Strategies

DOGE Hyperliquid Single-Instrument Slow Time-Series Momentum, Long-Short (DOGEUSDT.HYPERLIQUID Perp, 4H Bars — Directional by the Sign of a 5-Day Return with a Volatility Deadband, ATR-Trailing Exit, Flat Inside the Deadband, 3-Parameter)

Outcome: Abandoned

DogeHyperliquidSlowTsmomDeadbandLS

Outcome Summary

A slow single-name OHLCV momentum edge clears fees but reverses OOS and can't survive deflation; reframing onto another instrument just re-overfits.

Strategy report

Hypothesis

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV slow time-series-momentum (TSMOM) trend follower on DOGEUSDT.HYPERLIQUID perpetual, 4H bars. Deliberately chosen to fill the three furthest-from-target buckets that are also DATA-BACKED: (1) VENUE — HYPERLIQUID is 7.2% of the corpus vs a ≥20% target and is the most under-represented venue for which the factory actually collects data (Bybit and Polymarket have no collector in the data_infrastructure list / on_demand.py, so they die data_unavailable per L76); (2) DIRECTION — long_short is only 14.5% vs the ≤55% long-only cap, so a symmetric directional book pushes the mix the right way; (3) SYMBOL — DOGE fights the 62%-BTC recent concentration. This is a SINGLE-INSTRUMENT expression, explicitly the sanctioned fallback in L77/L88 which condemn only vol-targeted MULTI-perp baskets and Hyperliquid cross-sectional books (which crash at backtest_failed); one leg resolves fine. It is NOT a volume-confirmed-momentum clone (L91) and NOT a regime/efficiency/volatility gate bolted onto a momentum edge (L92) — there is a single signal (the sign of a slow return) and a deadband to suppress whipsaw noise, not a second 'confirmation' edge stacked on a first. Mechanism: compute the trailing 30-bar (5-day) log return r. Go/stay LONG when r > +d (deadband), SHORT when r < -d, FLAT when |r| <= d. Exit a position on an ATR-based trailing stop OR on a signal flip. Three parameters only: lookback (30 bars), deadband d (expressed in units of trailing ATR, ~0.5-1.0 sigma so it scales with volatility), and the ATR trailing multiple. Low parameter count directly attacks the overfit failure mode (381 deaths). Hyperliquid 4H has ~2.3 years of usable history (~5000 bars) — well inside the HL candle-depth limit for 4H, so this is NOT an HL sub-hourly history trap; a 5-day-lookback trend flips roughly every 4-12 days, yielding ~120-200 round trips over the window, enough for statistical power without churning sub-fee trades.

Direct implementation of the hypothesis: one signal (the sign/strength of a slow 5-day return), one volatility deadband, one ATR trailing exit, three tuned parameters plus the ATR period and risk fraction. No regime, volume or efficiency confirmation layer is stacked on the momentum edge. The deadband is expressed in ATR-sigma units rather than raw percent so it is scale-invariant across DOGE's volatility regimes (per the lesson against fixed absolute thresholds), and calculate_signal returns the raw continuous score every bar rather than a gated constant. Symbol correction: the hypothesis writes DOGEUSDT.HYPERLIQUID but Hyperliquid perps are catalogued as DOGEUSD.HYPERLIQUID (2024-01-22 to present, ~5,800 4H bars); the coded instrument is the same asset, venue and contract, just the catalog's real id. Venue is HYPERLIQUID as specified — required because the book is long+short and cannot live on a CASH spot venue. One deliberate reading of an ambiguity: the title says 'flat inside the deadband' while the mechanism paragraph says exit on trailing stop or signal flip; I gate ENTRY on the deadband but hold until a zero-cross or the ATR stop, because exiting whenever the score re-enters the deadband roughly doubles round trips and Hyperliquid's ~0.09% round-trip taker cost is the main killer of 4H trend books. Warmup is 49 bars out of ~5,800, and a 5-day-return sign flip on 4H DOGE fires every few days, so trade count over the sandbox window is comfortably above one and in the hypothesis's ~120-200 round-trip range.

Overfit with negative out-of-sample generalization. Walk-forward IS 1.98 -> OOS -0.418 (is_overfitted=true, 2 of 3 windows negative), and the last-20% holdout Sharpe is -1.24 on 22 trades — the edge reverses on unseen data, it does not merely decay. This is confirmed quantitatively by DSR=0.0015 / is_significant=false / fails programme FDR (4 of 268 kept): the optimized Sharpe is a best-of-N artifact of 225 trials, not a real edge (expected-max 2.27 > selected). Sensitivity shows 4 cliff parameters, so there is no robust region to iterate toward. This is a single-name OHLCV slow-TSMOM clone that misses the family survivor tier (PF>=2.0, Sharpe>=1.5, DD<20%) by a wide margin (PF 1.32, Sharpe 0.46) and joins several DOGE/HL trend-following siblings that died overfit/risk_reject/fee_edge. Not revise_hypothesis: reframing the same momentum mechanism onto another instrument re-overfits rather than fixing a premise flaw.

Implementation

Long-short slow time-series-momentum trend follower on the DOGE Hyperliquid perpetual, 4H bars, pure OHLCV. Each bar it computes the trailing 30-bar (5-day) log return and normalises it by the ATR-implied move over the same horizon (ATR%/close * sqrt(lookback)), producing a continuous, scale-free momentum score of roughly +-3. It goes LONG when the score exceeds +deadband, SHORT when it falls below -deadband, and stays flat inside the deadband so low-conviction chop does not generate fee-burning round trips. An open position is exited when the momentum score crosses zero (trend flip) or when price retraces atr_trail_mult ATRs from the favourable extreme reached since entry (ATR trailing stop). Sizing risks a fixed 2% of equity across the ATR stop distance, hard-capped at 1x equity notional so drawdowns cannot compound into larger positions. leverage=1.0 and sizing never exceeds equity notional, so no margin amplification is claimed or used. ATR and the return window are maintained incrementally (Wilder EMA + a bounded deque), so per-bar cost is O(1).

