Strategies

Hyperliquid AVAX–ATOM Daily Changepoint-Gated Slow Momentum and Fast Reversion

Outcome: Abandoned

HyperliquidAvaxAtomGpMomentumReversion

Outcome Summary

After two iterations, HyperliquidAvaxAtomGpMomentumReversion reached backtest review with functioning capital-relative sizing and long/short participation on AVAX and ATOM. Its 102 closed positions produced a -1.953% net return, -0.185 Sharpe, and -1.043% average net trade return. The reviewer abandoned it for negative expectancy and declined optimization because no measured profitable parameter region or specific implementation defect justified further iteration.

Strategy report

Hypothesis

Derived by the Research Lead from Kieran Wood, Stephen Roberts and Stefan Zohren's Slow Momentum with Fast Reversion: use online Gaussian-process changepoint severity to switch between persistent-trend exposure and short-horizon contrarian exposure. This is an interpretable operationalization, not a reproduction of the authors' learned LSTM policy or reported performance. Trade AVAX and ATOM independently, long or short, on 1-DAY bars. Preserve the paper's GP changepoint construction and 60-day volatility estimator; replace the learned policy with explicit regime rules and add turnover controls, protective stops and equity-relative exposure limits. These choices are registered below. Execute market orders at signal-bar closes; this is the registry's equivalent of next-bar-open execution on a continuous venue. Parameters remain frozen during each evaluation window and deployment.

Long-only exposure is the largest directional imbalance at 80.9%, while Hyperliquid represents 10.2% against its 20% target. This two-name long-short hypothesis addresses both and reduces single-instrument dependence. Corpus concentration is empty, so neither instrument is asserted to be underused. Daily bars avoid Hyperliquid's short intraday history. The mechanism differs from failed ATOM extreme-funding entries and generic RSI reversion: no supplementary-data conjunction is required, and reversion is conditional on online covariance disequilibrium. Options and cross-venue quotas remain unfilled because introducing option legs or venue arbitrage would change the source mechanism.

negative_expectancy: decline optimization of this explicit AVAX–ATOM daily GP-gated operationalization. The first measured backtest has negative realized expectancy across 102 positions, with average trade return -1.043% against the +0.15% review floor and PF 0.900. No measured profitable parameter region or specific implementation defect supports developer iteration. This rejects the registered operationalization on these targets, not the authors' learned policy or the broader changepoint mechanism; no optimization sweep has been evaluated.

Implementation

Trades AVAX and ATOM independently on Hyperliquid daily bars, switching between slow momentum and fast contrarian exposure using online GP changepoint severity and location. Uses volatility-scaled, equity-relative sizing, ATR protective stops, holding limits and cooldowns. Rejects entries with invalid normalized stops and immediately closes filled exposure when protection cannot be installed.

Preserves the previous GP, continuous signals, alignment, exposure limits and research parameter permissions. Fixes the critical QA defect by checking precision-normalized stop validity before entry and against actual fills; protection failures cancel remaining orders and immediately request closure, including late partial fills and rejected or denied stops. Synthetic stationary-plus-changepoint fitting benchmarks at permitted lookbacks 10/21/63 took 0.0934/0.0750/0.4724 seconds respectively on one standardized random window each. A trial triangular implementation was slower at 63 and altered the optimizer endpoint, so it was not adopted. These microbenchmarks do not resolve the measured 166.55-second sandbox runtime; engineering runtime review remains required without reducing model scope or data coverage. Sandbox negative expectancy warrants full-history analysis, rather than an economic verdict from this correctness iteration. The brief supplies previous code but no previous configuration; tunable initial values and solver controls in this output are explicit reconstructed baselines within the registered domains. No market history or final-exam data was read. The requested outbox file could not be written because the environment enforces a read-only filesystem; the complete call-bound JSON is returned here.

Verification Results

Request engineering profiling of likelihood evaluation and solver execution while preserving the registered GP construction, instrument coverage, history and parameter domains.

