Strategies

Hyperliquid ETH Daily Time-Series Momentum, LONG/SHORT (Single-Instrument HYPERLIQUID Perp, Pure OHLCV, Flip-Via-Flat, Fixed-Fractional Sizing — Direct ETH Analog of the PROMOTED HL BTC Daily TS-Momentum LS, Applying the ONE Validated HL Mechanism to the Other Clean Major on the Strong-Walk-Forward Daily Cadence, Pre-Registered Strong-WF Acceptance, Fills Long-Short + HL Venue, Tail-Safe, Low-Parameter)

Outcome: Abandoned

EthHyperliquidDailyTSMomentumLongShort

Outcome Summary

Built as a direct ETH copy of the promoted Hyperliquid BTC daily TS-momentum book — same construction so any difference would be the asset, not the code — this strategy aimed to extract ETH's weaker-but-present daily trend. Its raw numbers looked sibling-like (Sharpe 0.59, PF 1.55, +166% total), enough to clear the pre-optimization gate, but the edge was carried almost entirely by 2025 (+213%) and a handful of fat-tail days, with negative returns in every other year. Optimization confirmed the weakness rather than fixing it: deflated Sharpe 0.0 far below the 2.608 luck bar, PBO 0.76, an overfit walk-forward averaging -1.35 OOS, and a negative -0.45 holdout. The analyst abandoned it after one iteration, noting that the hypothesis had pre-registered exactly this outcome — ETH's daily trend is too weak to extract — making the final stage abandoned the predicted asset-selectivity result.

Strategy report

Hypothesis

A LONG-SHORT, SINGLE-INSTRUMENT, DAILY time-series (absolute) momentum strategy on ETHUSD.HYPERLIQUID — the direct instrument-analog of the PROMOTED HL BTC daily TS-momentum LS (Sharpe 0.68, walk-forward OOS 2.52, PF 1.59). It is grounded in the now-resolved HL cadence map: the ONLY validatable HL cadences are DAILY (~3yr, for edges that survive the thin 15-day holdout via a STRONG walk-forward) and 4H (~1.3yr, for turnover ideas); 1H (~7mo) and 15-min (~2mo) are data-walled (stop probing). The promoted BTC book proved that TS-momentum is the mechanism that DOES validate on HL daily — it survived the thin holdout precisely because its walk-forward OOS was strong (2.52). This hypothesis tests whether that single validated mechanism PORTS to the other confirmed-clean major (HL ETH sizes 64.5%; SOL is data-poisoned; BTC/ETH are the only clean instruments). It is deliberately the strong-WF daily route, NOT a sparsity-prone construction: TS-momentum on daily produced enough trades + strong WF on BTC, so the same on ETH is the legitimate HL-daily play (unlike the daily breakout/reversal that died on holdout sparsity). It is DISTINCT from: the failed HL ETH daily DONCHIAN BREAKOUT (a channel breakout with 1.05x whipsaw geometry — TS-momentum is a smoother sign-of-trailing-return signal, a different mechanism); the failed HL ETH daily REVERSAL (momentum is the opposite mechanism); and the pending HL 4H reversion books (different cadence + mechanism). It fills the dominant direction gap (long_only 85.9% vs <=55%) and under-weight HL venue (7.5% vs >=20%). HONEST CAVEAT / asset-selectivity: ETH trends weaker than BTC (its breakout had no edge), so this is a genuine test of whether TS-momentum extracts ETH's weaker-but-present daily trend; the construction is identical to the promoted BTC book so any difference is the asset, not the code. Bug-avoiding: pure OHLCV, ONE fixed-fractional size per entry, NO dynamic resizing, leverage 1.0, dual flat guard, flip-via-flat (clean LS path — HL BTC LS sized 97%, ETH expected similar). PRE-REGISTERED ACCEPTANCE: (1) STRONG-WF — like the BTC sibling, the edge must survive HL daily's thin holdout via a non-overfit walk-forward (IS/OOS ratio < 3, OOS Sharpe positive); if ETH momentum is too weak (breakeven/weak WF OOS), ABANDON — do NOT force/iterate (ETH momentum weakness is structural, not tunable, per the breakout finding). (2) SIZING — avg_position_pct ~95-100%, no impossible single-day returns; else escalate. Daily 4H depth not relevant — HL ETH daily has ~3yr. Low parameter count (momentum lookback, optional EMA confirm, ATR stop).