Verification Results

None required — the chosen behavior is consistent with the 'exit on signal flip or ATR stop' mechanism and is fee-sensible. Flagging only so the analyst is aware exits use a zero-cross, not a deadband re-entry.

The hypothesis title says 'flat inside the deadband', which could be read as exiting whenever the momentum score re-enters (-d, +d). The code instead gates ENTRY on the deadband but only EXITS on a full zero-cross (signal<0 for longs) or the ATR trailing stop. This is a defensible interpretation (explicitly documented in the config rationale, and the mechanism paragraph does say 'exit on trailing stop OR signal flip'), and it roughly halves fee-burning round trips, but it is a deliberate resolution of an ambiguity rather than a literal read of the title.

No change needed; confirmed intentional and correctly scoped.

min_bars_required() overrides a base-template method (static-analysis base_shadow warning). Here the override is deliberate and correct: it returns lookback+atr_period+5 = 49 bars, matching the strategy's actual warmup needs against ~5,800 available 4H bars.

Trend-follower that loses in the stressed vol tercile (return -24.98%, Sharpe -0.55, DD 36.8%) while winning in calm/normal (Sharpe ~0.80). Sandbox headline is positive (total_return 32.4%, PF 1.24, avg_trade_return_pct 0.56% > HL ~0.10% floor, 137 trades) but Sharpe CI [-0.97, 1.48] straddles zero and 4H DOGE TSMOM is whipsaw-prone. Analyst should confirm the edge survives full history and is not concentrated in a couple of trend windows (largest_win 23.5k vs avg_win 3.3k).

Backtest Review

Sharpe: 0.30
Total return: 32.40%
Max drawdown: 26.89%
Trades: 137
Win rate: 35.8%
Profit factor: 1.14

Mechanism faithfully implemented: 137 signaled = 137 submitted (no size/notional/cash drops), balanced 70 long / 67 short symmetric directional book, avg hold ~3d consistent with a 5-day-lookback trend flip — the trades ARE the stated slow-TSMOM-with-deadband + ATR-trail, not a mismatched signal.

Per-trade edge clears costs decisively: avg_trade_return_pct 0.56% vs the HYPERLIQUID ~0.15% floor; commission only 2.86% of gross.

Reliable, adequately powered sample: 5756 4H bars / 959 days, 137 round trips — enough for meaningful walk-forward/parameter sensitivity.

Positive expectancy: PF 1.14, total_return +32.4%, positive in 2024 (+10.2%) and 2025 (+8.7%); only a 3-parameter surface (low overfit surface area).

Low risk-adjusted return: Sharpe 0.30 with bootstrap CI [-0.86, 1.50] straddling zero — statistical significance is unresolved and the deflation stage will be the real test.

Regime fragility: loses in the stressed vol tercile (-24.6%, Sharpe -0.55, DD 36.5%) while profit sits in calm/normal; flat-to-negative through 2026 (rolling Sharpe deeply negative mid-2026).

Profit concentration typical of TSMOM but pronounced: largest_win 22.9k vs avg_win 3.2k; a single +22.8% day (2026-08-22) carries much of 2026.

Capacity/impact drag: impact_cost_pct 37.98% consumes over a third of gross PnL and capacity_usd is only ~$693k — the nominal edge shrinks at scale.

Analysis

Mechanism faithfully implemented and cleanly coded: 153 entries signaled = 153 submitted, balanced long/short book (69 long / 84 short), avg hold ~3.5d consistent with a slow-TSMOM flip. Trades ARE the stated slow-TSMOM-with-deadband + ATR-trail, no mismatch.

Per-trade edge clears the fee floor comfortably (avg_trade_return_pct 1.23% vs HYPERLIQUID ~0.15% floor); impact 24.8% of gross is within tolerance.

Optimized full-period drawdown 19.8% within the pre-registered 25% cap.

Walk-forward is overfit: avg IS Sharpe 1.98 collapses to avg OOS -0.418, with 2 of 3 OOS windows negative (-1.87, -0.57). is_overfitted=true (HARD gate).

Holdout is NEGATIVE: -1.24 Sharpe on 22 trades — the edge does not merely fade, it reverses on truly unseen data (HARD gate).

Deflated Sharpe 0.0015 (PSR 0.74, PBO 0.41), is_significant=false, and fails programme-level FDR (keeps 4 of 268). The selected Sharpe is indistinguishable from best-of-N noise: expected-max Sharpe 2.27 exceeds the point estimate.

Sensitivity flags 4 cliff parameters (deadband, lookback, atr_period, atr_trail_mult) — no stable parameter region to tune toward.

Single-name OHLCV TSMOM clone: this family's survivor bar (PF>=2.0, Sharpe>=1.5, DD<20%) is missed by a wide margin (PF 1.32, Sharpe 0.46). Multiple sibling DOGE/HL trend followers in the corpus died overfit/risk_reject/fee_edge.

Benjamini-Hochberg at q=0.10 over 268 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

Analyst abandon post-optimization on the hard overfit + negative-holdout gates, DSR 0.0015, FDR failure, 4 cliff params.

A single-instrument, long-short slow time-series-momentum trend follower on DOGEUSD.HYPERLIQUID 4H perps...

Initial +32.4% / Sharpe 0.30 over 137 trades; optimized to +75.9% / Sharpe 0.46, but WF IS 1.98 → OOS -0.42 and holdout Sharpe -1.24.

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