Strategies

Hyperliquid BCH–LTC Relative Shock Exhaustion: Hedged Multi-Day Convergence

Outcome: Abandoned

HyperliquidBchLtcShockExhaustion

Outcome Summary

HyperliquidBchLtcShockExhaustion tested whether large, volume-backed relative moves between BCH and LTC partly reverse once they stop accelerating, using equal-dollar long/short pairs on Hyperliquid 4h bars. Its first backtest returned +4.58% with Sharpe 0.37 and an average trade return of 0.164%, just above the 0.15% Hyperliquid floor, so the backtest review sent it to optimization as a test, not as evidence of robustness. Walk-forward optimization found no real edge: out-of-sample Sharpe averaged about 0.06 against about 2.6 in-sample, parameters were unstable with cliffs, and all the profit came from 2024. The analyst abandoned it after one iteration, describing it as overfit with no edge, the fifth BCH/LTC relative-shock attempt.

Strategy report

Hypothesis

Test whether unusually large, volume-backed relative advances between BCH and LTC partially reverse after the advance stops accelerating. Trade synchronized 4-HOUR perpetual bars, buying the relative laggard and shorting the leader with equal dollar notionals. Both exact bar directories exist locally. This is a conditional sector-relative convergence experiment, not an assumption that the price ratio is permanently cointegrated. Inspiration: [QuantInsti's Principal Component Analysis in Trading](https://blog.quantinsti.com/principal-component-analysis-trading/), specifically separating shared exposure from asset-specific movement; this proposal uses a transparent fixed log ratio rather than fitted PCA. Evaluate only the instruments' overlapping history.

Hyperliquid is the most underrepresented venue relative to its target: 8.5% versus 20%. A two-instrument long/short proposal also addresses excessive single-symbol and long-only concentration. corpus_concentration is empty, so individual instrument shares cannot be assessed; no claim of zero prior use is made. Unlike the failed DOGE inter-exchange spread, this experiment targets multi-percent differences between assets over days, rather than tiny same-asset venue mispricings. Unlike the failed ATOM rebound, it does not buy after an absolute-price bounce or exit at a rolling midpoint. The supplied success labels include negative Sharpes and are not treated as evidence of profitability.

Abandoned on measured evidence of no out-of-sample edge (overfit / no_edge), not merely on insufficient significance. Across 13 chronological walk-forward windows, OOS Sharpe averages about 0.06, against a 0.5 floor and inner IS of about 2.6. Two of three tunable parameters are sensitivity cliffs, and inner PBO is above 0.6 in 9 of 13 windows. The parameters selected in each window move across the whole search box. In the full-period backtest, all the profit is from 2024 (+9.5%); 2025 lost 3.8% and 2026 YTD lost 0.7%. The executed forecast IC is -0.033 (t=-0.34), with net edge per horizon of -0.34%. Fee-floor comparison: avg_trade_return_pct 0.164% vs the 0.15% Hyperliquid floor, the 0.18% economic floor and the 0.30% pre-registered target. PF is 1.09 against 1.15 pre-registered, and impact consumes 66% of gross. The small headline edge is too thin to survive costs once the 2024 regime is gone. This matches the prior BCH/LTC relative-dislocation attempt (no_edge, Sharpe -0.80). The verdict covers fixed-log-ratio BCH/LTC relative-shock exhaustion on Hyperliquid 4h. Research should stop proposing BCH/LTC relative-shock variants (now 5 attempts).

Implementation

Trades exactly synchronized BCH and LTC Hyperliquid four-hour perpetual bars. Fades a volume-backed relative shock once its one-bar relative advance slows, buying the laggard and shorting the leader with equal target dollar notionals. Requires 360-bar return correlation of at least 0.55. Closes both legs at 2 log-percentage points of convergence, 4 points of adverse spread movement, or 42 four-hour intervals; waits six intervals before re-entry. Uses bounded buffers and incremental correlation and volume statistics.

