ADASequentialHigherLowBreakoutLong

ADA 4H N-Bar Sequential Higher-Low Breakout Long (Rising-Support Staircase Pattern)

A long-only, single-instrument, single-timeframe strategy on ADAUSDT.BINANCE 4H bars (audit_stale issue RESOLVED). Implements a NEW structural mechanism analogous to the analyst-validated Two-Bar Sequential Higher-High Breakout (Sharpe 1.86, walk-forward IS/OOS ratio 1.11) but operating on LOWS instead of HIGHS — capturing defensive support-building rather than momentum acceleration. The mechanism: detect 3+ consecutive bars where each bar's LOW exceeds the prior bar's LOW (sequential higher-lows = rising support staircase). When this pattern resolves with the current bar breaking above the recent 20-bar Donchian high AND trend filter holds, enter long. Different signal than higher-highs: rising LOWS indicate buyers are systematically defending at higher levels even without making new highs — this is a more stealthy accumulation pattern that often precedes acceleration. Critically distinct from every existing ADA 4H mechanism: (a) NOT XRPTwoBarSequentialHigherHighBreakoutLong (uses LOWS not HIGHS, 3+ bars not 2 bars); (b) NOT double-bottom support reclaim (uses sequence of rising lows, not low-equality match); (c) NOT Donchian midpoint reclaim (uses sequence + channel breach, not midpoint pivot); (d) NOT inside-bar cluster (uses progressive HIGHER lows, not contracting ranges); (e) NOT statistical signals. SEQUENCE-BASED + RANGE-STATISTICS per analyst's explicit successful-pattern profile.

Iteration 2 addresses the optimization-stage statistical failure (PBO 0.73, deflated_sharpe 0.74 < 0.95, holdout ratio 0.37) — NOT a logic defect (the edge is robust: all OOS windows positive, sharpe_ci_low 1.51, sensitivity 0 cliffs). Root cause per the analyst: a 225-trial search over NINE parameters inflated the expected-max-under-noise, raising PBO and deflating Sharpe. The optimizer (walk_forward.py) builds its search space from every numeric key in config["parameters"], varying each +-50%. Fix = shrink that dict to ONLY the three genuine, monotone-response edge knobs (sequence_bars, donchian_period, stop_loss_pct). The six near-flat parameters are removed from the config so Optuna never suggests them; the strategy code falls back to self.parameters.get(name, default) whose defaults already equal the prescribed constants (sma_period=50, take_profit_pct=6, max_holding_bars=30, regime_breakdown_pct=5, per_trade_pct=0.20, min_notional=5), fixing them to constants with ZERO behavioral change. sequence_bars is kept at base 3 (the robust, higher-trade-count operating point, 214 trades / Sharpe 1.97) rather than the over-searched 4. Collapsing 9-D to 3-D directly lowers PBO and lifts DSR toward >=0.95 while preserving the real edge. No code changed; imports unchanged; all earlier layers stay green.

Failed every forward gate on attempt 2 of 2. Deflated Sharpe DSR=0.0349 (vs 0.95), with the optimized Sharpe 1.575 FAR below the 225-trial expected-max luck bar of 2.853 (is_significant=false, PBO 0.5609) — a 4H strategy's high effective-trial count makes the luck bar high and the edge does not clear it. Decisively, the time-ordered HOLDOUT FAILED and is NEGATIVE (holdout_sharpe -0.475, ratio -6.54 vs WF-OOS 0.073): the untouched recent window loses money. The walk-forward is is_overfitted=TRUE with avg IS 3.40 collapsing ~47x to avg OOS 0.073 (windows [-0.98, 0.90, 0.30]) — the best-of-225 config overfit the in-sample windows. The clean sensitivity grid (0 cliffs) and PSR 0.987 measure consistency, not significance, and cannot override a negative holdout plus DSR 0.035 (the PSR-vs-DSR trap). Not iterate: attempt 2 of 2 is exhausted and the binding failure is forward non-generalization (negative holdout, collapsing OOS), not an under-explored parameter region. Not revise_hypothesis: the sequential-breakout family is not a proven mechanism stranded on a dead target — its sibling (ADA Two-Bar Higher-High breakout) was likewise abandoned post-optimization for failing deflation, and this higher-low variant reaches the same end. FAILURE PATTERN: a strong, well-distributed base (Sharpe 1.81, positive 6/7 years, low kurtosis 4.4, CI-low +0.43) that looked like the session's best pre-optimization candidate STILL fails promotion — best-of-225 selection lands the optimized Sharpe (1.575) below the 2.85 luck bar (DSR 0.035), the walk-forward is overfit (IS 3.40 -> OOS 0.07), and the untouched forward holdout is NEGATIVE (-0.475). An attractive base with a clean sensitivity grid is not evidence of a promotable edge when deflation and a time-ordered holdout are applied; single-asset 4H sequential-breakout variants on ADA do not survive the multiple-testing and forward gates.

