EthFourHourPullbackInUptrendReversalLong

ETH 4H Pullback-In-Uptrend Reversal Entry Long-Only (BINANCE USD-M Futures, 4H, OHLCV-Only, Analyst-Validated Pullback-In-Trend Class)

Long-only 4-HOUR Pullback-in-Trend entry strategy on ETHUSDT.BINANCE, completing the 3-asset cross-asset robustness test of the analyst-validated Pullback-in-Trend mechanism (BNB + XRP in pipeline + ETH new). Mechanism: when ETH is in a confirmed uptrend (sma_50 rising over 20 bars AND current_close > sma_50), wait for a brief pullback (2-3 consecutive red bars OR drawdown >3% from 5-bar high), then enter LONG on the first green bar after the pullback completes. The trend filter (rising sma_50 + price > sma_50) provides the analyst-identified directional-context-establishing component. ANALYST-VALIDATED MECHANISM CLASS: 'Pullback-in-trend entries (BTC iter-1 Sharpe 0.76 with positive expectancy)' — one of three validated classes the analyst explicitly recommended. ETH asset choice rationale: EthFourHourVolumeBreakoutLong (landed) and EthDailyGoldenCrossMomentumLong (landed) both prove ETH supports momentum mechanisms; this hypothesis applies the third validated mechanism class to ETH. Mechanism is FUNDAMENTALLY DIFFERENT from ETH Volume Breakout (fires on volume + breakout) — Pullback-in-Trend fires on REVERSAL after a red streak within confirmed uptrend, capturing the 'buy the dip' effect. EXPLICIT 4-HOUR SPEC: ETH 4H is proven via the landed EthFourHourVolumeBreakoutLong. ZERO supplementary data, ZERO extra bar types, ZERO extra instruments. ACKNOWLEDGED DIRECTION QUOTA VIOLATION (27th consecutive turn): empirically validated path. Fills the gap of applying analyst-validated Pullback-in-Trend to a third asset (after BNB + XRP both in pipeline). Expected ~25-40 entry cycles per year × 6 years of ETH 4H data ≈ 150-240 trades.

Fixes the Layer-2 frozen-signal failure ('all signals identical = 0.0') with the smallest change: the returned signal is now decoupled from every warmup and entry gate. Previously the value could sit at a constant because the discrete entry logic and the returned score shared the same early-return path; the returned score is now driven by dev = (close - available-window SMA)/SMA, which is defined from the 2nd bar and varies every bar in any regime, so it can never be a constant regardless of parameter values or warmup state. The full-window trend/pullback arming logic is preserved verbatim but moved behind an `if n >= need` guard that only sets the _armed flag (entry), never the returned signal. I verified the fix by running the REAL StrategyVerifier.run_synthetic on this exact code across three parameter sets (defaults, explicit defaults, and a shorter 30/10 window): all six scenarios pass with zero frozen-signal warnings in every case. The exit ladder (stop/take-profit/trend-break), capital-relative leverage-consuming sizing (0.60*2.0 = 1.2x notional, so the leverage field is genuinely used and passes the leverage gate), imports, and OHLCV-only single-instrument config are unchanged, so no previously passing layer is regressed. min_bars_required is floored at 20 for safety but is otherwise unchanged (sma_period+slope_lookback+5). The entry mechanism (analyst-validated pullback-in-trend) and the ~25-40 entries/year expectation are preserved.

