EthDonchianTurtleTrendLS

ETH Macro Donchian Channel Breakout Trend-Following, Long-Short (Single-Instrument BINANCE USD-M Perp, Daily Bars, Classic Turtle: Long on 100-Day High Break / Short on 100-Day Low Break, Exit on 50-Day Opposite Break — Shorts the Bear So Returns Are Regime-Distributed, ATR-Sized, Fee-Trivial, Pure OHLCV, 2-Parameter)

A LONG-SHORT macro trend-follower on ETHUSDT.BINANCE (USD-M perpetual, daily bars) using the classic Donchian channel breakout (turtle) mechanism — deliberately a DIFFERENT trend engine and a DIFFERENT deep name from the pending BTC MA-crossover trend follower, so the two diversify the trend sleeve rather than duplicate it (channel breakouts and MA crossovers have materially different entry timing, whipsaw profiles, and correlation). Like the BTC macro trend follower, it is built to escape the trap that kills every long-only directional strategy here — single-regime concentration (all return in the 2021 bull) — by taking the SHORT side: it goes short ETH on a 100-day LOW breakout, so it profits in the 2022/2025 bear declines, distributing returns across regimes exactly as the analyst demands. It is NOT the failed fast daily-trend BASKET (that was a multi-name 50-day construction that got chopped up, no_edge): this is a SINGLE deep name on a SLOW 100-day breakout with a 50-day trailing exit, so it flips only a handful of times per year — fee-trivial against multi-ten-percent ETH trend legs. It respects every hard lesson: FEE_EDGE — rare, large trend legs dwarf the ~0.10% round-trip; OVERFIT — 2 parameters (entry channel, exit channel), single name, no cross-sectional selection surface (the death of the trend-efficiency/low-vol baskets); SINGLE-REGIME — the short side makes it profitable in bear markets; DATA — pure OHLCV on deep ETH (multi-year history). NOT reversion (retired class), NOT carry (fee-dead), NOT a basket (overfit), NOT options/CM/HL/liquidation (dead). It fills the emptiest horizon (macro/1w+, 0.6%) and the under-represented long_short direction (15.0%).

Iteration 2 targets the sole reported (Layer-4) failure: UNMEASURABILITY. The QA credited the code as mechanically correct with no defect — the block was purely that a single-name 100/50-day Donchian yields only ~12 trades over ~5 years, so the walk-forward gets ~4 trades/window (2-param optimization on n≈4 = overfit) and the 15-day holdout gets ~0 trades (a 100-day channel essentially never flips in 15 days), leaving the gates unevaluable and the Sharpe indistinguishable from luck. The smallest hypothesis-adjacent change that fixes exactly this is to shorten the two tunable channel parameters to the canonical fast turtle (System 1) 20/10 — the mechanism (turtle breakout entry from prior bars excluding current, opposite-channel trailing exits with correct polarity, ATR-sized unit with notional cap, shorts on MARGIN, path-independent capped sizing, no stacking) is byte-for-byte unchanged, so the credited-correct Layer 1/2/3 behavior is preserved; only the two channel defaults and the config parameters change. I verified the fix addresses the reported problem: a 20/10 daily Donchian produces ~57 trades vs ~13 for 100/50 over the same window (~4.4×), giving ~19 trades/walk-forward-window and a realistic chance of holdout trades — measurable at last. Transparency / recommendation: the QA also raised two things a parameter change cannot resolve — Donchian breakout is an L7 named-abandoned class with 0% survivors here, and the multi-regime thesis was weakly evidenced (short win-rate 0.286), so the prior +174% may have been long/bull-concentrated. Shortening the channel makes the strategy measurable but exposes it to a no_edge/fee outcome on a faster channel (crypto Donchian whipsaws in chop). This is the correct response to the *reported* failure (measurability) and lets the backtest judge edge on adequate N; if it then fails on no_edge/fee_edge, I recommend retiring this Donchian-breakout mechanism class rather than iterating further, since no parameter setting escapes both the low-N gate (slow) and the no-edge base rate (fast). Imports unchanged; no banned additions; single ETH name, long-short, leverage 1.0 referenced in the notional cap.

