. Basket exits coincide with a leg reaching its 5% protective stop, consistent with the unwind rule.

Closed-trade PnL reconciles to approximately +$172.79. Average net leg return is 0.847% and profit factor is 1.336, supporting further investigation.

Eighteen positions represent only three independent episodes, approximately 1.8 annually versus the predicted 15–30. Statistical evidence is extremely thin.

The short arm has only one losing episode. Sharpe is 0.162, and the predicted incremental benefit of synchronization gating remains unmeasured.

Stop-triggered exits occur after 4–20 hours rather than the 72-hour maximum. These observations cover only the visible development period through 2025-11-13.

Iteration History

Allow calculate_signal() to run from the first aligned bar (for example return 1 from min_bars_required()), retaining its existing internal readiness checks. This makes full readiness occur at window + rise + 2 bars: 176 at defaults and 260 at the maximum lookback. Verify readiness after exactly 260 aligned warmup bars for correlation_lookback=252 without changing the frozen research contract.

min_bars_required() delays initialization of the rolling signal history until window + rise + 2 primary bars have already arrived. FactoryStrategy._maybe_process_primary returns before calculate_signal() while below that minimum, including during warmup. calculate_signal() then starts collecting prices, window returns and rise+1 concentrations from scratch. With uninterrupted aligned bars, the first fully evaluable condition occurs at bar 351 for window=168, and bar 519 for window=252, rather than within the contract's 260-bar warmup. Thus each reset suppresses valid entry episodes inside the measured window, and changing the lookback also changes the unintended initial no-trade period. The passing smoke test does not establish compliance with the frozen boundary/warmup contract.

Outcome Summary

A synchronization gate on six names with only about 2.6 years of Hyperliquid history fires far too rarely (about 1.8 episodes a year against the predicted 15-30) to test anything, so a retest needs a wider point-in-time universe or a venue with longer history, such as Binance USD-M from 2020, or it should test the ungated factor-move-plus-agreement trigger on its own.

Backtest review sent it on to optimization, but at the analyzing stage it was closed as underpowered under the inconclusive policy: it produced 4 episodes in 2.6 years and none in the holdout. The pipeline said explicitly that this is not a verdict on the idea.

It used the dominant correlation factor across six Hyperliquid alt perpetuals (LINK, AAVE, UNI, AVAX, ATOM, DOT) on 4-hour bars to spot market-wide synchronization, then traded the basket in the common factor's direction for up to three days after synchronization rose, with 5% protective stops on each leg.

The full-history backtest (2024-03-26 to 2026-10-23) returned +17.3% total (6.7% CAGR) with a Sharpe of 0.13 (95% CI -1.08 to 1.24) and a profit factor of 1.34. It closed 18 positions, which came from only three six-leg basket episodes, and had a 67.9% max drawdown. The signal-research diagnostic rejected the forecast as carrying no information (IC 0.074, t=1.3).