AdaFourHourDoubleLowSupportReclaimLong

ADA 4H Double-Low Support Reclaim Long (Sequential Low-Equality Structural Pattern)

A long-only, single-instrument, single-timeframe strategy on ADAUSDT.BINANCE 4H bars (the most-reliably-fresh bar type). Implements a NEW STRUCTURAL MECHANISM consistent with the analyst's explicit guidance: 'STRUCTURAL signals (sequence-based or range-statistics-based) are what works on liquid crypto — drop momentum-magnitude signals.' The mechanism: SEQUENTIAL LOW EQUALITY (double-low support test). Detect when a recent bar's low approximately matches a prior local low within a small ATR-fraction tolerance — this constitutes a 'support test' or 'double bottom' structural event. Combined with a bullish-close reclaim (current bar closes above the matched low level) and a long-term uptrend filter, the strategy captures the moment when sellers have failed to break support TWICE — a structural signal that buyers are defending the level. Critically distinct from every existing ADA 4H mechanism: (a) NOT Donchian breach (uses LOW-equality, not high-channel breach); (b) NOT bar geometry (uses MULTI-BAR low comparison, not single-bar shape); (c) NOT range outlier (uses low-LEVEL match, not range size); (d) NOT autocorrelation/z-score/cumulative return (no momentum-magnitude math — pure structural sequence); (e) NOT gap/sequential-HH (uses LOW-low matching, not high-high or open-gap); (f) NOT volume signal (no volume needed — structural price-level signal). Mechanism is RANGE-STATISTICS-BASED per analyst's explicit recommendation. Different from the abandoned 'failed-breakdown-reclaim' (PF 0.45 wrong-direction) because this requires LOW-EQUALITY (sellers TWICE failed at same level), not breakdown-then-reclaim (sellers succeeded once then reversed). The thesis is fundamentally different: 'support held twice' vs 'breakdown false-faked'.

Iteration 2 fixes the over-restrictive single-bar gate that produced only 4 trades in 13,776 bars. Per the feedback: (1) DECOUPLED the support touch from the reclaim via an armed-state counter (self._armed_bars) — a double-low detected on bar T arms entry for the next arm_window (6) bars, so the reclaim no longer has to occur on the exact touch bar; (2) WIDENED tol_atr from 0.5 to 1.2 ATR so near-equal lows qualify as a support test; (3) RELAXED the second-touch test from 'literal lowest of last min_separation bars' (lo <= recent_min) to an APPROXIMATE match (lo <= recent_min + tol) so a bar that is near — but not exactly — the recent low still arms; (4) RELAXED the uptrend filter from a hard 'close > SMA-50' to 'close > SMA-50 OR SMA-50 flat-to-rising', removing the contradiction where a fresh support-test low must simultaneously sit above a rising trend SMA. The continuous ATR-normalized reclaim distance is still returned every bar (no gated-to-0.0), imports stay clean (numpy + nautilus only, preserving Layer 1), and the discrete arm/reclaim decision is cached for should_enter. These changes only affect trade generation (thresholds, arming, filter), so earlier passing layers are preserved while targeting the 40+ trade count needed for a meaningful optimization.

