Hyperliquid LTC–BCH–ETC PCA Basket Deviation Reversion

Hypothesis

Test the article's single concrete mechanism: buy or sell a signed, lowest-variance PCA basket when its standardized-price spread deviates, then close on convergence. Adapt the FX example to LTC, BCH and ETC Hyperliquid perpetuals using synchronized 4-HOUR bars. These related proof-of-work assets provide a testable shared-factor basket without using BTC. Replace broker-lot allocation with equity-relative quantities that reproduce the spread's price sensitivity. Fit statistics using preceding observations, excluding the signal bar, and freeze the component and its normalization during each trade. These are explicit adaptations drawn from the article's exposure-balancing and separate-training-window extensions. Market orders execute at the completed signal bar's close. Execution is sequential, with compensating closures after incomplete entry; atomic execution is not assumed. No profitability claim from the source is adopted.

The source contains one distinct trading mechanism, so only one hypothesis is proposed. Long-only exposure is the largest portfolio imbalance at 81%; a signed multi-leg basket addresses that imbalance and the excess single-symbol concentration. Hyperliquid is also below its 20% target. Four-hour bars avoid its short sub-hourly history limit. Corpus concentration was supplied empty, so individual-symbol underrepresentation cannot be verified. This is a three-asset PCA construction, distinct from the failed fixed ARB–OP pair, while retaining the submitted mechanism rather than replacing it with momentum or funding carry.

negative_expectancy (code: fee_edge/no_edge): first measured backtest shows avg_trade_return_pct -0.161% vs the 0.15% HYPERLIQUID floor, PF 0.917, Sharpe -0.79, DD 39% (vs 15% pre-registered), over 189 baskets / 567 legs on 732 days. Modeled impact is 150% of gross with capacity of only $44.5k, so shrinking size cannot produce a deployable edge; the signal's own forecast IC is -0.089 (t=-0.9) and returns worsen each year (2024 +15%, 2025 -12%, 2026 YTD -28%). Optimization cannot lift a basket-reversion signal with no measurable information above costs on these thin HL legs. QA's edge concern was borne out. Scope: LTC-BCH-ETC lowest-variance PCA standardized-price spread on Hyperliquid 4H; does not speak to PCA baskets on deeper-liquidity venues.

Implementation

Trades the signed lowest-variance PCA basket of LTC, BCH and ETC Hyperliquid perpetuals on synchronized four-hour closes. Fits preceding standardized prices, enters at absolute z-score 2 with projected convergence capture of at least 0.4% of gross notional, and exits on convergence, a 1% entry-equity basket loss, or 42 held bars. Gross exposure is capped at 60% of equity. Entry legs execute sequentially; failed or incomplete entry triggers compensating closures.

Preserves the version-2 frozen parameters and the proposed PCA deviation-reversion mechanism. Training excludes the signal bar and matches all three historical timestamps exactly. Component loadings, price normalization and spread normalization remain frozen throughout entry, holding and closure. Quantities are proportional to component loading divided by training price standard deviation, reproducing the spread's price sensitivity subject to venue quantity increments. The 1% risk fraction is interpreted as a basket loss budget, with initial sizing additionally limited by an entry_z-sized adverse spread move; this bar-close limit cannot guarantee the realized loss under gaps or execution delays. Refit cadence uses calendar-time four-hour buckets. No verification was run. The requested outbox file could not be written because the workspace permits read access only; the complete task-bound JSON is returned here.

Verification Results

Add an independent stale-feed watchdog that blocks entries and initiates compensating closure when synchronization remains unavailable, without substituting stale prices into the PCA signal.

Active-basket management requires all three current bars to align. The base-class synchronization barrier also defers calculate_signal when a leg is missing. During a feed gap, the basket loss check, holding counter and compensating-closure retry can therefore stop advancing while exposure remains open.

The sandbox reports -0.16735% average net return per closed leg and impact costs equal to 154.6% of gross profit. Assess liquidity-sensitive sizing and basket-level realized convergence capture in the full backtest. The 0.4% projected capture filter exceeds the documented 0.09% round-trip commission on gross notional, but does not account for impact or guarantee convergence.

Backtest Review

Sharpe

-0.79

Total return

-27.28%

Max drawdown

38.98%

Trades

567

Win rate

54.0%

Profit factor

0.92

Mechanism implemented as hypothesized: 189 three-leg baskets (567 closed legs), balanced 276 long / 291 short legs, preceding-window PCA fit, frozen normalization, sequential entry; no artifact signs (end_unrealized 0, accounting reconciles).

Adequate sample over 732 days of synchronized 4H bars; metrics_reliable true.

Negative expectancy on a reliable sample: avg_trade_return_pct -0.161% vs HYPERLIQUID floor +0.15%; PF 0.917; Sharpe -0.79 (CI -2.09 to 0.61); total return -27.3%; max DD 39.0% vs pre-registered 15% constraint; drawdown duration 462 days.

Costs swamp the gross spread capture: modeled impact $82.0k = 150% of gross, plus $13.4k commission; capacity_usd only $44.5k, i.e. at the backtest's 60%-gross sizing the edge is already past capacity, and any size small enough to avoid impact is toy scale.

Deteriorating through time, not regime-stable: annual +15.1% (2024) / -12.1% (2025) / -28.1% (Jan-Apr 2026); the only positive stretch is Nov-Dec 2024 (+19%) plus isolated months (Aug/Nov 2025). Loses in calm (-15.6%) and normal (-19.4%) vol terciles.

Executed-forecast diagnostic: IC -0.089 (t=-0.9), net edge -0.78% per horizon — no detectable information in the deviation signal.

Per-leg attribution: BCH -$36.9k, LTC -$15.3k, ETC +$24.9k — no consistent convergence capture across the basket.

Benchmark not meaningful (market-neutral basket); judged on absolute metrics. QA's edge concern (sandbox -0.167%/leg, impact 155% of gross) is borne out on the full history.

Outcome Summary

On thin Hyperliquid alt legs, impact costs for a three-leg basket can exceed the gross spread capture, so check capacity and impact against the expected convergence before building basket-reversion strategies there; this result says nothing about PCA baskets on deeper-liquidity venues.

At backtest review the analyst voted abandon for negative expectancy (fee_edge/no_edge): the -0.161% per-trade return was below the 0.15% Hyperliquid floor, and the 39% drawdown exceeded the pre-registered 15% limit. The signal's forecast IC was -0.089 (t=-0.9), estimated capacity was only $44.5k, and annual returns fell each year (2024 +15%, 2025 -12%, 2026 YTD -28%), so optimization and later stages never ran.

Mean-reversion on the lowest-variance PCA basket of standardized LTC, BCH and ETC Hyperliquid perpetual prices on 4-hour bars, entering when the spread z-score deviated (entry_z 2.0) and exiting on convergence (exit_z 0.5), with the fit using only preceding bars and normalization frozen during each trade.

Over 732 days (189 three-leg baskets, 567 closed legs) it returned -27.3% with Sharpe -0.79 (CI -2.09 to 0.61), profit factor 0.917, avg_trade_return_pct -0.161% and max drawdown 39.0%. Modeled impact ($82.0k) was 150% of gross PnL, plus $13.4k in commission.

Analysis

From Novice to Expert: Trading Multi-Symbol Basket