EthFourHourDonchianBreakoutLongShortDailyRegimeFilter

ETH Perp 4H Multi-Timeframe Trend-Following Donchian Breakout Long-Short (1D Regime Filter, ATR-Trailed)

A long-SHORT, single-instrument, single-venue, OHLCV-ONLY TREND-FOLLOWING strategy on ETHUSDT.BINANCE perpetual futures that trades 4-HOUR Donchian channel breakouts ONLY in the direction of the higher-timeframe (1-DAY) trend. This is a deliberate pivot into the one mechanism class that BOTH has documented structural edge (the trend / time-series-momentum risk premium, robust across all asset classes and especially strong in crypto) AND has demonstrably LANDED on this exact engine/data path (the BTC 4H Donchian breakout long-short was a confirmed pipeline success), in contrast to this session's graveyard of simple intraday price-pattern momentum/reversal and buy-the-dip mean-reversion strategies, ALL of which the analyst showed are empirically dead after fees on BTC OHLCV (sweep-fade −99.7%, vol-expansion-momentum −19.9%, buy-the-dip MR −43%). The differentiating, under-represented element is the MULTI-TIMEFRAME alignment: a 1D EMA regime gate permits LONG breakouts only in a daily uptrend and SHORT breakouts only in a daily downtrend, filtering out the counter-trend 4H breakouts that cause the whipsaw losses and fee drag that killed naive breakout/expansion strategies. It is long-short (attacking the portfolio's 86% long-only concentration), single-asset (structurally avoiding the multi-leg basket exposure bug that cannot be capped on futures), 4H (firing dozens of times per year so all three walk-forward OOS windows and the holdout are populated — no event-sparsity), simple (4 core parameters to resist overfitting), and uses only deep-history ETH OHLCV (no options/HL/COIN-M/cross-venue/supplementary dependencies, all confirmed unvalidatable). It avoids every dead mechanism: no mean-reversion, no single-bar pattern, no cross-sectional reversal, no calendar, no multi-leg-on-futures.

Implements the analyst-endorsed trend / time-series-momentum mechanism class that has landed on this exact engine (BTC 4H Donchian LS), pivoted to ETH and differentiated by a higher-timeframe regime gate. Single-instrument long/short via the standard FactoryStrategy hooks; the 1D regime is a slower-timeframe context leg supplied through extra_bar_types (ETHUSDT 1-DAY): because its timeframe differs from the primary 4H, the base does not treat it as a sync leg and the cross-leg alignment barrier never stalls the 4H signal, while on_extra_bar maintains the daily closes and _update_regime sets +1/-1 from EMA(fast) vs EMA(slow). extra_instruments lists ETHUSDT.BINANCE index-paired with the daily extra_bar_type so the on-demand collector fetches the daily series (and the runner also loads it from the catalog by parsing the bar-type instrument), avoiding data_unavailable. The 1D gate permits longs only in a daily uptrend and shorts only in a daily downtrend, discarding the counter-trend 4H breakouts responsible for the whipsaw and fee drag that killed naive breakout/expansion strategies. calculate_signal returns the continuous 4H channel position so Layer-2 frozen-signal detection passes, while the discrete breakout + regime decision lives in should_enter; exits use a symmetric ATR trailing stop from the position's running extreme (no asymmetric TP/SL) plus a regime-flip exit, with side/extreme reconstructed from the open position if state is lost. Sizing is capital-relative (2% risk over the 3x-ATR stop) capped at 50% notional, clearing the ETHUSDT $20 min-notional, with multi-ATR trend capture well above the ~0.10% round-trip futures fee; firing dozens of times per year populates all walk-forward windows. Long+short attacks the portfolio's long-only concentration; leverage stays 1.0 referenced only as a non-amplifying sizing multiplier so the unused-leverage gate does not apply. Four tunable parameters resist overfitting; no options/HL/COIN-M/cross-venue/supplementary dependencies.