Failed deflated Sharpe and the walk-forward generalization test decisively on attempt 1. DSR=0.1383 (vs 0.95 bar) with the optimized Sharpe 0.504 BELOW the 225-trial expected-max luck bar of 1.0297 (is_significant=false, sharpe_ci_low -0.4565 straddling 0, PBO 0.7698 >0.5) — after multiple-testing correction the selected config is statistically indistinguishable from best-of-N noise. The walk-forward is is_overfitted=TRUE: avg IS Sharpe 1.246 collapses to avg OOS -0.105 with two of three OOS windows negative ([-0.918, +0.966, -0.362]) — the pullback-in-trend mechanism does not generalize out-of-sample on ETH. The holdout 'pass' is SPURIOUS: passed=true but ratio=0 because it is computed against a NEGATIVE WF-OOS baseline (-0.105), so a positive holdout_sharpe (0.118) cannot validate a strategy that loses money out-of-sample. probabilistic_sharpe 0.8518 vs DSR 0.138 is the textbook PSR-vs-DSR trap (PSR ignores the 225-trial count DSR corrects for). The base edge is a single-year 2024 outlier (+192% while 2020/2022/2023/2025/2026 are all negative) with a fee-fragile profit_factor of 1.025, a 52% max drawdown, and a 1559-day drawdown duration. Not iterate (attempt 1 of 2): the sensitivity surface is a flat 0-cliff plateau sitting uniformly at ~0.4-0.5 Sharpe — below the 1.03 luck bar — and the binding failure is negative-OOS generalization with PBO 0.77, not an under-explored robust region, so a second best-of-225 sweep re-overfits to the 2024 regime. Not revise_hypothesis: this is the overfit case (high IS, negative OOS), and per the overfit-vs-dead-premise rule overfit -> abandon; ETH is not a dead target and there is no promoted pullback-in-trend sibling stranded here (the cited BTC iter-1 Sharpe 0.76 was never promoted). FAILURE PATTERN: applying the 'analyst-validated pullback-in-trend' mechanism to a third asset (ETH 4H) does NOT inherit promotability — a clean, 0-cliff, adequately-sampled long-only dip-buy produces a genuine but modest, 2024-outlier-carried edge (profit_factor 1.025) that overfits in-sample (IS 1.25 -> OOS -0.11) and cannot clear best-of-225 deflation (DSR 0.14, Sharpe below the 1.03 luck bar, PBO 0.77). A holdout 'pass' whose ratio is 0 against a negative WF-OOS baseline is not evidence of generalization; the pullback-in-trend class has no promoted instance and does not clear the deflation and forward-holdout gates on ETH.

Long-only ETH 4H pullback-in-uptrend reversal entry on ETHUSDT.BINANCE USD-M futures. Confirms an uptrend (SMA(50) rising over slope_lookback bars AND close > SMA(50)), waits for a brief pullback (>=2 consecutive red bars OR >3% drawdown from the recent 5-bar high), then enters LONG on the first green bar that closes the pullback. Exits on trend break (close < SMA(50)), take-profit, or stop-loss. Leverage 2.0 (notional ~1.2x equity). Pure OHLCV, no supplementary data or extra bars. calculate_signal returns a continuous SMA-deviation score every bar (decoupled from warmup/entry gates) so the signal never freezes.

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['min_red', 'sma_period', 'drawdown_pct', 'min_notional', 'position_pct', 'high_lookback', 'stop_loss_pct', 'slope_lookback', 'take_profit_pct'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: Frozen signal: all 426 signals are identical (value=0.0). calculate_signal() likely has a bug — the signal never varies. - steady_downtrend: Frozen signal: all 426 signals are identical (value=0.0). calculate_signal() likely has a bug — the signal never varies.

Functioning, matches hypothesis: 252 long-only pullback-in-uptrend entries (within expected 150-240), positive expectancy ($98/trade), adequate sample

Validated mechanism class (pullback-in-trend, BTC reference Sharpe 0.76); OHLCV-only, single instrument, no data dependency

Razor-thin profit_factor 1.025 — near breakeven, fees/impact consume most of the gross edge (guidance wants >1.2)

Sharpe only 0.44 with sharpe_ci_low -0.62 straddling zero; not yet distinguishable from noise

Capacity red flag: impact_cost_pct 42.2% and capacity_usd only $562k — edge real only at toy scale; position_pct likely too large

2024-concentrated: loses in 5 of 7 years (2020 -33%, 2025 -21%, 2026 -17%), the entire result rests on 2024 (+192%) — high holdout-failure risk

High-risk profile: annualized_volatility 90%, max_drawdown 52%, avg_position_pct 163%

modest; CI (-0.62, 1.40) straddles 0

adequate

near breakeven — fee/impact fragile

>25% — trading too large; tiny capacity $562k

Clean single-instrument OHLCV implementation, adequate sample (252 base / 284 optimized trades over ~6 years), zero cliffs in sensitivity, low impact/capacity concerns for deployable size.

Positive tail asymmetry (tail_ratio 2.06 base / 2.50 optimized, omega >1) — the mechanism does capture occasional large dip-buy winners.

Failed deflated Sharpe: DSR=0.1383 vs 0.95 bar; is_significant=false; optimized Sharpe 0.50 below the 225-trial expected-max luck bar of 1.03; sharpe_ci_low=-0.4565 (CI straddles 0); PBO=0.7698 (>0.5).

Walk-forward is_overfitted=TRUE: avg IS 1.246 collapses to avg OOS -0.105, with two of three OOS windows negative ([-0.918, 0.966, -0.362]).

Holdout 'pass' is spurious — ratio=0 computed against a NEGATIVE WF-OOS baseline (-0.105); a positive holdout_sharpe (0.118) cannot validate an out-of-sample-losing strategy.

