Strategies

ETH Daily Donchian Channel Breakout Long-Only (BINANCE USD-M Futures, Daily, OHLCV-Only)

Outcome: Abandoned

EthDailyDonchianBreakoutLong

Outcome Summary

EthDailyDonchianBreakoutLong ported the validated BTC daily Donchian breakout mechanism onto ETH, cleanly implementing a long-only Turtle breakout that produced 200% total return over 36 trades and earned an 'optimize' verdict at backtest review. Optimization sharpened the in-sample picture — Sharpe 0.64, profit factor 2.11, no sensitivity cliffs, and a non-overfit walk-forward averaging 0.55 OOS — but the forward gates broke it. The holdout collapsed to a negative Sharpe of -2.20 (ratio -4.011), the deflated Sharpe of 0.32 fell below the best-of-225 luck bar, and PBO 0.68 flagged overfitting, all traceable to an edge concentrated in 2021 and decaying to negative in 2022-2026. The analyst abandoned it on its first iteration, noting the same bull-concentrated, holdout-failing signature seen in earlier AVAX/DOGE/ADA breakout siblings.

Strategy report

Hypothesis

A long-only, single-instrument, single-venue, OHLCV-ONLY breakout strategy on ETHUSDT.BINANCE 1D bars. Same template as landed BtcDailyDonchianBreakoutLong — applied to ETH. Enter long when close breaks above the 20-day Donchian channel high; exit when close drops below the 10-day Donchian channel low. Classic Turtle-trader mechanism. Per factory directive: ETHUSDT.BINANCE 1D bars are factory-approved. Three parameters total.

Direct application of the landed BtcDailyDonchianBreakoutLong template to ETH, as the hypothesis specifies. The entry fires only on a STRICT new-high breach of the prior 20-day high (look-ahead-safe prior-window slicing), capturing breakout momentum from consolidation ranges; the exit uses the classic 10-day Turtle trailing low plus a hard percentage stop. calculate_signal returns 1.0 on the discrete breakout and otherwise a continuous tanh-damped proxy of the close-vs-Donchian-high gap so the signal varies bar-to-bar (passing the Layer-2 frozen-signal detector) while the binary entry is gated in should_enter. Routed to the BINANCE USD-M futures venue per the hypothesis (factory-approved ETHUSDT.BINANCE 1D bars); leverage is 1.0 and the logic is long-only. No supplementary, cross-asset, or calendar logic.

Failed deflated Sharpe: DSR=0.3221 (vs 0.95 bar), expected-max=0.8352 over 225 trials — the optimized Sharpe of 0.6355 is BELOW the best-of-N luck bar, statistically indistinguishable from selection noise (is_significant=false). The decisive forward gate is the HOLDOUT, which FAILED catastrophically with a NEGATIVE Sharpe -2.2046 (ratio -4.011 vs WF-OOS 0.55) — the recent one-shot forward window loses money badly. PBO 0.6812 (>0.5) and sharpe_ci_low -0.799 (CI straddles 0) confirm the selection is overfit. Although sensitivity PASSED (0 cliffs) and walk-forward is_overfitted=FALSE with all 3 OOS windows positive, those measure in-sample smoothness/selection, not forward generalization — and they cannot override a negative holdout and a DSR below the luck bar. Root cause is bull-concentration with recent decay: the entire edge lives in 2021 (+164% base / +82% optimized) while every other year is flat-to-negative (2022 -10.8%, 2023 -5.7%, 2025 -2.6%, optimized 2026 -12.0%), with the headline still leaning on fat-tail pump days (2021-03-22 +87% single day; return_kurtosis 656 base / 261 optimized, skew 24) — which is exactly why the time-ordered holdout collapses. Not iterate (attempt 1 of 2): sensitivity is already clean (0 cliffs), so the optimizer is on its robust plateau with no under-explored region to tune toward; the failure is bull-concentration + sub-luck-bar deflation + a negative recent holdout, so a second best-of-225 sweep re-selects the same 2021-carried config that fails the 2026 forward window. Not revise_hypothesis: the Donchian breakout mechanism is already validated and landed on BTC (the cited sibling); ETH is the bull-concentrated/decaying arm of that family's asset-selectivity, not a proven mechanism stranded on a dead target. FAILURE PATTERN: long-only daily Donchian/Turtle breakout on a crypto major (ETH) optimizes to a smooth, cliff-free sensitivity grid with a positive walk-forward average (avg OOS 0.55, is_overfitted=FALSE), yet still fails deflated Sharpe (0.32, optimized Sharpe 0.64 below an 0.835 luck bar) and a deeply NEGATIVE holdout (-2.20) because the edge is concentrated in 2021 and decays to negative in 2025-2026. This is the same bull-concentrated, holdout-failing, fat-tail-carried signature as the abandoned AVAX/DOGE/ADA volume-and-Donchian breakout siblings — a clean sensitivity surface and a positive in-sample walk-forward do NOT rescue a breakout edge whose time-ordered forward (holdout) window loses money.