Failed deflated Sharpe: DSR=0.4337 (vs 0.95 bar), expected-max=1.2426 over 225 trials — the selected optimized Sharpe of 1.166 is BELOW the best-of-N luck bar, so it is statistically indistinguishable from selection noise (is_significant=false), with PBO 0.6415 (>0.5) and the optimized sharpe_ci_low collapsed to 0.001 (from the base 0.45). This is the rare case where strong consistency gates (holdout PASSED at ratio 1.443 / holdout_sharpe 0.738, walk-forward is_overfitted=FALSE with avg OOS 0.511, sensitivity 0 cliffs) are overridden by the multiple-testing gate: the PSR=0.9943 / DSR=0.4337 split is the diagnostic tell of best-of-225 selection inflation. The expected-max luck bar (1.24) is high because the strategy's Sharpe varies enormously across regimes (excellent 2020-2024, then 2026 -6.4% with rolling Sharpe collapsing to -7.9), and a ~1.1-1.2 Sharpe cannot clear that bar. Not promote: DSR below 0.95, is_significant=FALSE, Sharpe below the luck bar, and PBO>0.5 are each disqualifying, and a passing holdout does not override a deflated-Sharpe failure (the holdout window predates the 2026 decay). Not iterate (attempt 1 of 2): the sweep was thorough, not under-explored — sensitivity is flat with 0 cliffs and PBO 0.64 indicates overfit selection, with the best Sharpe already below the luck bar; the luck bar is set by intrinsic cross-regime return variance (one OOS window -0.23, another +1.50) that parameter clamping cannot reduce, so a second sweep re-finds configs below the same bar. Not revise_hypothesis: the Donchian-breakout mechanism is already validated/landed on a sibling; this is the same mechanism on the named asset failing the multiple-testing gate, not a sound mechanism stranded on a dead target. FAILURE PATTERN: even the best-constructed trend-following breakout of the session (multi-timeframe regime-gated, long-short, non-overfit walk-forward, passing holdout, 0 cliffs) can still fail deflated Sharpe when the strategy's large cross-regime Sharpe variance sets an expected-max luck bar (1.24) above the honest selected Sharpe (1.166) over a 225-trial search — a modest, regime-variable trend edge on a single alt cannot clear best-of-N deflation regardless of how clean its consistency gates are. The PSR=0.99 / DSR=0.43 split, the Sharpe below the luck bar, and PBO 0.64 are the decisive tells; passing holdout and non-overfit walk-forward measure forward consistency, not multiple-testing significance.

Long/short multi-timeframe trend-following on ETHUSDT.BINANCE USD-M futures: 4-hour Donchian channel breakouts taken only in the direction of the 1-day EMA trend regime. When the daily EMA(20) is above EMA(50) it goes long on a 4H close above the prior 20-bar high; when EMA(20) is below EMA(50) it goes short on a close below the prior 20-bar low. It exits on a symmetric 3x-ATR trailing stop (trailing the position's running extreme) or a daily regime flip against the position. The daily regime is supplied via a slower-timeframe extra bar leg, so it filters direction without blocking the 4H signal. Risk-based sizing (2% equity risk over the ATR stop) capped at 50% notional, leverage 1.0, OHLCV-only. Four tunable parameters: donchian_period, daily_ema_fast, daily_ema_slow, atr_stop_mult.

Positive CI lower bound: Sharpe 1.02 with sharpe_ci_low 0.4501 (does NOT straddle 0) — the strongest breakout base of the session; PSR 0.9998

NO recent decay (unlike all abandoned breakout arms): positive in 6/7 years with 2024 +36.9%, 2025 +37.5%, 2026 +36.8% — the 1D regime gate + long-short filters counter-trend whipsaws

Strong alpha (+0.217), PF 1.72, max_DD 15.6%, 393 trades, capacity $216M; landed mechanism class (BTC 4H Donchian LS)

Fat tails are legitimate trend-following convexity (skew +8.4, distributed across many years), not one-outlier artifact

High return_kurtosis 109.8 — must be stress-tested by the deflated-Sharpe gate post-opt

Absolute Sharpe 1.02 comes with high annualized_volatility (29.9%); 2023 was -2.0%

end_unrealized_pct 63.9 (~8% of the +768% headline is open MTM) — minor, judge on realized track record

