Strategies

Hyperliquid SOL Daily Time-Series Momentum Long-Only (Simple TSMOM on Less-Crowded Venue)

Outcome: Abandoned

HyperliquidSolDailyTimeSeriesMomentumLong

Outcome Summary

This strategy applied canonical long-only daily TSMOM to SOL on Hyperliquid, explicitly engineered to dodge the prior ADA 4H optimization collapses by switching timeframe, instrument, and venue, and notably with clean sizing (123% ≈ intended 120%, no over-exposure defect). Its headline looked promising (+77.4%, Sharpe 1.18, Sortino 17.7), but the entire edge came from a single suspected-bad daily bar on 2024-04-17 showing an impossible +459% return, with every other year losing and unmistakable artifact signatures (skew 9.67, kurtosis 95, 730% vol, -7.88 Sharpe CI low, 14.7% win rate). The analyst abandoned it at the backtest-review gate on its second iteration as a data artifact rather than a real edge — recommending the data team verify the 2024-04-17 HL SOL OHLC — and noted the ~3-year/34-trade sample was too thin to validate anyway, so it never advanced to optimization or risk review.

Strategy report

Hypothesis

A single-instrument, single-timeframe, single-venue, long-only time-series momentum strategy on SOLUSD.HYPERLIQUID DAILY bars. Implements the canonical Moskowitz-Ooi-Pedersen (2012) TSMOM formulation: when trailing N-day return is positive AND price is above its long moving average, hold a long position; otherwise hold no position. EXPLICITLY ADDRESSES ALL FOUR ANALYST-RECOMMENDED FIXES from the seven prior ADA 4H optimization collapses: (a) DAILY timeframe (~800-1000 bars over 2-3 years of HL data, vs ~13,774 ADA 4H bars — orders of magnitude smaller Optuna search space, which the analyst's META observation identified as the root cause of the 7-for-7 walk-forward failure pattern); (b) DIFFERENT INSTRUMENT — SOL has higher momentum-continuation than ADA per cross-sectional studies (Hong & Stein 1999, applied to crypto by Liu-Tsyvinski 2021); (c) longer-timeframe TSMOM closer to the original Moskowitz-Ooi-Pedersen daily/monthly formulation; (d) DIFFERENT VENUE — Hyperliquid has different participant mix than Binance (more native crypto whales, less retail), reducing the over-arbitraged risk that killed the cross-venue BTC→ADA lead-lag. ALSO hits the biggest data-available quota gap: HYPERLIQUID at 5.8% vs ≥20% target (14.2% deficit). ZERO conceptual overlap with anything in the 670-strategy portfolio — no SOL.HYPERLIQUID strategy exists today, no daily-TSMOM strategy exists, and SOL on HL has not been touched in this session. Avoids EVERY recently-identified failure pattern: (1) NOT ADA 4H; (2) NOT a microstructure / tick-data strategy; (3) NOT a bar-derived microstructure proxy; (4) NOT a single-trade-dependency setup (TSMOM at daily timeframe has 30-80 trades/year, no lottery dynamics); (5) NOT a regime-fragile breakout (TSMOM captures persistent trends, not channel breaks).

Iteration 2 fix for the Layer-2 'frozen signal (value=-1.0)' failure in steady_downtrend. Root cause: when the TSMOM long condition failed, calculate_signal returned a constant -1.0, so in a persistent downtrend (long_ok always False) every signal was identical. Fix: the not-long branch now returns a continuous, non-positive, bar-varying value -1.0 + tanh(min(trailing_ret, MA-distance)). Because at least one of trailing_ret or MA-distance is <= 0 whenever long_ok is False, this value is always <= -1 (<= 0), so the exact entry/exit contract is preserved: signal > 0 iff the canonical long condition holds. should_enter (signal>0), should_exit (signal<=0), and position_size (signal<=0) are unchanged, so Layer-3 trade behavior is byte-identical to the version that passed Layer 1. The long branch (1 + trailing_ret), warmup guards, stop, and leverage-consuming sizing are untouched, and no imports changed (numpy only), preserving the Layer-1 pass. Venue/leverage unchanged: HYPERLIQUID perp at leverage 2.0 with position_size genuinely consuming leverage.

