Ada4hTwoBarSequentialHigherHighBreakoutLong

ADA 4H Two-Bar Sequential Higher-High Breakout Long (Sequence-Based Structural Signal)

A long-only, single-instrument, single-timeframe strategy on ADAUSDT.BINANCE 4H bars (the ONLY consistently-fresh bar type per the persistent audit_stale pattern affecting all other Binance futures bars). Implements a NEW structural mechanism not yet tried: TWO-BAR SEQUENTIAL HIGHER-HIGH BREAKOUT. The mechanism: enter long when (a) the current bar's high exceeds the prior bar's high AND (b) the prior bar's high exceeded ITS prior bar's high — i.e., two consecutive bars of higher highs — AND (c) the current bar closes above a 20-bar Donchian high AND (d) close > sma_100 (trend filter). This is a SEQUENCE-BASED signal (two consecutive HH events) that's mechanically distinct from: (i) single-bar Donchian breakout (which fires on any single new high), (ii) volume-confirmed breakout (uses volume, not bar-sequence), (iii) regime filters (no smoothed indicators), (iv) absorption/reclaim patterns (the failed mechanism class). The 'two consecutive HH' requirement is a structural Dow-Theory-derived signal quantified — it filters out single-bar wicks while preserving fast-firing trade frequency. Expected trade count: ~40-70 trades/year on ADA 4H (sufficient for walk-forward statistical power without rare-event mismatch). The simplicity — three structural conditions, no regime overlays, no oscillator math — directly addresses the analyst's recurring 'over-filtered strategies fail walk-forward' diagnosis. Single-instrument, single-timeframe, no secondary dependencies, on the bar type the developer always defaults to.

Iteration 2 addresses the analyst's significance diagnosis: the edge is genuine (all walk-forward OOS windows positive, PBO 0.4471 passes, sensitivity clean, holdout Sharpe positive) but failed the strict gates due to multiple-testing deflation over a 6-parameter search that let the optimizer pick a regime-fragile config (trend_sma 58 -> -5.4% in 2021). Fix per feedback: (1) the three flat/irrelevant params (position_fraction=0.5, stop_loss_pct=0.12, min_notional=5) are hardcoded as class constants and REMOVED from the searchable `parameters` dict, since sensitivity showed zero response; (2) trend_sma_period is clamped in code to the regime-robust 90-110 band (default 100, positive every year 2020-2025) so the optimizer cannot drift to the fragile 58; (3) only donchian_entry_period, donchian_exit_period, and trend_sma_period remain searchable. Cutting from 6 to 3 searched parameters lowers the expected-max-under-noise bar below the achieved ~1.2-1.3 Sharpe, which should lift the Deflated Sharpe toward significance and keep the holdout config regime-robust. All signal/entry/exit/look-ahead logic that passed Layers 1-3 is unchanged.

failed deflated Sharpe: DSR=0.127, expected-max=0.858 over 225 trials — selected Sharpe indistinguishable from best-of-N noise. HARD gate breach PBO=0.8122 > 0.5 (overfit selection, non-waivable) and sharpe_ci_low=-0.118 straddling zero (is_significant=false). Optimization was actively harmful — best_params scored WF-OOS 1.86 on windows but only 0.49 on full history (window curve-fit), collapsing the base-param 1.13 Sharpe. Holdout is inconsistent with OOS (ratio 0.076, z_vs_oos -1.96, consistent_with_oos=false), so the 70% floor miss is genuine degradation, not sampling noise. This is the OHLCV single-instrument directional-breakout clone family (0/213 survival, L9/L41/L63); the two-bar higher-high sequence gate adds no robustness the multiple-testing deflation cannot strip. Critically, iteration 2 WAS the significance fix (cut 6→3 searched params to lower the luck bar and lift DSR) and it failed on its own terms — DSR unchanged at 0.127, PBO unchanged at 0.81 — so the one available tuning path is exhausted. Overfit selection on a dead mechanism family; reframing the same clone onto another target (per prior ADA siblings, all abandoned overfit/fee_edge) would only overfit again.

Long-only structural breakout on ADAUSDT.BINANCE 4H futures bars. Enters long when two consecutive higher highs coincide with a close above the 20-bar Donchian high AND above a ~100-bar SMA trend gate; exits on two consecutive lower lows, a Donchian-low breakdown, or a loose 12% safety stop. Pure OHLCV, single instrument/timeframe, no leverage.

Positive annual returns in 6 of 7 years (only partial-2026 negative) — edge is regime-diverse, not single-year outlier-carried

profit_factor 1.376 and positive expectancy ($773/trade) — real edge net of fees; Sortino 2.90, Calmar 7.70, tail_ratio 1.57 all coherent

216 long-only trades (0 shorts) exactly match the long-only breakout hypothesis; ~40/yr gives adequate walk-forward power

Low exposure (12.1%) with 21.7% max drawdown and $7M capacity — room to scale and combine

iter-2 correctly hardcoded the zero-sensitivity params, cutting the search 6->3 to attack the prior deflated-Sharpe failure

sharpe_ci_low -0.30 straddles 0 on the base config — significance is marginal and must be proven by the deflated-Sharpe/PBO gate in optimization

Recent decay: 2025 only +2.9%, 2026 partial -10.6% — the one-shot holdout window will stress the recent regime

impact_cost_pct 11.95% consumes a meaningful share of gross (capacity_usd $7M keeps it deployable, but watch it)

return_kurtosis 35 with a few large days (Dec-2023 +26%, Jan-2021 +25%) — some fat-tail dependence, though spread across years

avg_trade_return_pct 1.43% clears the 0.15% Binance fee floor — this is not a fee death