Closest-to-promotable of the session: holdout PASSED (ratio 1.443, holdout_sharpe 0.738), walk-forward is_overfitted=FALSE (avg OOS 0.511, 2 of 3 windows positive), sensitivity PASSED with 0 cliffs

Default config is genuinely significant (base sharpe_ci_low 0.45); deep capacity ($106M), correct trend-following payoff

Failed deflated Sharpe: DSR 0.4337 (<0.95), is_significant=FALSE; optimized Sharpe 1.166 is BELOW the expected-max luck bar 1.2426 over 225 trials — indistinguishable from best-of-N noise

PBO 0.6415 (>0.5, overfit-leaning); optimized sharpe_ci_low 0.001 (collapsed from the base 0.45 — the selected config is barely significant)

PSR 0.9943 / DSR 0.4337 split is the diagnostic tell of selection inflation; high luck bar is driven by cross-regime variance (2026 -6.4%, rolling Sharpe -7.9) that parameter clamping can't fix

max_consecutive_losses 15; recent decay despite the passing (older-window) holdout

EthFourHourDonchianBreakoutLongShortDailyRegimeFilter was a deliberate pivot to the one mechanism with documented structural edge and a landed sibling (BTC 4H Donchian long-short), adding a 1-day EMA regime gate to filter the counter-trend whipsaws that killed the session's naive breakout and mean-reversion strategies. It produced the strongest base of the session — Sharpe 1.02 with a positive CI lower bound, profit factor 1.72, strong alpha, deep capacity, and crucially no recent decay — and then cleared the consistency gates in optimization (passing holdout, non-overfit walk-forward, zero sensitivity cliffs). But it failed the deflated-Sharpe gate: its optimized Sharpe of 1.166 fell below the 1.243 best-of-225 luck bar, making it statistically indistinguishable from selection noise, with PBO 0.64 and a collapsed optimized CI lower bound confirming the inflation. The analyst abandoned it on its first iteration, concluding that a modest, regime-variable trend edge on a single alt cannot clear best-of-N deflation no matter how clean its consistency gates, and that a second sweep would re-find configs below the same bar.

Even the cleanest trend-following breakout — multi-timeframe regime-gated, long-short, non-overfit walk-forward, passing holdout, zero cliffs — can fail deflated Sharpe when its large cross-regime Sharpe variance raises the expected-max luck bar above the honest selected Sharpe over a 225-trial search; passing consistency gates measure forward consistency, not multiple-testing significance, and a passing (older-window) holdout does not override a DSR failure.

The analyst issued an 'abandon' verdict at the post-optimization ANALYZING stage: despite passing the consistency gates, it failed the deflated-Sharpe multiple-testing gate (DSR 0.4337 < 0.95, is_significant=FALSE) because the optimized Sharpe 1.166 sat below the expected-max best-of-225 luck bar of 1.2426, with PBO 0.6415 (>0.5) and the optimized sharpe_ci_low collapsing to 0.001 — the PSR-0.99/DSR-0.43 split being the diagnostic tell of selection inflation.

A long-short, single-instrument, OHLCV-only trend-following strategy on ETHUSDT.BINANCE USD-M futures that traded 4H Donchian channel breakouts only in the direction of a higher-timeframe 1-day EMA regime (long breakouts in a daily uptrend, short breakouts in a daily downtrend), exited via a symmetric ATR trailing stop or a daily regime flip, using four core parameters to capture the time-series-momentum risk premium while filtering counter-trend whipsaws.

The base backtest was the strongest of the session: Sharpe 1.02 with sharpe_ci_low 0.45 (not straddling zero), profit factor 1.72, alpha +0.217, 15.6% max drawdown, 393 trades, $216M capacity, positive in 6/7 years with no recent decay (2024 +36.9%, 2025 +37.5%, 2026 +36.8%); optimization passed the consistency gates — holdout PASSED (ratio 1.443, holdout Sharpe 0.738), walk-forward is_overfitted=FALSE (avg OOS 0.511), sensitivity 0 cliffs — but the optimized Sharpe was only 1.166.