Long-only single-instrument ADAUSDT.BINANCE 4H rising-support staircase + Donchian breakout. Detects sequence_bars consecutive strictly-higher-lows (rising support = stealthy accumulation), and enters long when the current close breaks above the prior donchian_period-bar Donchian high while holding above the 50-bar SMA trend filter. Exits on stop_loss_pct hard stop, 6% take-profit, 30-bar time stop, or a 5%-below-SMA regime breakdown. Signal logic and all trading behavior are unchanged from the analyst-validated edge; this revision only shrinks the optimizer search surface.

Large sample (215 long trades over 6+ years); trades faithfully implement the hypothesis (0 shorts, rising-higher-low + Donchian breakout), entry_diag 215->215 with 0 dropped

Consistent across every regime: positive in 6 of 7 years including the 2022 bear (+10.3%) — genuinely diversified, NOT 2021-outlier-concentrated like the Donchian siblings

Well-behaved distribution (kurtosis 4.4, skew 1.0, tail_ratio 1.75) — the edge is spread across many small trades, not carried by outlier days

Statistically healthy start: Sharpe 1.81 with sharpe_ci_low +0.43 (CI does NOT straddle 0), PSR 0.9956; low max_drawdown 8.8%, clean avg_position_pct 21%

Modest absolute return (CAGR 8.3%) — but exposure is only 13.4% and max_DD 8.8%, so there is headroom for the optimizer to scale per_trade_pct given the strong edge-to-risk

Market impact drag: impact_cost_pct 4.98% of gross, capacity ~$40M — watch sizing into ADA liquidity

Sibling caution: the Two-Bar Higher-HIGH variant was abandoned post-optimization for failing deflated Sharpe (1.29 vs luck bar 1.18); best-of-225 deflation is the real test and is not guaranteed here

long-only rising-low breakouts

long breakout

Clean sensitivity grid (0 cliffs); large sample (215-232 trades); base looked well-distributed (positive 6/7 years, kurtosis 4.4). Trades match the hypothesis (long-only rising-low + Donchian breakouts)

Fails deflated Sharpe: DSR 0.0349 (vs 0.95), optimized Sharpe 1.575 FAR below the 225-trial expected-max luck bar 2.853, is_significant=false, PBO 0.5609

HOLDOUT FAILED and is NEGATIVE: holdout_sharpe -0.475, ratio -6.54 vs WF-OOS 0.073 — the untouched recent window loses money

Walk-forward is_overfitted=TRUE: avg IS 3.40 collapses to avg OOS 0.073 (windows [-0.98, 0.90, 0.30]) — the optimized config overfit the in-sample windows

PSR 0.987 vs DSR 0.035 is the multiple-testing trap; a clean sensitivity grid does not rescue a negative holdout and sub-luck-bar deflation