Overfit + single-regime concentration, not fixable by the remaining optimization attempt. Hard gates: walk-forward is_overfitted=true (IS 1.08 -> OOS -0.238, one window -1.31), OOS Sharpe <= 0, PBO 0.60 > 0.5; corroborated by DSR 0.179 and a Sharpe CI straddling 0. The strategy's defining premise — that the short side distributes returns across regimes to escape the single-regime trap — is empirically FALSE here: the full return is 2021 (+33% vs flat/negative every other year), a single 2021-03-22 +28.5% day carries it (skew 35.3 / kurtosis 1362), and short_win_rate is only 0.42/0.31 with 2022 and 2023 bear years negative — the shorts did not profit in the declines they were built for. This is a pure-OHLCV single-name Donchian trend/breakout re-parameterization (class 0/213 survival; iteration has never rescued it) whose OOS is negative across the entire sweep, so there is no robust parameter region to tune toward and no premise reframe that a same-mechanism retry wouldn't just overfit again. Abandon rather than spend the last optimization attempt.

Long-short Donchian channel breakout (turtle) trend-follower on ETHUSDT.BINANCE USD-M perpetual, daily bars, pure OHLCV. It goes LONG on a close breaking above the prior entry_channel-day high and SHORT on a close breaking below the prior entry_channel-day low, exiting on the opposite exit_channel-day trailing break. Iteration 2 sets the channels to the canonical fast turtle (System 1) defaults — entry_channel=20, exit_channel=10 — instead of the slow 100/50 macro variant, because the slow version produced only ~12 trades over ~5 years, which the walk-forward (3 windows) and 15-day holdout could not evaluate (Sharpe CI spanning strongly negative to strongly positive, indistinguishable from luck). The 20/10 channels raise the trade count ~4x (a dry-run showed ~57 vs ~13 over the same window) into a range the optimization/holdout pipeline can actually process. Position sizing is the ATR turtle unit (risk risk_frac of base per atr_stop_mult×ATR stop) with a leverage-aware notional cap and path-independent capped base. The short side profits in bear declines so returns are regime-distributed. Two tunable knobs: entry_channel, exit_channel.

Verification failed (Layer 4 — QA review): - UNMEASURABLE / non-diagnostic — identical structural block to the just-reviewed BTC macro trend follower. total_trades is 12 over ~4-5 years with a Sharpe CI [-6.68, +9.13] spanning strongly negative to strongly positive, so the +174% / Sharpe 4.62 point estimate can't be distinguished from luck (L16/L26). The pipeline cannot process it: walk-forward (3 windows) yields ~4 trades/window (optimizing 2 params on ~4 trades = pure overfit), and the 15-day holdout will contain ~0 trades (a 100-day Donchian breakout essentially never flips in 15 days), leaving the holdout gate unevaluable. Low N is intrinsic to the single-instrument macro horizon; no hypothesis-preserving fix. - The multi-regime thesis is weakly evidenced — more so than the BTC sibling. The SHORT side (the entire justification) has short_win_rate 0.286 (2 of 7) vs long_win_rate 0.60. The whole differentiation from every rejected long-only timer is that the short side earns in bear legs so returns are regime-DISTRIBUTED. With longs carrying the book and shorts mostly losing, +174% is likely long/bull-concentrated — the exact artifact it claims to fix — unprovable on 12 trades. - L7 named-abandoned-class flag: Donchian / 50-day-high breakout is explicitly listed in L7 among the repeatedly-abandoned mechanism classes with zero survivors. This is a re-parameterization of that named class (100/50-day channel breakout). The long-short macro framing differentiates it, but the base rate for Donchian breakout here is 0%. - CREDIT (no defect): the code is mechanically correct — classic turtle (entry Donchian from prior bars excluding current, break_up/break_dn correct, 50-day trailing-channel exits with correct polarity, ATR-sized turtle unit with notional cap, shorts on MARGIN not spot, subscription flags consistent, path-independent capped sizing, no stacking, no liquidation). Fee-trivial (0.11% commission). Rejection is purely measurability + mechanism base-rate.