Direct instrument-analog of the promoted HL BTC daily TS-momentum LS, with identical construction so any difference is the asset (ETH's weaker daily trend), not the code. Iteration-2 fix for the Layer-3 timeout: the strategy is explicitly a DAILY-cadence book, so the bar_type is pinned to 1-DAY (ETHUSD.HYPERLIQUID-1-DAY-LAST-EXTERNAL) — running the daily strategy on its intended ~1,095-bar daily series instead of any finer interval is the dominant fix for the 300s wall-clock blowup. Additionally, per-bar cost was already bounded by fixed-maxlen deques; it is now reduced to a single numpy conversion per bar (closes and TRs each converted once and slices reused), so the per-bar constant factor is smaller and provably independent of history length. All signal logic, thresholds, sizing, and risk controls that passed Layers 1-2 are unchanged.

Failed every forward and multiple-testing gate, and the hypothesis's own pre-registered abandon trigger fired. DSR=0.0 (vs 0.95 bar) with the optimized Sharpe 0.459 FAR BELOW the expected-max best-of-225 luck bar of 2.608 (is_significant=false, sharpe_ci_low=-0.95 straddling 0). Walk-forward is_overfitted=TRUE: avg IS 0.114 collapses to avg OOS -1.347 with windows [-3.18, -2.38, +1.52] (two deeply negative), and the HOLDOUT FAILED and is NEGATIVE (holdout_sharpe -0.452, ratio 0) — the untouched recent window loses money. PBO=0.7632 (>0.5). The base headline is regime-concentrated outlier-carry — 2025 +213% against 2023 -16.6%/2024 -6.2%/2026 -6.3%, driven by ~3 fat-tail days (kurtosis 137->192, skew ~10) — and optimization made Sharpe WORSE (0.587 -> 0.459). Not iterate (attempt 1 of 2): there is no robust region to tune toward (OOS negative in 2 of 3 windows, holdout negative, only a lookback cliff), the best-of-225 Sharpe sits well below the luck bar, and the hypothesis explicitly pre-registered ABANDON — not iterate — if ETH momentum proved weak (it stated ETH momentum weakness is structural, not tunable, confirmed by the dead ETH breakout sibling). Not revise_hypothesis: the TS-momentum mechanism is already validated/promoted on the HL BTC daily sibling, so ETH is not a sound mechanism stranded on a dead target — it is asset-selectivity (ETH's daily trend is structurally weaker than BTC's), with no other clean major to reframe toward (SOL is data-poisoned, BTC is already promoted). FAILURE PATTERN: the HL-daily TS-momentum mechanism that clears the bar on BTC (Sharpe 0.68, WF-OOS 2.52) does NOT port to ETH — ETH's weaker daily trend yields a regime-concentrated, outlier-carried edge whose best-of-225 selected Sharpe (0.459) falls far below the luck bar (2.61), failing deflated Sharpe (0.0) with an is_overfitted walk-forward (avg OOS -1.35) and a negative holdout (-0.45). A clean, low-parameter, sibling-identical construction does not rescue an asset whose underlying trend is too weak to extract; this is the predicted asset-selectivity outcome, and the pre-registered acceptance correctly routes it to abandon rather than a second wasted sweep.

Implementation

Long/short single-instrument daily time-series (absolute) momentum on ETHUSD.HYPERLIQUID. Computes a vol-normalized, t-stat-like momentum score over a lookback window; goes long when the score is strongly positive and price is above a trend SMA, short when strongly negative and below the SMA. Exits on momentum reversal, trend-SMA cross, a frozen ATR-multiple stop, or a percentage stop. Flips are routed through flat. One fixed-fractional position per entry, leverage 1.0, pure OHLCV.

Verification Results

Verification failed (Layer 3 — sandbox backtest): smoke test exceeded the 300s wall-clock limit. This almost always means per-bar work that scales with history — e.g. rescanning the full funding/supplementary series, or rebuilding a list and calling min()/sorted() inside calculate_signal()/on_bar() on every bar. Precompute sorted timestamp arrays ONCE in __init__ and use bisect, or cache lookups keyed by timestamp, so per-bar cost is O(log n) not O(n).

Backtest Review

Sharpe: 0.59
Total return: 166.53%
Max drawdown: 38.36%
Trades: 64
Win rate: 23.4%
Profit factor: 1.55

Functioning signal with an adequate sample for daily momentum: 64 trades (37 long / 27 short) over ~864 daily bars, zero dropped/size-zero/min-notional entries.

