EthHyperliquidTrendPullbackContinuationLS

ETH Hyperliquid Trend-Pullback Continuation, Long-Short (Single-Instrument ETHUSD.HYPERLIQUID Perp, 4H Entry + 1D Trend Regime Gate, BUY Shallow Pullbacks in Uptrends / SHORT Shallow Rallies in Downtrends, ATR-Trailing, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV trend-CONTINUATION strategy on ETHUSD.HYPERLIQUID perpetual futures. It is deliberately differentiated from the two momentum strategies already in this factory's pipeline (BTC dual-timeframe momentum confluence, and my SOL Hyperliquid Donchian BREAKOUT) by using the OPPOSITE entry trigger within the same winning family: instead of entering on a breakout to new highs/lows, it enters on a shallow COUNTER-move (pullback/rally) BACK toward the trend after a temporary overshoot — i.e. 'buy the dip in an uptrend, sell the rip in a downtrend.' The higher-timeframe daily trend defines the ONLY allowed direction (long-only in daily uptrends, short-only in daily downtrends), so it never fades a trend the way the dead VWAP-reversion / weekend-overshoot / wick-fade strategies did — those all had structurally inverted payoffs; this one keeps the favorable geometry (enter at a better price WITHIN the trend, ride the continuation leg, trail the winner). It targets the two most under-represented portfolio buckets simultaneously: HYPERLIQUID venue (5.8% vs 20% quota) and long_short direction (13.7% vs target). Pure OHLCV, no supplementary feed (best-surviving tag family at 4%). Uses 4H entry bars (~2.3 years usable HL history) + 1D regime bars (full history) — both above the HL sub-hourly history wall, so it will NOT be abandoned for insufficient data. It structurally avoids every recent death mode: no options-expiry cadence, no cross-venue spot-lookup timing bug, no funding fee-floor problem, no liquidation-feed 90-day wall.

Implements the hypothesis directly: 4H entry + 1D regime gate, buy-the-dip in uptrends / sell-the-rip in downtrends, with-trend only, ATR trailing exit, long-short on Hyperliquid, pure OHLCV, three tunable parameters. Four points worth flagging. (1) I validated data and the FULL compound gate on real catalog data before submitting, since over-selective conjunctions are the top verification-loop killer: ETHUSD.HYPERLIQUID has 4-HOUR bars from 2024-01-22 (5,514 bars) plus full daily history, the shipped signal spans roughly -2.4 to +2.8 at the 5th/95th percentile (so a 1.0-ATR threshold sits mid-distribution, not in the tail), and the complete gate — regime AND the shallow-pullback band — is satisfied on 186 long and 254 short bars inside the 365-day Layer-3 sandbox window. Trade count is not a risk here; if anything the risk is over-trading, which is why entries only fire when flat and a 12-hour cooldown follows every exit. (2) The signal is one continuous variable in natural ATR units — displacement from the prior swing range's midpoint — returned every bar and thresholded in matching units in should_enter, so there is no gated constant-default pattern and no clipping. (3) ETHUSD.HYPERLIQUID appears in both extra_instruments and extra_bar_types deliberately: InstrumentResolver.resolve_extras iterates extra_instruments and pairs bar types positionally, so an empty instrument list would silently drop the daily feed, pin the regime at 0 and yield zero trades. Because the daily leg is a slower timeframe it is excluded from the base class's same-timeframe alignment barrier and cannot deadlock the 4H signal; daily bars are deduped by timestamp because the catalog holds repeats for some timestamps. If the daily feed is ever absent the strategy trades nothing rather than degenerating into an unfiltered dip-buyer — no silent mechanism swap. (4) On the hypothesis's core worry, that this resembles the dead fade family: it is structurally not a fade, because the regime gate makes every long a long inside a daily uptrend, but I would flag one honest risk for the analyst — buying dips inside a trend inevitably has a lower average entry-to-stop distance than a breakout, so the 3-ATR trail and the shallow-band ceiling are doing the work of distinguishing 'pullback' from 'trend break', and the ceiling (locked at 3x the entry threshold) is the parameter I would scrutinise first if the payoff geometry comes out unfavourable. Risk controls follow the compounding-blowup lesson: 1% of equity risked per trade, gross notional hard-capped at 1x equity, no leverage set or referenced, and the cooldown is anchored to bar timestamps rather than a bar counter so it is identical across restarts, backtests and paper replay.