Fails deflated Sharpe and the forward holdout decisively on optimization attempt 1. DSR=0.0677 (vs 0.95 bar) with the optimized Sharpe 1.628 FAR below the 225-trial expected-max luck bar of 2.7975 (is_significant=false, sharpe_ci_low=0.0944 hugging 0) — after multiple-testing correction the selected config is statistically indistinguishable from best-of-225 noise; probabilistic_sharpe 0.9812 vs DSR 0.0677 is the textbook selection-inflation trap. Decisively, the untouched time-ordered HOLDOUT FAILED and is NEGATIVE (holdout_sharpe=-2.106, ratio=-1.343 vs WF-OOS 1.568) — the recent regime loses money, realizing the decay flagged at pre-optimization review (rolling Sharpe sliding to ~-8 by 2026; optimized annual 2025 -0.6%, 2026 -0.1%). Sensitivity FAILED with 6 of 14 cliff parameters (lookback, risk_pct, sma_trend, atr_period, position_pct, max_hold_bars) — a knife-edge surface with no robust region to tune toward. The benign signals (walk-forward is_overfitted=false, all three OOS windows positive) measure in-sample walk-forward consistency, not significance or forward-generalization, and the attractive optimized headline (Sharpe 1.63, max_dd 3.8%) is a best-of-225 mirage on that cliff surface. Not iterate: the binding failures are a negative forward holdout and a 6-cliff sensitivity minefield whose best-of-225 selection lands far below the luck bar — there is no robust parameter region, so a second sweep re-selects another sub-significant, cliff-perched config while the recent regime keeps losing. Not revise_hypothesis: ADA 4H is a fine target and the double-low structural pattern is not a proven mechanism stranded on a dead instrument — it is a modest, 2021-concentrated, decaying edge failing multiple-testing deflation, so there is nothing validated to re-target. FAILURE PATTERN: a clean, non-overfit-flagged single-asset ADA 4H structural pattern with positive walk-forward OOS windows still fails promotion because its best-of-225 optimized Sharpe (1.63) cannot clear the 2.80 expected-max luck bar (DSR 0.068), it sits on a 6-of-14-cliff sensitivity surface, and its untouched forward holdout is deeply NEGATIVE (-2.106) from genuine recent decay. A flattering low-drawdown optimized refit and a passing walk-forward do not rescue an edge that fails deflated Sharpe, the forward holdout, and sensitivity simultaneously.

Long-only ADAUSDT.BINANCE 4H structural double-low support-reclaim strategy. Detects a time-separated double-low (sellers test the same support twice within an ATR-fraction tolerance) and ARMS; while armed for up to arm_window bars, enters long on a bullish reclaim (close > prior support low and close > open) in a trend-permissive regime (close > SMA-50 OR SMA-50 flat-to-rising). Exits on a structural stop below the matched support, ATR stop, ATR target, or a time stop. Capital-relative sizing capped at 18% equity notional and reduced to keep per-trade risk <= 1.5% of equity.

Clean, functioning backtest with NO artifacts: sane daily returns, avg_position_pct 17.6% within the 18% cap, normal kurtosis (8.4)

Healthy sample: 193 trades over 6.5 years — sufficient for meaningful walk-forward/sensitivity analysis

Fee-viable edge: ~0.31% net per trade (above the ~0.15% BINANCE futures threshold), profit_factor 1.11, low max_drawdown 13%

Trades match the hypothesized structural mechanism (long-only double-low support reclaim); novel structural signal class as recommended

Weak headline: Sharpe 0.375, CAGR 1.6%, thin profit_factor 1.11 — a marginal edge

Clear regime decay: positive 2020-2023 but negative every year since (2024 -1.1%, 2025 -7.5%, 2026 -3.1%); rolling Sharpe falls from +7 (2021) to -8 (2026)

Recent-window holdout is the key risk — the edge is concentrated in early bull years and has faded, so it may fail the post-optimization holdout and deflated-Sharpe gates

Walk-forward not overfit-flagged (is_overfitted=false, avg IS 2.98 vs avg OOS 1.57) with all three OOS windows positive ([1.23, 2.18, 1.29])

Clean data path, adequate sample (193 base / 135 optimized trades), fee-viable per-trade edge, and a low optimized max_drawdown (3.8%)

Fails deflated Sharpe decisively: DSR=0.0677 (<0.95), optimized Sharpe 1.628 far below the 225-trial expected-max luck bar of 2.7975, is_significant=false, sharpe_ci_low=0.0944 hugging 0

Untouched time-ordered HOLDOUT FAILED and is NEGATIVE: holdout_sharpe=-2.106, ratio=-1.343 vs WF-OOS 1.568 — the recent regime loses money, realizing the pre-optimization decay

Sensitivity FAILED with 6 of 14 cliff parameters (lookback, risk_pct, sma_trend, atr_period, position_pct, max_hold_bars) — no robust region

