EthHyperliquidDualTfEmaConfluenceLS

ETH Hyperliquid Dual-Timeframe Momentum Confluence, Long-Short (Single-Instrument ETHUSD.HYPERLIQUID Perp, 4H Primary + 1D Confirm, Flat on Disagreement, ATR-Trailing Exit, 3-Parameter) — the factory's ONE surviving edge, moved onto the most under-represented data-available venue

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV momentum-confluence trend follower on ETHUSD.HYPERLIQUID (perpetual), 4H primary bars + 1D trend confirm. This is a DELIBERATE, evidence-driven port of the ONLY mechanism family that has ever reached paper-stage in this factory — dual-timeframe momentum confluence long-short (BTC Sharpe 1.99, ADA 1.60, ETH volume-confirmed 1.05, all currently in paper_stage) — onto the venue the portfolio is most starved of. Rationale for every choice below is anti-clone and anti-quota-violation. It is single-instrument (NOT a basket — L88 shows HL multi-instrument baskets crash backtest_failed with null metrics; single-leg HL bars resolve fine). It is 3-parameter (fast EMA span, EMA ratio, ATR-stop multiple) to respect that overfit is the #2 killer (376 deaths) and the surviving siblings are all 3-parameter. It is ETH, an asset on which THIS mechanism has already survived deflation on Binance — so this tests venue-robustness of a REAL edge rather than porting a proven recipe onto a marginal alt (LTC/DOT/BNB/LINK ports all died overfit/negative-expectancy, so instrument novelty is a liability here; venue novelty on a proven asset is the safer axis).

This is the deliberate venue port the hypothesis asks for: the dual-timeframe momentum-confluence long-short family is the only mechanism in this factory that has reached paper stage (BTC 1.99, ADA 1.60, ETH volume-confirmed 1.05), and ETH is an asset on which it already survived deflation on Binance -- so the only novel axis here is the venue (Hyperliquid), not the asset or the mechanism. Implementation choices map one-to-one onto the hypothesis: SINGLE INSTRUMENT (no HL basket -- multi-leg HL baskets have crashed with null metrics; a single ETHUSD.HYPERLIQUID leg with a second bar type of the SAME instrument resolves cleanly, and because 1-DAY is a slower timeframe than the 4-HOUR primary it is not a sync leg, so the base class's cross-leg alignment barrier never blocks the signal). THREE PARAMETERS exactly as specified (fast EMA span, EMA ratio, ATR-stop multiple), with every risk/execution constant locked out of self.parameters so it cannot become an inert optimizer dimension that inflates PBO. FLAT ON DISAGREEMENT and ATR-TRAILING EXIT are implemented literally. The daily confirm is derived from the same two spans rather than two new ones, which is the only way to express a genuine 1-DAY confirm without breaking the 3-parameter budget. Warmup is ~36 daily bars (~216 4H bars) out of ~5,700 available 4H bars / ~960 daily bars of Hyperliquid history (2024-01-22 onward), roughly 4% of the window, so the sandbox has ample room to produce trades; entry is a two-condition state gate (not a rare conjunction), and the sibling port on BTCUSD.HYPERLIQUID with the same entry discipline produced 323 trades over this venue's history. Venue is HYPERLIQUID futures/MARGIN because the strategy goes short; leverage stays 1.0 and sizing therefore never references a leverage multiplier it does not use.

Overfit venue port: walk-forward is_overfitted=TRUE (IS 0.74 -> OOS 0.288), OOS Sharpe 0.288 < 0.5, PBO 0.69 > 0.5, and holdout carries only 9 trades. Quantitatively confirmed by failed deflated Sharpe: DSR=0.025, expected-max=1.585 over 225 trials — the selected Sharpe is indistinguishable from best-of-N noise (PSR 0.70, fails programme FDR). The sensitivity grid is uniformly mediocre (0.23-0.53) with no robust region to tune toward, so iteration cannot recover an edge that is not present. The dual-TF momentum-confluence mechanism is proven on Binance BTC/ADA siblings, but this ETH-on-Hyperliquid port specifically failed deflation — matching the abandoned-overfit outcome of the near-identical prior ETH HL / ETH Binance / BTC HL / SOL HL ports in the failure store. Not a salvageable-elsewhere premise (same asset, shorter-history venue), so abandon rather than revise.

