EthVolatilitySqueezeBreakoutLS

ETH Volatility-Squeeze Breakout Momentum — Long-Short, Single-Name: Buy Volatility at Its Cheapest by Entering the Directional Break Out of a Bollinger-Width SQUEEZE (Bottom-Percentile Consolidation), Ride the Expansion with an ATR Trailing Stop, Flat When Volatility Re-Contracts (ETHUSDT.BINANCE USD-M, 4H, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV TREND-CAPTURE strategy whose entry TRIGGER is a volatility-REGIME transition, not a momentum crossover — deliberately in the factory's ONE demonstrated survivor family (directional trend, per the BTC dual-timeframe winner) but with a materially different construction that avoids the dead momentum-confluence template (L62) and every reversal/carry/positioning family that keeps dying. Mechanism: volatility in crypto clusters, and a deep contraction (Bollinger-band width in its bottom percentile = a 'squeeze') reliably PRECEDES a large directional expansion. We do nothing during the squeeze, then enter in the direction of the range break when volatility releases, and ride the expansion with an ATR trailing stop. This structurally attacks the #1 killer (fee_edge): squeeze-release moves on ETH run multiple percent, so per-trade capture is an order of magnitude above the ~0.10% round-trip cost, and squeezes are occasional so turnover — and fee drag — stay low. It is NOT the cross-sectional reversal that just died (opposite sign, single-name, trend not reversion), NOT a dual-timeframe confluence clone (single timeframe, volatility trigger), NOT OI/funding/positioning driven, and needs no supplementary feed (pure OHLCV, full ETH history, data-safe).

Iteration 3 targets the ONE exit path that exists. I read the base template: `_process_primary` calls `should_exit` and, on True, `_submit_exit` -> `close_position`; nothing else in the framework ever closes a position, and `_submit_entry` attaches no bracket. So a ~1.25-bar average hold can only be produced by (a) an exit branch firing right after entry, or (b) an exit followed by an immediate re-entry on the next bar, which chains one breakout into many one-bar round trips. Iteration 2 tightened (a) branch-by-branch and the holding time did not move, which points at (b) — my should_enter re-fired on every bar the score merely REMAINED above threshold, so each exit was instantly followed by a new position on the same squeeze. This iteration closes (b) with two independent brakes: entry now requires the score to CROSS the threshold (self._prev_score, captured at the top of calculate_signal, must be below it) and a reentry_cooldown_bars=6 window after any exit. It also makes (a) impossible instead of improbable: min_hold_bars is now the FIRST test in should_exit and short-circuits the trail, the re-contraction exit and the time stop alike — one guard instead of three separate per-branch grace windows that each had to be correct — with only a hard_stop_pct=6% catastrophic stop allowed to bypass it, so bounded per-trade risk survives the longer hold. Both changes also cut turnover (the crossing rule and cooldown must reduce the 158-trade count), which directly attacks the fee-viability failure: fewer trades, each held >= 24h through a squeeze expansion, is the only geometry in which per-trade capture clears the ~0.10% round trip. The signal computation, imports and class structure that passed Layers 1-3 are untouched; every change is confined to should_enter, should_exit and the sizing stop basis. One honest note: if the next sandbox again reports ~1-bar holds, that result is inconsistent with this code — min_hold_bars mathematically forbids any exit before 6 bars unless price moves 6% against the entry — and the run should be checked for a stale/cached strategy build before the strategy itself is blamed.