PSR 0.9812 vs DSR 0.0677 is the classic multiple-testing selection-inflation trap; PBO 0.42

Optimized annual returns are tiny outside 2021 (+14%): 2020 +0.4%, 2022 +0.6%, 2023 +2.6%, 2024 +0.1%, 2025 -0.6%, 2026 -0.1% — the edge is concentrated early and decayed, rolling Sharpe negative into 2026

Negative information ratio (-0.51) — no risk-adjusted edge vs its benchmark

The entry gate is far too restrictive — 4 trades in 13,776 4H bars means the edge cannot be evaluated and optimization would fit 4 data points (worthless). The core problem is requiring too many conditions on the SAME bar, two of which fight each other (a 'fresh recent low' simultaneously in a confirmed uptrend above SMA-50). Loosen so the signal fires often enough to assess (target ~40-60+ trades over the period): (1) Decouple the touch from the reclaim — ARM when a time-separated double-low is detected (current low within tol_atr of the prior support min), then ENTER on the reclaim (close > prior_low and close > open) within the next N bars, rather than demanding the touch+reclaim on one bar. (2) Widen tol_atr from 0.5 to ~1.0-1.5 so near-equal lows qualify. (3) Relax the 'second touch must be the literal lowest of the last min_separation bars' to an approximate match within the tolerance. (4) Consider relaxing the uptrend filter (e.g. close > SMA50 OR SMA50 flat-to-rising) since requiring a fresh low above a rising trend SMA is contradictory at support tests. Re-backtest and confirm the trade count is high enough (40+) to support a meaningful optimization before proceeding.

This strategy replaced an earlier failed-breakdown-reclaim idea with a distinct double-low support-reclaim mechanism on ADAUSDT 4H, and its first backtest earned an 'optimize' verdict on a marginal-but-fee-viable edge (Sharpe 0.375, PF 1.11) despite visible regime decay after 2023. Optimization produced a superficially strong refit (Sharpe 1.63, max drawdown 3.8%), but the statistical gates exposed it: deflated Sharpe 0.068 against a 0.95 bar, an optimized Sharpe well below the 225-trial expected-max luck bar of 2.80, a decisively negative untouched holdout (-2.106), and a 6-of-14-cliff sensitivity surface. The analyst read this as the textbook selection-inflation trap (PSR 0.98 vs DSR 0.07) on a genuinely decaying, 2021-concentrated edge, and abandoned it after two iterations rather than re-sweeping a knife-edge surface. Later risk-review and promotion stages were never reached.

A clean, non-overfit-flagged structural pattern with positive walk-forward OOS windows can still be a mirage: a flattering low-drawdown optimized refit does not rescue a 2021-concentrated, decaying edge that fails multiple-testing deflation and a negative forward holdout.

The analyst issued an 'abandon' verdict after the first optimization attempt because the edge failed deflated Sharpe, the forward holdout, and sensitivity simultaneously — a best-of-225 selection that couldn't clear the luck bar, sat on a knife-edge cliff surface, and lost money in the recent regime.

A long-only, single-instrument ADAUSDT 4H structural strategy that armed on a sequential double-low support test (a bar's low matching a prior local low within an ATR-fraction tolerance) and entered on a bullish reclaim close above that level within a trend-permissive regime, exiting via ATR target/stop, a structural stop below support, or a time stop.

The initial 193-trade backtest showed a marginal but fee-viable edge (Sharpe 0.375, profit factor 1.11, total return 17.5%, 42% win rate, 13% max drawdown) with clear regime decay — positive 2020–2023 but negative every year since. After optimization the refit looked attractive (Sharpe 1.63, max drawdown 3.8%, PF 1.50) but failed the gates: deflated Sharpe 0.068 (vs 0.95), optimized Sharpe 1.63 far below the 225-trial luck bar of 2.80, a negative untouched holdout (Sharpe -2.106, ratio -1.343), and 6 of 14 sensitivity cliff parameters.