Long-short, single-instrument, pure-OHLCV dual-timeframe EMA momentum confluence on the ETHUSD.HYPERLIQUID perpetual. The fast leg is an EMA pair (span fast_ema_span vs round(fast_ema_span*ema_ratio)) on 4-HOUR bars; calculate_signal returns their normalised spread (ema_fast-ema_slow)/ema_slow every bar. The confirm leg applies the IDENTICAL EMA pair to 1-DAY bars of the same instrument (a daily bar is 6x a 4H bar, so the same spans give a 6x slower regime leg for zero extra parameters). Entry: both spreads positive -> LONG, both negative -> SHORT, disagreement -> FLAT. Exit: the daily confirm flipping against the held side, or an ATR trailing stop seeded at 1.5*ATR from entry and ratcheting favourably to best_price -/+ trail_atr_mult*ATR. Sizing risks 1% of equity across the initial 1.5*ATR stop with gross notional capped at 50% of equity; leverage 1.0. Exactly three tunable parameters (fast_ema_span, ema_ratio, trail_atr_mult) -- ATR period, initial-stop multiple, risk %, notional cap, min notional and size precision are locked class constants so the optimizer sees a 3-dimensional surface rather than nine mostly-inert dimensions.

No action strictly required: should_enter() independently self-guards on self._f_n and self._d_n >= slow_ema_span, so entries cannot fire before both legs are warm regardless of whether the base calls this method. Left as-is is safe.

min_bars_required() shadows a framework method (Layer-1 base_shadow warning). If the base does not actually invoke this override, engine-level warmup gating may not apply.

Correctness is clean, but this is a Hyperliquid dual-timeframe momentum port, and every prior HL dual-TF momentum port in the record has died at analysis (0/N: SOL/ETH HL confluence, SOL HL dual-TF vol-gate, SOL HL TSMOM) despite the mechanism surviving on Binance majors. The sandbox Sharpe is 0.62 with CI [-0.53, 1.66] straddling zero and the 'normal' vol tercile is slightly negative (-1.2%), so the venue-robustness thesis is unproven. First thing the analyst should check on full history: whether OOS/deflated Sharpe survives, and whether per-trade edge holds above the 0.045% HL taker cost on a decisive trade count.

Per-trade edge clears the venue cost floor with wide margin: avg_trade_return_pct 0.739% vs the 0.15% Hyperliquid taker floor — this is NOT a fee-edge death.

Decisive, balanced sample: 158 trades (72 long / 86 short) over 958 days; 158 signaled = 158 submitted, so the trades faithfully implement the dual-timeframe confluence long-short mechanism (no direction/timing contradiction).

Positive expectancy and profit_factor 1.227; total_return +41.6% (CAGR 11.4%) with only 10.4% end_unrealized, so the headline is mostly realized. Max drawdown 27.1% is well within review tolerance.

Metrics flagged reliable; benchmark_meaningful=false correctly noted (single-leg long-short vs equal-weight basket), so judge on absolute risk-adjusted metrics.

Sharpe only 0.625 with bootstrap CI [-0.565, 1.765] straddling zero — statistical significance is unproven and the win rate is low (0.32).

The 'normal' vol tercile is slightly negative (Sharpe 0.034, -1.2% return); the edge is carried by a handful of large trend days (2024-03-05 +16%, 2025-05-17 +15%, 2025-07-23 +16%, 2026-08-28 +15%).

Class prior is adverse: every prior Hyperliquid dual-TF momentum port has died at ANALYSIS (overfit/deflation), and this sits far below the only surviving tier of this family (ADA/BTC at Sharpe>=1.5, PF>=2.0). Deflation risk at optimization is high.