Dead-flat single-name breakout whose iteration-3 fix did not take and which is capacity-dead — not worth 2 hours of optimization. Over 463 trades: Sharpe 0.012 with bootstrap CI [-0.72, 0.71] straddling zero, profit_factor 1.019, PSR 0.507, total_return +2.36% over 6.7 years (CAGR 0.26%), expectancy +$3.77/trade — no measurable edge. The iteration-3 minimum-hold/crossing/cooldown changes were meant to end the ~1.25-bar scalp holds, but avg_holding_period is still ~4h26m (about one 4H bar), so the mechanism is still scalping, not capturing trends. It is fee-fragile (avg_trade_return_pct 0.154% sits exactly on the 0.15% USD-M floor while commission is 17.9% of gross) and capacity-dead (impact_cost_pct 74.6%, capacity_usd only ~$180k). This is the L30/L63 single-name channel/squeeze breakout graveyard, and the recent regime is negative (2026 -2.5%, rolling Sharpe deeply negative through 2026), so the holdout would be weak-to-negative. No parameter change turns a PF-1.02, zero-Sharpe, one-bar-hold, fee-floor breakout into an edge, and it is untradeable at scale regardless. Failure pattern: no_edge/fee_edge/capacity single-name squeeze breakout, churn fix ineffective (L30/L63).

Long/short ETHUSDT USD-M 4H trend-capture. Every bar it scores z * comp, where z = (close - Bollinger mid)/sd on the prior-bar band (how far price broke out) and comp = 1 - percentile rank of the tightest band width in the last few bars within a 240-bar width history (how compressed the consolidation was just before). Entry fires only when |score| CROSSES entry_score (previous bar below it) and no exit occurred within the last reentry_cooldown_bars, so one squeeze release yields one trade. Once open, the position is held for a hard minimum of min_hold_bars (24h) — only a hard_stop_pct catastrophic stop can close it sooner — after which it is managed by an ATR trailing stop sized to post-breakout volatility (max of current ATR, the breakout bar's true range, 2.5% of price), a volatility re-contraction exit when band width falls into its bottom quartile, and a 120-bar time stop. Sizing risks 2% of equity across the effective stop distance, capped at 50% of equity notional.

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Verify realized average holding period in the full backtest; if genuinely ~1 bar, check for a stale/cached strategy build before faulting the code.

Sandbox reports avg_holding_bars=0.0 and avg_holding_period='4h 35m' (~1 bar), which contradicts min_hold_bars=6 (24h) being the first, short-circuiting test in should_exit. The holding metric in the report is unreliable; the code enforces the minimum hold correctly on inspection.

Analyst to confirm net-of-impact per-trade edge and capacity at intended deployment capital.

impact_cost_pct=37.9%, turnover=53.7, capacity_usd~$697k in the sandbox indicate meaningful impact/fee drag despite the low-turnover design intent.

Differentiated construction (volatility-regime trigger, not momentum crossover); clean incremental implementation; contained drawdown (15.8%)

No edge: Sharpe 0.012 with CI [-0.72, 0.71] straddling zero, PF 1.019, PSR 0.507, total_return +2.36% over 6.7 years (CAGR 0.26%)

Iteration-3 hold fix failed: avg_holding_period still ~4h26m (~1 bar) — still scalp geometry, not trend capture

Fee-fragile: avg_trade_return_pct 0.154% on the 0.15% floor, commission 17.9% of gross

Capacity-dead: impact_cost_pct 74.6%, capacity_usd only ~$180k; L30/L63 single-name breakout graveyard, recent regime negative (2026 -2.5%)

Verification failed (Layer 4 — QA review): - THE CORE MECHANIC ('ride the expansion with an ATR trailing stop') IS NOT REALIZED. The trail distance is atr_trail_mult * ATR where ATR is a lagging Wilder(14) average. During the squeeze ATR is compressed to tiny values; on the breakout it moves only ~1/14 toward the large true range, so at ENTRY the trail is calibrated to PRE-breakout volatility and far too tight for the POST-breakout expansion just entered. should_exit also bumps self._peak to the current bar's HIGH before comparing to close, so one wide expansion bar with an upper wick trips 'close <= peak - 3*ATR(compressed)' and exits the next bar. Sandbox confirms from three angles: avg_holding_period '4h 42m' (~1.2 bars vs the tens-of-bars/2.3-day holds the hypothesis claims); avg_win 406 ~= avg_loss 334 (near-symmetric, NOT trend-capture geometry); avg_trade_return_pct 0.023%, PF 0.81, total_return -6.2%. The strategy chokes off every trade ~1 bar in and never rides an expansion, so the implemented behaviour does not match the stated strategy. - FEE-VIABILITY FAILURE, CONFIRMED IN THE SANDBOX (checklist #7). Realized avg_trade_return_pct 0.023% is ~1/6 of the 0.15% floor and below the ~0.10% round-trip cost, so the strategy is a net loser (total_return -6.2%, Sharpe -0.54, PF 0.81, expectancy -39/trade over 158 trades) -- the direct consequence of the ~1-bar holds. The developer's fee-viability defence was an engine-free replay claiming +0.535%/trade and 2.3-day holds; the real engine delivered +0.023%/trade and ~1-bar holds, so that justification is contradicted by the actual backtest.

