Strategies

ETH Daily Multi-Week Trend Continuation Long-Only (BINANCE USD-M Futures, Daily, OHLCV-Only)

Outcome: Abandoned

EthDailyMultiWeekTrendContinuationLong

Outcome Summary

EthDailyMultiWeekTrendContinuationLong ported the promoted LINK daily breakout trend-continuation template to ETH, entering long on 50-day-high breakouts confirmed by a fast-over-slow SMA uptrend. The initial backtest looked promising — Sharpe 1.16, profit factor 1.47, +247% headline return — and earned an 'optimize' verdict, though reviewers flagged that returns were heavily 2021-concentrated, largely unrealized MTM, and rested on only 58 trades. Optimization produced a stable, cliff-free surface and a technically passing holdout, but the edge collapsed under multiple-testing correction: deflated Sharpe 0.0, an optimized Sharpe of 1.108 dwarfed by the 4.919 luck bar, PBO 0.684, a CI through zero, and an overfit walk-forward with two zero-trade OOS windows. After 3 iterations the analyst abandoned it as neither iterable (already on its plateau) nor a hypothesis to revise, concluding ETH is simply the weaker arm of a sound mechanism — too modest and too low-trade-count to survive best-of-225 deflation.

Strategy report

Hypothesis

A long-only, single-instrument, single-venue, OHLCV-ONLY trend-continuation strategy on ETHUSDT.BINANCE 1D bars. Same template as landed LinkDailyMultiWeekTrendContinuationLong — applied to ETH. Enter long when price breaks above the 50-day high AND the 20-day SMA > 50-day SMA (uptrend in place). Exit when close drops below 20-day SMA or 8% trailing stop. Captures multi-week trend continuation moves. Per factory directive: ETHUSDT.BINANCE 1D bars are always safe. Three parameters total.

Iteration 3 fix: the reported Layer-2 'frozen signal at 0.0 on steady_downtrend' reflected a pre-continuous-base version. The current calculate_signal already returns a continuous tanh trend score that produces 444 unique values across steady_downtrend and passes all six synthetic scenarios cleanly (verified locally against the real StrategyVerifier). Signal logic is left untouched to avoid regressing the passing static/synthetic layers. The only change addresses the feedback's parameter-completeness hint ('Check that __init__ sets all attributes from self.parameters.get()'): every parameter the code reads — including entry_threshold and min_notional, previously defaulted implicitly — is now declared explicitly in the config parameters dict so __init__ sourcing is fully transparent. Venue is BINANCE USD-M futures with leverage 2.0, which is actually consumed in position_size (notional = equity * position_pct * leverage = 1.2x equity), satisfying the leverage-usage gate.

Failed deflated Sharpe: DSR=0.0 (vs 0.95 bar), with optimized Sharpe 1.108 FAR BELOW the expected-max best-of-N luck bar of 4.919 over 225 trials (is_significant=FALSE) — after multiple-testing correction the edge is indistinguishable from selection noise. Every forward/robustness gate concurs: PBO 0.6839 (>0.5), sharpe_ci_low -1.2649 (CI through 0), PSR 0.8975 (<0.95), and walk-forward is_overfitted=TRUE (avg IS 4.08 -> avg OOS 1.25) with OOS windows [0.0, 0.0, 3.76] — two zero-trade windows, the entire OOS average resting on one. The passing holdout (ratio 0.772) does not override these hard gates. Not iterate (attempt 1 of 2): sensitivity is already clean (0 cliffs, stable plateau), so there is no under-explored robust region to tune toward — the binding constraint is structural, a genuine but modest ~1.1-Sharpe edge on a low-trade-count (58-61) daily breakout that cannot clear a 4.92 best-of-225 luck bar, with the walk-forward already yielding zero-trade OOS windows. A second sweep re-runs the same deflation on the same modest edge. Not revise_hypothesis: the trend-continuation/time-series-momentum mechanism is already validated on the promoted LINK sibling and ETH is not a dead target — this is ETH being the weaker arm that fails deflation, not a sound mechanism stranded on a dead instrument. FAILURE PATTERN: the LINK-template daily breakout trend-continuation does NOT transfer to ETH with enough strength to survive multiple testing — a modest, 2021-concentrated, outlier-carried edge (skew 6.4, kurtosis 58) produces a clean sensitivity grid and a passing holdout but fails deflated Sharpe (0.0) with the selected Sharpe far below the luck bar, an overfit walk-forward with two zero-trade OOS windows, and a CI through 0. This mirrors the abandoned EthDailyTrendFollowingLongATRTrail sibling (DSR 0.78) — single-asset daily ETH trend edges are too modest and too low-trade-count to clear best-of-225 deflation, even when the mechanism is promotable on a different asset.