failed multiple-testing robustness: PBO=0.7302 (>0.5, HARD fail — parameter selection more likely than not overfit) with deflated_sharpe=0.0313 and is_significant=false, over 225 trials. The optimized full-sample Sharpe (1.18) sits below both the 1.5 promotion floor and the expected-max-by-luck bar (2.118), and optimization actively DEGRADED the default config (Sharpe 1.42→1.18, DD 5.1%→9.0%, avg_trade 1.78%→1.09%, turnover 2.3x) by selecting atr_trail_mult 1.62 outside the clean sensitivity region. The mechanism is a legitimate two-sided trend-pullback with edge above the fee floor and a clean single holdout, but on only ~2.5yr of HL data its returns are carried by a handful of outlier days (kurtosis 13.5) and the whole single-instrument OHLCV HL trend/momentum family has zero survivors (SOL Donchian, SOL momentum, AVAX Kaufman all abandoned overfit). The only tuning lever — restoring atr_trail_mult toward the default ~3.0 region — yields Sharpe ~1.42 (still below the 1.5 floor) and cannot lower a surface-level PBO of 0.73, so there is no articulable path to a robust, promotable edge from the second optimization pass. Not revise_hypothesis: the failure is statistical overfit on thin data, not a salvageable-mechanism-on-wrong-target premise problem.

Long-short 4H trend-pullback continuation on the ETHUSD Hyperliquid perpetual, gated by the 1-DAY trend regime. The daily close vs its 50-day SMA sets the only permitted direction; within it, the strategy buys shallow dips (4H close 1.0–3.0 ATRs below the midpoint of the prior 20-bar swing range) in daily uptrends and shorts shallow rallies of the same depth in daily downtrends, staying flat whenever the regime is unclear or the counter-move is too deep to be a pullback. Exits use a 3x-ATR chandelier trail from the extreme reached since entry, plus an immediate exit if the daily regime flips against the position, with a 12-hour calendar-anchored cooldown after each exit to prevent re-entry churn. Sizing risks 1% of equity across the initial stop with gross notional capped at 1x equity; leverage 1.0. Exactly three tunable parameters (pullback_atr, regime_sma_window, atr_trail_mult).

Require OOS/walk-forward Sharpe clearly above zero and stable across windows before promoting; don't treat one strong sandbox year as sufficient given the class prior.

Pure-OHLCV class prior (L7) is unfavorable, and 'buy the shallow dip in an uptrend' is a cousin of the abandoned 'RSI-oversold-with-trend-filter' pullback class. However, this is a genuine WITH-TREND continuation, not a fade: the 1-DAY regime gate opens longs only in daily uptrends and shorts only in daily downtrends, and realized geometry is favorable (avg_win $1128 vs avg_loss $608, positive skew), so it lacks the inverted-payoff signature that killed the VWAP/overshoot/wick fade family. Positive, above-floor sandbox, materially unlike the negative Donchian-breakout sibling.

Judge on the full HL history (~100+ trades) plus walk-forward; confirm the favorable geometry and positive expectancy persist across ETH up- and down-trend regimes.

Single-year sample. Strong result (total_return +14.0%, Sharpe 1.14, PF 1.85, avg_trade_return_pct +1.29%, probabilistic_sharpe 0.92) but over 54 trades / 365 days, and the Sharpe CI lower bound (-0.51) dips below zero. A trend-continuation edge needs multiple trend cycles; the full ~2.3yr history + walk-forward must confirm it is not a single-regime artifact.

Check sensitivity around the pullback band structure; fragility to the 3x ceiling would indicate overfit work rather than a robust effect.

Developer's own honest flag, seconded: dip-buying inside a trend has a smaller entry-to-stop distance than a breakout, so the shallow-band ceiling (MAX_PULLBACK_MULT, locked at 3x the entry threshold) and the 3-ATR trail carry the payoff geometry. If OOS geometry degrades this ceiling is the first thing to scrutinize — but it is a LOCKED constant (good for overfit-resistance; a poor choice needs a code change, not re-optimization).