Implements conditional relative-shock exhaustion without fitting PCA or assuming permanent cointegration. Shock and spread exits are measured as 100 times changes in log(BCH/LTC). Leader volume is the latest rolling six-bar block divided by the average of the preceding 30 nonoverlapping six-bar blocks. Exhaustion means the latest leader-direction relative log return is smaller than the preceding return. Both legs contribute only at identical timestamps, restricting signals to overlapping history. Registered fixed values and tuning permissions are preserved. Risk sizing allocates at most 20% of equity per leg and 40% gross, targeting 0.8% equity loss at the spread stop; spread stops execute at synchronized bar closes and cannot guarantee that loss ceiling. Separate leg orders are not atomic, and unpaired positions are flattened at the next eligible callback. Holding and cooldown durations use timestamp-anchored four-hour slots. Verification remains the pipeline's responsibility. The requested outbox file could not be written because this session's filesystem permissions are read-only; the complete bound JSON is returned here.

Verification Results

Document this as a bar-close risk target rather than a guaranteed loss ceiling. If intrabar protection is required, implement coordinated pair protection without duplicating exits.

The adverse-spread stop is checked only at synchronized four-hour closes. Intrabar moves and gaps can exceed the nominal 4% spread stop and the targeted 0.8% equity loss.

Add execution-event handling for rejected or partially filled legs and a clock-driven feed watchdog before live deployment.

Leg submissions are sequential, and orphan repair runs on bar callbacks. A rejected or delayed leg can leave directional exposure until another eligible callback. The stale-feed guard also requires incoming bars; it cannot detect both feeds stopping together.

Sandbox impact consumed 66.57% of gross profit, and funding cash flows were not modeled. Full-history evaluation should examine these costs; neither establishes a correctness defect.

Backtest Review

Sharpe: 0.37
Total return: 4.58%
Max drawdown: 8.18%
Trades: 146
Win rate: 60.3%
Profit factor: 1.09

Trade details confirm 73 paired entries over 753 days: opposite directions, equal entry notionals and synchronized exits. Frequency is plausible for conditional relative-shock convergence.

Net return is +4.58%, PF 1.091 and average leg return 0.1635%, above the 0.15% Hyperliquid review floor. No open-position profit drives the headline.

Funding is modeled and folded into PnL; trade records show approximately +$179 net funding, resolving QA's sandbox omission.

QA's impact concern is borne out: modeled impact consumes 65.68% of gross profit, leaving fragile economics.

Sharpe is 0.371 with a confidence interval spanning zero; 2025 and partial 2026 lost money. Optimization is warranted as a test, not evidence of robustness.

The effective sample is 73 pair events, not 146 independent trades. Evaluate paired outcomes and preserve chronological validation.

Entry-condition timing was not independently checked against underlying bars.

The bound outbox file could not be written because this session permits only filesystem reads.

Outcome Summary

A fixed-log-ratio BCH/LTC relative-shock edge that only made money in 2024 and lost most of its gross profit to impact does not hold up; after five attempts, research should stop proposing BCH/LTC relative-shock variants.

The backtest review sent it to optimization, but the analyst then abandoned it for having no out-of-sample edge: across 13 walk-forward windows, out-of-sample Sharpe averaged about 0.06 against a 0.5 floor, while in-sample Sharpe was about 2.6. Two of the three tunable parameters showed sensitivity cliffs, the probability of backtest overfitting (PBO) was above 0.6 in 9 of 13 windows, and the forecast information coefficient was -0.033 (t=-0.34).

On synchronized Hyperliquid 4-hour perpetual bars, it waited for an unusually large, volume-backed move in the BCH/LTC log ratio to stop accelerating, then bought the laggard and shorted the leader in equal dollar amounts, betting the gap would partly close.

The full-period backtest (2024-03-29 to 2026-04-20) made +4.58% over 146 legs (73 paired trades), with Sharpe 0.37, profit factor 1.09, 60.3% win rate, 8.18% max drawdown and an average trade return of 0.164% of notional. All of the profit came in 2024 (+9.5%); 2025 lost 3.8%, 2026 year-to-date lost 0.7%, and modeled market impact took about 66% of gross profit.

Backtest and paper results are hypothetical. Trading involves risk of loss.