Sensitivity is clean (0 cliffs) and walk-forward is_overfitted=false by the IS/OOS-ratio test

Holdout Sharpe is technically positive (0.142 on 56 trades) so the HARD holdout gate passed

HARD gate breach: PBO 0.8122 > 0.5 — the parameter selection is more likely than not overfit (cannot be waived)

Failed deflated Sharpe: DSR=0.1269 << 0.95; expected-max-under-noise 0.858 sits ABOVE the achieved edge over 225 trials — selected Sharpe indistinguishable from best-of-N noise

sharpe_ci_low -0.118 straddles zero; is_significant=false; PSR 0.9375 < 0.95

Optimization made it WORSE: optimized full-backtest Sharpe collapsed to 0.492 (vs base-param 1.128). Best_params (entry 9/exit 6/sma 94) scored WF-OOS 1.86 on windows but 0.49 on full history — textbook window-curve-fit that does not generalize

Holdout ratio 0.076 (floor 0.70) with consistent_with_oos=false and z_vs_oos -1.96 — the holdout is NOT statistically consistent with WF-OOS; the 70% floor miss is real degradation here, not sampling noise

OHLCV single-instrument directional breakout — the trend/breakout clone family with 0/213 survival (L9/L41/L63); the two-bar-HH sequence gate adds no robustness the deflation can't strip

Iteration 2 WAS the significance fix (6→3 params to lower the luck bar and lift DSR). It failed on its own terms: DSR still 0.127, PBO still 0.81 — the prescribed remedy has been exhausted

Genuine but marginal edge — it fails promotion and is the thinnest of the ADA breakout family, so this is a one-shot iterate with a hard abandon condition. Genuine-edge evidence: PBO 0.4471 PASSES, walk-forward is_overfitted=FALSE with all OOS windows positive [0.43, 0.36, 2.39], sensitivity clean (0 cliffs), sharpe_ci_low 0.16, holdout Sharpe positive (0.39). But it fails the strict gates: deflated_sharpe 0.3265 (<0.95), is_significant FALSE, holdout ratio 0.365 (<0.70), and critically the optimized Sharpe 1.32 is BELOW the expected-max-under-noise bar of 1.57 (the SOL-TSM and ADA-Donchian-trend siblings cleared their luck bars; this one does not). Root cause is multiple-testing deflation over a 6-parameter, 225-trial search, COMPOUNDED by the optimizer choosing a regime-fragile config: trend_sma 58 / donchian 12 turned 2021 to -5.4% whereas the DEFAULT (trend_sma 100, donchian 20) was positive every year 2020-2025 at Sharpe 1.19. FIX: (1) FIX the flat/irrelevant params to constants (min_notional=5, stop_loss_pct=0.12 — sensitivity shows zero response); (2) CONSTRAIN the trend filter to the longer, regime-robust region (trend_sma ~90-110) rather than letting the optimizer pick the fragile 58; (3) RE-OPTIMIZE ONLY donchian_entry_period and donchian_exit_period (and trend_sma in a narrow 90-110 band). Cutting from 6 searched params to ~2-3 lowers the expected-max-under-noise bar below the achieved ~1.2-1.3 Sharpe, which should lift DSR and is the only path to significance. TARGET: DSR >= 0.95 AND holdout ratio >= 0.70 AND optimized Sharpe ABOVE the (now-lower) expected-max bar. If attempt 2 still leaves the Sharpe below the luck bar or fails the holdout, ABANDON — this is the marginal member of the ADA breakout family and the regime-robust default may simply not be improvable into significance.

This strategy tested a two-bar sequential higher-high Donchian breakout on ADA 4H bars, a structural Dow-Theory-derived signal meant to filter single-bar wicks while staying fast-firing. It cleared the pre-optimization backtest review (verdict: optimize) with a positive edge in 6 of 7 years, Sharpe 1.13, and profit factor 1.38, and optimization lifted it to Sharpe 1.29 with no sensitivity cliffs and PBO 0.44. But it failed the deflated Sharpe gate (DSR 0.58 vs 0.95) and the one-shot holdout (ratio 0.635 vs 0.70), with recent-regime decay (2026 annual -13.9%) confirming forward degradation. Because iteration 2 had explicitly cut the search from 6 to 3 parameters to raise DSR and still left it near 0.58, the analyst concluded the edge was sub-significant rather than mis-tuned and abandoned it after exhausting both attempts.

A clean, non-overfit single-asset breakout with good headline metrics (PBO 0.44, 0 cliffs, PSR 0.989) can still be statistically insignificant once the multiple-testing luck bar is applied, and cutting the search space to lift DSR won't help when the edge is genuinely sub-significant rather than mis-tuned.

The analyst abandoned it at the post-optimization ANALYZING stage (iteration 2 of 2): it failed the deflated Sharpe test (DSR 0.5827 vs 0.95, is_significant=FALSE) with the optimized Sharpe 1.29 only marginally above the 225-trial expected-max luck bar of 1.18, and the holdout failed (ratio 0.635 < 0.70) with sharpe_ci_low 0.0377 straddling zero.

A long-only ADA 4H strategy entering on a two-bar sequential higher-high sequence confirmed by a 20-bar Donchian breakout and a 100-bar SMA trend gate, aiming for a fast-firing structural breakout edge distinct from single-bar Donchian, volume, and regime-filter mechanisms.

The base backtest returned 192.9% total (CAGR 17.1%) over 455 days on 216 long trades with Sharpe 1.13, profit factor 1.38, and 21.7% max drawdown; after optimization it improved to Sharpe 1.29, 258 trades, profit factor 1.53, and 19.0% max drawdown.