Verification failed (Layer 4 — QA review): - THE ITERATION-2 FIX DID NOT CHANGE TRADE TIMING, SO THE CORE MECHANIC IS STILL NOT REALIZED. The sandbox is near-identical to the failed iteration 1: total_trades 158 vs 158, win_rate 0.3987341772 vs 0.3987341772 (byte-identical), long/short 85/73 vs 85/73, avg_holding_period 5h 0m (~1.25 bars) vs 4h 42m. Only the PnL magnitudes shrank (avg_win 406->262, avg_loss 334->212, avg_position_pct 41%->26%) because sizing widened the stop and cut quantity. Entries/exits fire on the same bars as before. Decisive proof: the new code suppresses the trail for held < trail_grace_bars (3) and re-contraction for held < 8, so should_exit CANNOT close before ~bar 3 -- yet holds are still ~1.25 bars. The binding exit is NOT the rewritten trail; positions close ~1 bar in via another path (almost certainly opposite-signal flips after the band recenters post-breakout). 'Ride the expansion for tens of bars' is still not what the strategy does. - FEE-VIABILITY FAILURE, STILL PRESENT (checklist #7). avg_trade_return_pct 0.0178% is ~1/9 of the 0.15% floor and below the ~0.10% round-trip cost; the strategy stays a net loser (total_return -3.69%, Sharpe -0.53, PF 0.82, expectancy -23/trade over 158 trades). avg_win 262 ~= avg_loss 212 is the geometry of ~1-bar scalps, not trend capture. The developer's own stated success criterion for this iteration (convert ~1.2-bar holds into multi-day right-skewed geometry lifting capture over the floor) is unmet by this submission's sandbox.

Verification failed (Layer 4 — QA review): - SANDBOX BEHAVIOUR IS INCONSISTENT WITH THE SUBMITTED CODE, SO 'RIDE THE EXPANSION' CANNOT BE VERIFIED. min_hold_bars=6 is the first test in should_exit and short-circuits the trail, re-contraction and time-stop branches, so the ONLY exit permitted before 6 bars (24h) is the hard_stop_pct=6% catastrophic stop. Yet avg_holding_period is 6h 31m (~1.6 bars on 4H). The entry changes are demonstrably live (158 -> 65 trades, which only the new crossing+cooldown can produce), so the same-file min-hold is live too. For a ~1.6-bar average under a live 6-bar minimum, the -6% catastrophic stop must fire on the large majority of trades within ~1 bar -- not credible (would need ~80% of breakouts to move -6% in a bar or two) or the metric doesn't reflect the build. If the catastrophic stop really dominates, the entry is systematically catching FAILED breakouts that reverse hard -- a mechanism failure. Either way the hypothesis's core 'hold then ride' mechanic is not demonstrably realized; the developer's own flagged check reproduced. - NO EDGE, AND WORSE THAN BEFORE. The churn fix worked (158 -> 65 trades, one trade per squeeze), which makes the result informative: with the scalp churn removed, the squeeze-breakout premise shows negative expectancy. PF 0.57 (down from 0.82), Sharpe -0.90 (down from -0.53), total_return -5.24%, win_rate 0.34, avg_trade_return_pct -0.122% (now outright negative). Forcing the longer minimum hold converted small scalp losses into larger trend-fighting losses -- squeeze releases on ETHUSDT 4H do not reliably precede a ridable expansion net of fees. Three iterations, monotonically worse (Sharpe -0.53, -0.53, -0.90).