133 trades over ~2.5 years — adequate sample for meaningful parameter sensitivity, not noise-fit

avg_trade_return_pct 1.78% is far above the HYPERLIQUID 0.15% fee floor — real edge net of costs

Trades implement the hypothesis: genuinely two-sided (52 long / 81 short), both legs profitable (long WR 0.50, short WR 0.46) — a regime-gated LS system, not a closet long or an unfiltered fade

Healthy risk-adjusted profile: Sharpe 1.42 (CI [0.44, 2.40] excludes 0), PF 2.16, max_dd 5.1%, calmar 12.6, positive alpha with near-zero beta (0.008)

Clean execution: entry_diag shows 133/133 submitted, 0 dropped for size/notional/cash

Return distribution is fat-tailed (kurtosis 13.5, skew 1.78); a handful of large up-days (e.g. 2025-05-16 +10%, 2025-07-23 +11.7%) contribute heavily — robustness across walk-forward OOS windows must be confirmed, not assumed

max_drawdown CI is wide ([4.78, 15.35]) — downside is less certain than the 5.1% point estimate suggests

benchmark_meaningful=false (LS market-neutral), so alpha/IR vs the basket are not the right yardstick — judge on absolute risk-adjusted metrics, which are solid

Legitimate two-sided trend-pullback mechanism that implements its hypothesis (52-59 long / 81-101 short), not a fee-edge or artifact death; avg_trade_return_pct 1.09-1.78% comfortably above the HYPERLIQUID 0.15% fee floor

Low/negative beta (0.008 raw / -0.027 optimized) with positive alpha (0.13-0.15) and consistent WF-OOS sharpes [2.32, 2.17, 1.26]; holdout passed (ratio 1.12); sensitivity grid clean with no cliffs

HARD FAIL — PBO 0.7302 > 0.5: the parameter selection is more likely than not overfit; promote is impossible and unwaivable

Deflated Sharpe 0.0313 (needs 0.95), is_significant=false, Sharpe CI low 0.1175 — the selected config over 225 trials is best-of-N noise

Optimized full-sample Sharpe 1.18 is below BOTH the 1.5 floor AND the 2.12 expected-max-by-luck bar; optimization degraded the default (Sharpe 1.42→1.18, DD 5.1%→9.0%, avg_trade 1.78%→1.09%, turnover 18.5→41.8) by choosing atr_trail_mult 1.62 outside the clean sensitivity region

Only ~2.5yr of HL data (919 days); returns carried by ~4 outlier days (kurtosis 13.5, skew 1.78) — the overfit signature

Every sibling in this single-instrument OHLCV HL trend/momentum family (SOL Donchian, SOL momentum, AVAX Kaufman) was abandoned for overfit — zero survivors via this path

This strategy proposed a differentiated, two-sided trend-pullback edge on ETH Hyperliquid perps — buying dips in uptrends and shorting rips in downtrends, gated by a daily regime filter — and its default backtest delivered a legitimately attractive profile (Sharpe 1.42, PF 2.16, 5.1% max DD, 1.78% avg trade return across 133 balanced trades). On that strength the analyst approved it for optimization at the backtest-review gate. But the 3-phase optimization exposed overfit: PBO 0.7302, deflated Sharpe 0.0313, and an optimized full-sample Sharpe of just 1.18 that sat below both the promotion floor and the expected-max-by-luck bar, with tuning actively degrading the default. With only ~2.5 years of Hyperliquid history, returns carried by a few outlier days, and every sibling in this single-instrument OHLCV trend family already abandoned for the same reason, the analyst found no path to a robust promotable edge and abandoned it after one iteration.

A mechanically sound edge with a clean initial backtest can still be unpromotable when the parameter search shows high PBO on thin data — confirm robustness before optimizing, and treat the entire single-instrument OHLCV Hyperliquid trend/momentum family (zero survivors) as a structurally overfit-prone target.

The analyst abandoned it at the post-optimization ANALYZING stage on a hard multiple-testing failure: PBO 0.7302 (>0.5), deflated Sharpe 0.0313, is_significant=false over 225 trials, with the optimized Sharpe 1.18 below both the 1.5 floor and the 2.12 expected-max-by-luck bar. The single holdout passed (ratio 1.12) but could not offset a surface-level overfit signature on only ~2.5 years of data whose returns leaned on a handful of outlier days.

A long-short, single-instrument trend-continuation strategy on ETHUSD.HYPERLIQUID perps that bought shallow 4H pullbacks in daily uptrends and shorted shallow rallies in daily downtrends, gated by a 1-day regime SMA and exited via an ATR chandelier trail.

The default backtest was strong — Sharpe 1.42, total return ~59.9%, profit factor 2.16, max drawdown 5.1% over 133 genuinely two-sided trades (52 long / 81 short) with avg_trade_return_pct 1.78%, well above the 0.15% Hyperliquid fee floor. Optimization degraded it: Sharpe fell to 1.18, drawdown rose to 9.0%, avg_trade_return_pct dropped to 1.09%, and turnover more than doubled.