Direct code-identical analog of the PROMOTED HL BTC daily TS-momentum LS; Sharpe 0.586 and PF 1.55 sit in the same range as the promoted sibling (0.68), so the mechanism is validated and the test is legitimate.

The apparent 'impossible single-day returns' are per-trade realized returns booked on ~23-day exit dates, not daily marks — sizing (leverage 1.0, fixed-fractional) is fine, so the pre-registered SIZING escalation does not fire.

Low win-rate + fat-tailed wins (PF 1.55, +166% total) is the expected TS-momentum profile, not a defect.

Edge is almost entirely concentrated in 2025 (+213%) with 2023/2024/2026 all negative — strong single-regime concentration risk.

sharpe_ci_low -0.54 straddles 0 and naive expectancy is negative; the edge is outlier-dependent (kurtosis 137, skew 9.9).

information_ratio -0.38 vs buy-hold; ETH's daily trend is genuinely weaker than BTC's, exactly the asset-selectivity risk the hypothesis pre-registered.

Analysis

Clean, low-parameter, bug-avoiding construction (pure OHLCV, single instrument, fixed-fractional sizing, leverage 1.0, flip-via-flat) — identical to the promoted BTC sibling, so any difference is the asset not the code

Sensitivity surface is mostly flat with only 1 cliff (lookback); ATR/min_notional/stop/exit params show no cliffs

Plausible position sizing (avg_position_pct ~48-63%, no impossible single-day returns), good capacity (~$53-76M)

Failed deflated Sharpe decisively: DSR=0.0, is_significant=false, optimized Sharpe 0.459 FAR below the expected-max best-of-225 luck bar of 2.608 — indistinguishable from selection noise

Walk-forward is_overfitted=TRUE: avg IS 0.114 collapses to avg OOS -1.347 with windows [-3.18, -2.38, +1.52] — two deeply negative, edge rests on a single positive window

Holdout FAILED and is NEGATIVE: holdout_sharpe -0.452 (ratio 0) — the untouched recent window loses money

PBO=0.7632 (>0.5) and sharpe_ci_low=-0.95 (CI straddles 0) on both base and optimized configs

Entire headline return is regime-concentrated outlier-carry: 2025 +213% vs 2023 -16.6%, 2024 -6.2%, 2026 -6.3%, driven by 3 fat-tail days (return_kurtosis 137->192, skew ~10); optimization made Sharpe WORSE (0.587 -> 0.459)

Pre-registered abandon trigger fired: hypothesis committed to ABANDON (not iterate) if WF OOS is weak/breakeven — ETH momentum weakness is structural, not tunable

Outcome Summary

A validated, code-identical mechanism does not automatically port across assets — ETH's structurally weaker daily trend yields only a regime-concentrated, outlier-carried edge that collapses out-of-sample, vindicating a pre-registered abandon-on-weak-walk-forward rule over a second wasted optimization sweep.

The backtest-review gate said 'optimize' but the post-optimization analyst returned an 'abandon' verdict on attempt 1: the strategy failed every forward and multiple-testing gate (DSR 0.0 vs the 2.608 luck bar, negative holdout, overfit walk-forward), its headline rested on a single 2025 regime (+213% against negative 2023/2024/2026), and the hypothesis's own pre-registered abandon trigger fired since ETH's momentum weakness is structural, not tunable; the final stage is abandoned.

A long/short, single-instrument daily time-series (absolute) momentum strategy on Hyperliquid ETH perpetual — a code-identical analog of the promoted HL BTC daily TS-momentum book — testing whether the one validated HL-daily mechanism ports to the other clean major, with a pre-registered commitment to abandon (not iterate) if ETH's weaker trend produced a weak walk-forward.

The initial backtest showed Sharpe 0.59 (CI low -0.54, straddling zero), profit factor 1.55 and 166.5% total return over 64 trades (37 long/27 short) but a 23.4% win rate, negative expectancy, and a fat-tailed, outlier-dependent profile (skew 9.9, kurtosis 137); optimization made it worse (Sharpe 0.46, DSR 0.0, PBO 0.76, is_overfitted true with avg OOS Sharpe -1.35 across windows [-3.18, -2.38, +1.52]) and the holdout failed and was negative (holdout Sharpe -0.45).

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