The edge is genuine and robust (sensitivity 0 cliffs; walk-forward is_overfitted=FALSE with all OOS windows positive [2.97, 2.45, 0.80], avg OOS 2.07; sharpe_ci_low 1.51 excludes 0; holdout still positive at 0.76). It does NOT clear promotion because the 225-trial search over NINE parameters inflated PBO to 0.73 and left deflated_sharpe at 0.74 (<0.95), and the holdout retained only 37% of WF-OOS Sharpe. The fix is to SHRINK the parameter surface so the optimizer deflates less and generalizes better: 1) FIX the near-flat (irrelevant) parameters to constants — the sensitivity grid shows their Sharpe barely moves across +-20%: set min_notional=5, regime_breakdown_pct=5, sma_period=50, take_profit_pct=6, max_holding_bars=30, per_trade_pct=0.20. Removing these from the search directly lowers the expected-max-under-noise bar and PBO, and should lift DSR toward/over 0.95 since the underlying edge is real. 2) RE-OPTIMIZE ONLY the genuine edge parameters: sequence_bars and donchian_period (optionally stop_loss_pct). These are the only knobs the sensitivity grid shows real, monotone Sharpe response to. 3) PREFER the higher-trade-count / more-robust region. The optimizer pushed sequence_bars 3->4, cutting trades 214->155 and worsening the holdout (ratio 0.37). Keep sequence_bars=3 (test 2-3 only); the base 3-bar config had 214 trades, Sharpe 1.97, and sharpe_ci_low 0.68 — a more generalizable operating point. The goal of the re-run is DSR >= 0.95 AND holdout ratio >= 0.70 AND PBO <= 0.5, not a higher raw Sharpe. If attempt 2 still fails DSR/holdout after the parameter-surface reduction, abandon — the recent-year decay (2026 negative) would then indicate the edge is fading rather than just over-searched.

ADASequentialHigherLowBreakoutLong adapted the analyst-validated two-bar higher-high breakout onto LOWS, entering when a rising-support staircase of 3+ higher lows resolved into a Donchian-high breakout in an uptrend, capturing stealthy accumulation rather than momentum acceleration. Its base backtest looked like the session's strongest candidate — Sharpe 1.81, profit factor 1.44, positive in 6 of 7 years including the 2022 bear, low kurtosis, an 8.8% max drawdown, and a Sharpe CI clear of zero — earning an 'optimize' verdict. But optimization exposed the same fragility as its abandoned higher-high sibling: the walk-forward was overfit (IS 3.40 → OOS 0.07), it failed deflated Sharpe (DSR 0.035, optimized 1.575 below the 2.85 luck bar), and the untouched forward holdout was negative (-0.475). On the second and final iteration, with the binding failure being forward non-generalization rather than a tunable region, the analyst abandoned it, concluding that a clean base and sensitivity grid cannot rescue a strategy that loses money out-of-sample once deflation and a time-ordered holdout are applied.

An attractive, well-distributed base (Sharpe 1.81, positive 6/7 years, low kurtosis, CI clear of zero) with a clean sensitivity grid is not evidence of a promotable edge — best-of-225 selection can still land below the multiple-testing luck bar with a negative forward holdout, and single-asset 4H sequential-breakout variants on ADA repeatedly fail these forward gates.

The analyst abandoned it at the post-optimization ANALYZING stage on attempt 2 of 2: it failed every forward gate — deflated Sharpe (DSR 0.0349 vs 0.95, optimized Sharpe 1.575 far below the 225-trial luck bar of 2.853, PBO 0.5609), an overfit walk-forward (avg IS 3.40 collapsing ~47× to avg OOS 0.073), and a decisively negative time-ordered holdout (holdout_sharpe -0.475, ratio -6.54).

A long-only, single-instrument ADAUSDT.BINANCE 4H breakout strategy that detected a rising-support staircase — 3+ consecutive bars of strictly higher lows (stealthy accumulation) — and entered long when that sequence resolved with a break above the prior 20-bar Donchian high while price held above its long SMA, exiting on hard stop, profit target, time stop, or SMA regime breakdown.

The base backtest was the session's most attractive pre-optimization candidate: total return +74.2% over 215 long trades (0 shorts), Sharpe 1.81 with sharpe_ci_low +0.43 (CI clear of zero), profit factor 1.44, positive in 6 of 7 years including the 2022 bear (+10.3%), well-distributed (kurtosis 4.4), with only 8.8% max drawdown; optimization lowered it to Sharpe 1.575 / +58.4% return.