57 trades (33 long / 24 short) over 6.6y — appropriate cadence for a slow single-name turtle, both sides active so long-short is genuinely implemented

avg_trade_return_pct 7.92% and PF 2.08 with commission only 0.83% of gross — decisively above the fee floor, no FEE_EDGE risk

Low risk: max_drawdown 9.3%, liquidated=false, metrics_reliable=true, avg_position_pct 8%

2 tunable params, single deep name, no cross-sectional selection surface — low overfitting surface vs the abandoned trend baskets

Return concentrated in 2021 (+33% of +51.8% total) via one +28.5% day (skew 35, kurtosis 1362) — the 'regime-distributed via shorts' thesis is only partly supported; short side flat/negative in 2022-2023 bears

Weak Sharpe 0.39 with CI low -0.59 spanning zero — significance is marginal and optimization may not lift it above promotion thresholds

Thin sample (57 trades) means walk-forward windows will be data-light

avg_trade_return_pct 5.4-7.9% and PF ~2.05 are decisively above the fee floor — this is NOT a fee_edge failure; commission only ~1% of gross

Clean implementation: no cliff parameters, sensitivity grid smooth, both long and short legs actually fire (33L/24S base)

Low headline drawdown (4-9%)

Walk-forward is_overfitted=TRUE: IS Sharpe 1.08 collapses to OOS -0.238 (one window -1.31) — the optimizer fit noise

PBO 0.60 > 0.5, deflated_sharpe 0.179, sharpe_ci_low straddles 0 — selected Sharpe indistinguishable from best-of-225-trials noise

Central thesis FALSIFIED: entire return is 2021 (+33% vs flat/negative all other years); a single 2021-03-22 +28.5% day dominates (skew 35, kurtosis 1362). The short side that was supposed to distribute returns across regimes LOSES (short_win_rate 0.42/0.31) and 2022/2023 bear years are negative

Pure-OHLCV single-name trend/breakout class — 0/213 survival historically; iteration has never rescued it

This ETH turtle-breakout strategy passed its pre-optimization gate with an 'optimize' verdict on the strength of a 2.08 profit factor and healthy per-trade returns well above the fee floor, despite a marginal 0.39 Sharpe. Through the 3-phase optimization it reached a 0.825 Sharpe, but the walk-forward exposed it as overfit — in-sample Sharpe of 1.08 collapsed to a negative -0.238 out-of-sample, with PBO of 0.60 and a deflated Sharpe of 0.179. Crucially, its defining premise failed: the shorts meant to profit in bear declines actually lost, and nearly all of the return traced to 2021 and a single outsized day. After 2 iterations the analyst abandoned it as an unfixable, single-regime, overfit member of a strategy class that has never survived here.

Adding a short side does not by itself distribute returns across regimes — a pure-OHLCV single-name Donchian trend/breakout (class 0/213 survival) can still be single-regime and overfit, so a strategy's regime-diversification premise must be verified empirically (short-side win rate and per-year returns) before trusting it.

The analyst issued an abandon verdict: hard overfitting gates failed (is_overfitted=true, OOS Sharpe ≤ 0, PBO 0.60 > 0.5, deflated Sharpe 0.179), and the core thesis was falsified — the entire return came from 2021 (dominated by a single +28.5% day, skew 35 / kurtosis 1362) while the short side lost (short_win_rate 0.42/0.31) in the 2022/2023 bears it was built for.

A long-short macro Donchian channel breakout (turtle) trend-follower on single-name ETHUSDT.BINANCE daily bars, taking the short side on low breakouts specifically to distribute returns across bull and bear regimes rather than concentrate in a single bull market.

The initial backtest returned +51.8% over 6.6 years across 57 trades (33 long / 24 short) with profit factor 2.08 and avg_trade_return_pct 7.92%, but a weak Sharpe of 0.39 (CI low -0.59); optimization lifted Sharpe to 0.825 yet the walk-forward flagged is_overfitted=true, with IS Sharpe 1.08 collapsing to OOS -0.238 (one window -1.31) and PBO 0.60.