Impact cost is 7.76% of gross PnL — non-trivial for the DEX venue; watch capacity/impact after tuning.

Per-trade edge clears the Hyperliquid cost floor: avg_trade_return_pct 0.62% (optimized) vs the 0.15% taker floor — this is NOT a fee-edge death.

Sensitivity is clean: no cliff parameters, only 3 tunable dimensions, balanced long/short sample.

Mechanism family is genuinely proven on siblings (BTC/ADA/ETH-volume dual-TF confluence in paper_stage) — the idea is sound in the abstract.

Walk-forward flags is_overfitted=TRUE: IS avg Sharpe 0.74 collapses to OOS avg 0.288, with OOS windows scattered -2.99 / +0.59 / +3.26 (pure noise, not a consistent edge).

OOS Sharpe 0.288 < 0.5 hard validity floor; PBO 0.69 > 0.5 (more likely than not overfit); holdout only 9 trades < 10 minimum (thin, and the vault is self-admittedly underpowered at ~30 expected trades).

Deflated Sharpe 0.025 and PSR 0.70 — the selected Sharpe is statistically indistinguishable from best-of-N noise over 225 trials (expected-max 1.585); is_significant=false; fails programme-level FDR (candidate_p 0.975, keeps 4 of 270).

Uniformly mediocre sensitivity grid (0.23–0.53 everywhere) — no robust high-Sharpe region to tune toward; best params pinned at the edge of range (fast_ema_span=24).

This exact dual-TF-confluence venue-port family has died repeatedly: ETH HL (overfit, 2 iters), BTC HL (fee_edge), ETH Binance (overfit), SOL HL (risk_reject) — the ports do not add robustness the deflation can't strip.

Benjamini-Hochberg at q=0.10 over 270 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.

EthHyperliquidDualTfEmaConfluenceLS-36614476ab

This strategy took the factory's only mechanism to ever reach paper-stage — long-short dual-timeframe EMA momentum confluence — and ported it onto the under-represented Hyperliquid venue on ETH, an asset where the mechanism had already survived deflation on Binance. The initial backtest cleared the pre-optimization gate (Sharpe 0.625, PF 1.23, +41.6%, per-trade edge well above the fee floor) and earned an 'optimize' verdict, but optimization exposed it as overfit: in-sample Sharpe of 0.74 collapsed to 0.288 out-of-sample, PBO hit 0.69, the holdout held just 9 trades, and the deflated Sharpe of 0.025 confirmed the selected result was noise across 225 trials. The analyst abandoned it after one iteration, noting that a uniformly mediocre sensitivity grid (0.23–0.53) left no robust region to tune toward and that near-identical prior ETH/BTC/SOL Hyperliquid ports had all died the same way.

Porting a genuinely proven mechanism family onto a new asset/venue does not by itself add robustness — deflation strips the edge when the sensitivity grid is uniformly mediocre and the venue's history is too short to power out-of-sample inference for a strategy this slow.

The post-optimization analyst verdict was abandon: walk-forward flagged is_overfitted=TRUE (IS avg Sharpe 0.74 collapsing to OOS 0.288 with windows scattered -2.99/+0.59/+3.26), OOS Sharpe 0.288 < 0.5, PBO 0.69 > 0.5, holdout carried only 9 trades (< 10), and deflated Sharpe 0.025 with programme-FDR failure made the result indistinguishable from best-of-N noise over 225 trials.

A deliberate venue port of the factory's one paper-stage-reaching mechanism family — a single-instrument, long-short dual-timeframe (4H primary + 1D confirm) EMA momentum-confluence trend follower with an ATR trailing exit — moved onto ETHUSD.HYPERLIQUID perps to test venue-robustness of a proven edge.

The initial backtest looked passable: +41.6% total return (CAGR 11.4%), Sharpe 0.625, profit factor 1.23 over 158 trades (72 long / 86 short) with avg per-trade return 0.74% clearing the 0.15% fee floor. After optimization it degraded — Sharpe 0.38, profit factor 1.11, +19.3% return over 126 trades — and validation failed.