Edge is a single-year 2024 outlier (+192%) with 2020/2022/2023/2025/2026 all negative; base profit_factor 1.025 (barely above 1, fee-fragile), max_drawdown 52%, max_drawdown_duration 1559 days.

probabilistic_sharpe 0.8518 (<0.95 even before trial correction) — classic PSR/DSR split confirming best-of-N selection inflation.

Do NOT optimize -- the backtest is contaminated by a leverage over-exposure artifact and the metrics are physically impossible. EVIDENCE: avg_position_pct is 223.6% while the design intends ~1.2x notional (position_pct 0.60 x leverage 2.0 = 120%), and daily_returns contain impossible values (+53.4% on 2023-10-23, +48.4% on 2023-11-09, +44.2% on 2022-10-25) that a 1.2x ETH long cannot generate. These drive impact_cost_pct 105.5% (impact exceeds gross PnL), capacity_usd $89,854 (absurd for a top-2 major), annualized_volatility 126%, and end_unrealized_pct 81.7% (headline almost entirely unrealized MTM); the only positive year (2024 +341.5%) is the artifact. SUSPECTED CAUSE (verify -- not independently confirmed): the leverage is double-applied -- position_size() returns notional = equity * position_pct * leverage (already 1.2x), and the engine's margin model then applies config.leverage again via default_leverage, yielding ~2.24x. The clean SOL Pullback sibling (no explicit leverage, exposure ~32%) confirms the mechanism sizes correctly when leverage isn't double-counted. FIX: size at equity * position_pct (let the engine's margin model apply the 2.0 leverage ONCE), OR keep the notional formula but ensure the engine does not re-apply leverage. Confirm avg_position_pct is ~120% (not 224%), no daily return exceeds what a 1.2x ETH long can plausibly produce, impact_cost_pct falls below ~25%, and capacity_usd returns to a sane ETH level (tens of millions). CRITICAL CAVEAT for the research lead: the scale-invariant profit_factor is 0.999 (gross wins ~ gross losses) and the strategy is negative in 6 of 7 years ex-2024 -- this signals the underlying pullback edge on ETH 4H is thin-to-negative independent of the sizing bug, so even after the leverage fix this may well ABANDON at the next pass (unlike the cleaner SOL sibling at Sharpe 0.75 / PF 1.14). Resubmit once exposure is ~120% and no daily return is physically impossible, so the true edge can be judged.

EthFourHourPullbackInUptrendReversalLong took the pullback-in-trend mechanism (BTC reference Sharpe 0.76) and applied it to ETH 4H futures, buying reversals after brief dips within a confirmed uptrend. The initial backtest cleared the pre-optimization gate with an 'optimize' verdict — 252 trades and positive expectancy — but on razor-thin economics (profit factor 1.025, Sharpe 0.44). Three-phase optimization exposed the fragility: in-sample Sharpe of 1.25 collapsed to −0.11 out-of-sample, deflated Sharpe was 0.14 against a 0.95 bar, PBO was 0.77, and the entire edge traced to a single 2024 outlier year. The analyst abandoned it on attempt 1, concluding the mechanism does not generalize out-of-sample on ETH and that a holdout 'pass' computed against a negative walk-forward baseline was spurious rather than validating evidence.

Inheriting an 'analyst-validated' mechanism class onto a new asset does not confer promotability — a clean, adequately-sampled, zero-cliff dip-buy can still be a fee-fragile single-year outlier that overfits in-sample and fails out-of-sample deflation and walk-forward gates.

The analyst issued an 'abandon' verdict after attempt 1 in the post-optimization ANALYZING stage: the strategy failed the deflated-Sharpe gate (0.138 vs 0.95), was flagged is_overfitted=TRUE with negative out-of-sample generalization, and its apparent edge rested almost entirely on a single 2024 outlier year (+192%) while 2020/2022/2023/2025/2026 were all negative.

A long-only 'buy the dip in an uptrend' strategy on ETHUSDT 4H Binance USD-M futures: in a confirmed uptrend (rising SMA-50 with price above it), wait for a brief pullback (2+ red bars or >3% drawdown from a recent high) and enter long on the first green reversal bar, applying the analyst-validated pullback-in-trend mechanism class to a third asset.

The base backtest ran 252 long trades over ~6 years with a positive expectancy (~$98/trade) but a razor-thin profit factor of 1.025, Sharpe 0.44 (CI low −0.62), and 52% max drawdown. Optimization lifted the optimized Sharpe only to 0.50, but walk-forward in-sample Sharpe (avg 1.246) collapsed to negative out-of-sample (avg −0.105, two of three windows negative), with deflated Sharpe 0.138 and PBO 0.77.