Repeated online stationary and changepoint GP fits produced a 223.85-second sandbox runtime, exceeding the supplied 120-second baseline target. Optimization multiplies this computational cost.

The sandbox reports -0.424% average net return per closed position and profit factor 0.982. Full-history analysis should assess momentum and reversion regimes separately against Hyperliquid's approximately 0.09% round-trip taker commissions plus impact. These smoke-test results do not establish a correctness failure.

Backtest Review

Sharpe: -0.19
Total return: -1.95%
Max drawdown: 6.04%
Trades: 102
Win rate: 44.1%
Profit factor: 0.90

102 closed positions across both instruments over 1333 evaluated days; capital-relative sizing and long/short participation are functioning.

Accounting reconciles, with no liquidation or end-position unrealized profit.

Average net trade return is -1.043% versus the supplied Hyperliquid review floor of +0.15%; PF is 0.900, Sharpe -0.185 and total return -1.953%.

Trade details confirm losses on both instruments: AVAX -$1330 and ATOM -$623; both long and short positions lose overall.

QA's economic concern is borne out. Regime labels are absent from trade records, so momentum-versus-reversion attribution cannot be independently established.

The one-day forecast diagnostic is uninformative (IC -0.0021, t=-0.08), though it does not directly test the longer holding horizons. Benchmark comparison is not meaningful for this specification.

Iteration History

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] The sandbox reports -1.249% average trade return and PF 0.900 across 117 closed positions. These warrant full-history analysis but do not establish a correctness failure. - [critical] A long fill receives no protective stop when stop_atr_multiple × entry ATR is greater than or equal to its fill price. Sizing permits this case, but the resulting nonpositive stop_level silently bypasses stop submission. After a large price decline, historical ATR can exceed this bound. The position then retains only signal/time exits, contrary to the specified protective-stop behavior. (line 369) — fix: Reject long entries whose intended, precision-normalized stop price is nonpositive. Revalidate against the actual fill price and immediately close any filled exposure for which a valid protective stop cannot be installed. - [warning] The sandbox took 166.55 seconds against the 120-second baseline target at cpd_lookback=21. The permitted 63-day GP window increases dense matrix costs substantially and requires engineering runtime review. (line 140) — fix: Profile GP fitting at every permitted lookback and optimize the numerical implementation while preserving the registered model, data coverage, and parameter permissions.

Reject long entries whose intended, precision-normalized stop price is nonpositive. Revalidate against the actual fill price and immediately close any filled exposure for which a valid protective stop cannot be installed.

A long fill receives no protective stop when stop_atr_multiple × entry ATR is greater than or equal to its fill price. Sizing permits this case, but the resulting nonpositive stop_level silently bypasses stop submission. After a large price decline, historical ATR can exceed this bound. The position then retains only signal/time exits, contrary to the specified protective-stop behavior.

Profile GP fitting at every permitted lookback and optimize the numerical implementation while preserving the registered model, data coverage, and parameter permissions.

The sandbox took 166.55 seconds against the 120-second baseline target at cpd_lookback=21. The permitted 63-day GP window increases dense matrix costs substantially and requires engineering runtime review.

Outcome Summary

This explicit rule-based operationalization failed on these targets; the result does not establish failure of the authors’ learned policy or the broader changepoint mechanism.

Backtest review returned abandon for negative expectancy, with average trade return below the +0.15% review floor and no measured profitable parameter region or specific implementation defect supporting iteration. Optimization and subsequent analyst and risk reviews were not reached.

Trade AVAX and ATOM independently on Hyperliquid daily bars, using Gaussian-process changepoints to switch between slow momentum and fast reversion with volatility-based sizing and protective stops.

Across 1333 evaluated days and 102 closed positions, net return was -1.953%, Sharpe was -0.185, and maximum drawdown was 6.041%. Average net trade return was -1.043%, profit factor was 0.900, and both instruments lost money overall.

Analysis

Slow Momentum with Fast Reversion: A Trading Strategy Using Deep Learning and Changepoint Detection

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