Implementation

Long-only single-instrument Donchian channel breakout on ETHUSDT.BINANCE (USD-M perp) 1-DAY bars. Classic Turtle-trader mechanism: enter long when the close breaks strictly above the prior 20-day Donchian channel high; exit when the close drops strictly below the prior 10-day Donchian channel low, with a hard 8% stop-loss safety net. OHLCV-only, no supplementary data, leverage 1.0. Three core parameters (entry_window, exit_window, stop_pct). Positions sized at 40% of account equity.

Backtest Review

Sharpe: 0.34
Total return: 200.20%
Max drawdown: 23.91%
Trades: 36
Win rate: 38.9%
Profit factor: 1.43

Trades implement the hypothesis exactly: 36 long entries, 0 shorts, 36 submitted with 0 dropped (min-notional/size) — clean long-only Donchian breakout

Strategy is functional with a real trend-follower profile: PF 1.43, Sortino 2.08, Calmar 3.43, win rate 38.9% (few big wins, many small losses — expected for Turtle)

36 trades over 6.3 years is adequate sample for daily-bar parameter optimization

Same mechanism as the landed BtcDailyDonchianBreakoutLong sibling — family validated on a liquid major; ETH is a reasonable next instrument

Returns heavily front-loaded into 2021 (+164.5%); most other years negative (2020 -11.1%, 2022 -10.8%, 2023 -5.7%, 2025 -2.6%, 2026 -12.4%) — recent decay risk

Headline carried by extreme tails (kurtosis 656, skew 23.9, single +87% day); Sharpe only 0.34 with sharpe_ci_low -1.45 (CI straddles 0)

max_drawdown 23.9% with a 1078-day drawdown duration — long underwater stretches

long-only breakout, 0 shorts

Analysis

Sensitivity PASSED with 0 cliffs — parameter surface is smooth

Walk-forward is_overfitted=FALSE with all 3 OOS windows positive (avg 0.55)

Optimized PF 2.11, max_DD 17.1%, capacity ~$2.2B — clean-looking in-sample profile

Deflated Sharpe 0.3221 (is_significant FALSE); optimized Sharpe 0.6355 is BELOW the expected-max luck bar of 0.8352 over 225 trials — indistinguishable from selection noise

Holdout FAILED catastrophically: ratio -4.011, holdout Sharpe -2.20 (NEGATIVE) vs WF-OOS 0.55 — recent forward window loses money badly

PBO 0.6812 (>0.5) and sharpe_ci_low -0.799 (CI straddles 0)

Bull-concentrated and decaying: edge lives in 2021 (+82% optimized) with 2022/2023/2025 flat-to-negative and 2026 -12.0%

Fat-tail carried: return_kurtosis 261 (656 base), skew 14.5, one +87% pump day (2021-03-22)

Outcome Summary

A smooth, cliff-free sensitivity surface and a positive in-sample walk-forward do not rescue a breakout edge whose returns are concentrated in one bull year (2021) and decay to negative in the recent time-ordered holdout window.

The analyst abandoned it after optimization: the holdout failed catastrophically (ratio -4.011, Sharpe -2.20 vs walk-forward OOS 0.55), the deflated Sharpe of 0.32 sat far below the 0.95 bar with the optimized Sharpe 0.64 below the 0.835 best-of-225 luck bar, and PBO 0.68 confirmed overfit — none of which a clean sensitivity grid or positive walk-forward could override.

A long-only Turtle-style Donchian channel breakout on ETHUSDT daily futures bars — enter long when the close breaks above the 20-day channel high, exit on a drop below the 10-day channel low or a hard stop, reusing the template from the landed BTC sibling strategy.

The initial backtest returned 200.2% total (CAGR 11%) over 36 trades with profit factor 1.43, win rate 38.9%, and max drawdown 23.9%, but Sharpe was only 0.34 with a CI low of -1.45. After optimization the profile looked cleaner in-sample (Sharpe 0.64, PF 2.11, max DD 17.1% over 35 trades), yet the one-shot holdout produced a negative Sharpe of -2.20.

Backtest and paper results are hypothetical. Trading involves risk of loss.