Suspected backtest/data artifact: the 2024-04-17 daily return of +459.1% is impossible for the stated ~1.2x long position (avg_position_pct 123%, leverage 2.0) — it would require SOL to rise ~380% in a single day, which did not occur. This is almost certainly a bad SOLUSD.HYPERLIQUID daily bar (spurious print) marked-to-market on the open long (symptom stated; the bar was not independently verified — recommend the data team check the 2024-04-17 HL SOL daily OHLC before any retry). That one day IS the entire edge: largest_win $426,648, monthly 2024-04 +438%, annual 2024 +247%, while every other year loses (2023 -7%, 2025 -42%, 2026 -7%) with a 14.7% win rate. The artifact signature is unmistakable — skew 9.67, kurtosis 95.0, annualized_volatility 730%, tail_ratio 0.085, sharpe_ci_low -7.88 — so the headline Sharpe 1.18 / Sortino 17.7 are outlier-carried, not a real edge. The hypothesis explicitly claimed to avoid single-trade-dependency, but this is the purest single-trade-dependent result possible. Separately, HL's ~3-year history (996 bars, 34 trades) is too short to validate a low-frequency daily-TSMOM edge, matching prior HL daily-trend collapses on walk-forward OOS. Not worth 2 hours of optimization (the optimizer would simply fit parameters to one freak bar). NOTE: sizing is clean (123% ~ intended 120%, metrics_reliable=true) — this is a data artifact, not the BINANCE over-exposure defect; HL sizes correctly.

Implementation

Long-only canonical daily TSMOM (Moskowitz-Ooi-Pedersen 2012) on SOLUSD.HYPERLIQUID, pure OHLCV. Holds long while the trailing 90-day return is positive AND price is above its 100-day SMA; otherwise flat. Exits on signal flip (trend lost) or a wide 20% catastrophe stop. calculate_signal() returns 1 + trailing_return when the long condition holds and a continuous non-positive score (-1 + tanh(min(trailing_return, MA-distance))) otherwise, so the signal varies every bar. Leverage 2.0 (~1.2x equity notional) on Hyperliquid.

Verification Results

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['long_ma', 'min_notional', 'position_pct', 'stop_loss_pct', 'lookback_period'] Check that __init__ sets all attributes from self.parameters.get(). - steady_downtrend: Frozen signal: all 396 signals are identical (value=-1.0). calculate_signal() likely has a bug — the signal never varies.

Backtest Review

Sharpe: 1.18
Total return: 77.41%
Max drawdown: 78.85%
Trades: 34
Win rate: 14.7%
Profit factor: 1.18

Sizing is clean: avg_position_pct 123% ~ intended 120%, 0 dropped entries, metrics_reliable — HL sizes correctly (no over-exposure defect)

Mechanism is implemented as specified (canonical long-only daily TSMOM); single-instrument/venue, no staleness risk

SUSPECTED DATA ARTIFACT: 2024-04-17 daily return +459.1% is impossible for a ~1.2x long (would need SOL +380% in a day) — almost certainly a bad HL SOL daily bar

Entire result is that one day: largest_win $426,648, 2024-04 +438%, 2024 +247%; all other years lose (2023 -7%, 2025 -42%, 2026 -7%)

Artifact signatures: skew 9.67, kurtosis 95, annualized_vol 730%, tail_ratio 0.085, win_rate 14.7%, sharpe_ci_low -7.88

max_drawdown 78.8%, max_consecutive_losses 11

Only ~3 years of HL history (996 bars, 34 trades) — too short to validate a daily-trend edge; prior HL daily-trend strategies collapsed on walk-forward OOS

34 trades / ~3yr HL

adequate sample

+459% day at 1.2x lev

physically possible

Outcome Summary

A spectacular headline Sharpe driven by one impossible bar is a data-integrity red flag, not an edge — artifact signatures (extreme skew/kurtosis, 730% vol, sub-zero Sharpe CI, 14.7% win rate) should be checked first, and a strategy explicitly designed to avoid single-trade dependency can still become the purest single-trade-dependent result via a spurious print; also, ~3 years of venue history is too thin to validate daily trend-following.

The backtest-review analyst issued an 'abandon' verdict: the result is a suspected Hyperliquid data artifact (a +459% day physically impossible at the stated leverage) that single-handedly creates the edge, making the headline Sharpe outlier-carried rather than real; separately, Hyperliquid's ~3-year history (996 bars, 34 trades) is too short to validate a low-frequency daily-TSMOM edge, matching prior HL daily-trend OOS collapses.

A single-instrument, long-only canonical daily time-series momentum (TSMOM) strategy on SOLUSD.HYPERLIQUID (2.0x leverage, ~1.2x notional) — holding long while the trailing 90-day return was positive and price was above its 100-day SMA, otherwise flat — deliberately chosen to fix prior ADA 4H walk-forward collapses via a daily timeframe, different instrument, and a less-crowded venue while filling the Hyperliquid quota gap.

Headline metrics looked strong but distorted — 34 trades, +77.4% total return, Sharpe 1.18, Sortino 17.7 — yet with a 14.7% win rate, 78.8% max drawdown, sharpe_ci_low -7.88, skew 9.67, kurtosis 95 and 730% annualized volatility; a single suspected-bad bar (2024-04-17, an impossible +459.1% daily return for a 1.2x long) produced the $426,648 largest win and carried the entire result, while every other year lost (2023 -7%, 2025 -42%, 2026 -7%).

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