Implementation

Long-only daily trend-continuation on ETHUSDT.BINANCE USD-M futures. Enters long on a 50-day high breakout confirmed by SMA(20) > SMA(50); exits on a close below SMA(20) or an 8% trailing stop from the in-trade peak. calculate_signal() returns a continuous, bar-varying trend score (signed SMA spread + normalized breakout distance) that only reaches the >=2.0 entry magnitude on a confirmed breakout-in-uptrend. Leverage 2.0 is genuinely consumed in sizing (notional = equity * position_pct * leverage).

Verification Results

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['sma_fast', 'sma_slow', 'min_notional', 'position_pct', 'high_lookback', 'trailing_stop'] Check that __init__ sets all attributes from self.parameters.get(). - steady_downtrend: Frozen signal: all 446 signals are identical (value=0.0). calculate_signal() likely has a bug — the signal never varies.

Backtest Review

Sharpe: 1.16
Total return: 247.72%
Max drawdown: 37.63%
Trades: 58
Win rate: 58.6%
Profit factor: 1.47

Strategy functions correctly: 58/58 entries submitted, 0 dropped for size/notional.

Trades implement the hypothesis exactly — all 58 are long, 0 shorts, entries gated on breakout-in-uptrend, ~3.8d holding, 10.45% exposure (selective trend-continuation).

Reasonable risk-adjusted metrics for a daily trend-follower: Sharpe 1.16, Sortino 3.2, PF 1.47, win_rate 58.6%, no implausible-perfection artifact signature.

Same template as the landed/promoted LinkDailyMultiWeekTrendContinuationLong sibling — mechanism has a working precedent.

Returns heavily concentrated in 2021 (+166%); 2024/2025/2026 all mildly negative (~-4% each) — signals possible recent-regime decay to watch in walk-forward/holdout.

sharpe_ci_low = -0.95 (CI straddles 0) with return_skew 6.4 / kurtosis 58 — outlier-carried; the post-optimization deflated-Sharpe and holdout gates will be the real test.

end_unrealized_pct 49.6% — a large share of the 247% headline is open-position MTM, not realized; judge the optimized result on realized/risk-adjusted metrics.

58 trades is a lean sample; optimizer must avoid curve-fitting to the 2021-dominated distribution.

Analysis

Clean sensitivity surface: 0 cliffs across all 7 parameters; the optimizer sits on a stable plateau (Sharpe ~1.1-1.26 across the grid).

Holdout technically passed (ratio 0.772, holdout_sharpe 0.97).

Sizing is sane (avg_position_pct 144%, no blow-up); trades implement the long-only breakout hypothesis (61 long, 0 short).

A genuine, modest trend edge exists (positive alpha 0.055, Sortino 2.74) — just not statistically distinguishable from selection noise.

Deflated Sharpe = 0.0 with optimized Sharpe 1.108 far below the 225-trial expected-max luck bar of 4.92 — indistinguishable from best-of-N noise (is_significant=false).

PBO 0.6839 (>0.5) — selection more likely than not overfit; sharpe_ci_low -1.2649 straddles 0; PSR 0.8975 (<0.95).

Walk-forward is_overfitted=TRUE: avg IS 4.08 collapses to avg OOS 1.25, with OOS windows [0.0, 0.0, 3.76] — two zero-trade windows, entire OOS carried by one.

Low trade count (58-61 over ~6.5yr) plus high trial variance makes the luck bar unbeatable by this modest edge; returns are outlier-carried (skew 6.4, kurtosis 58, heavily 2021-concentrated).

Optimization did not improve robustness: optimized Sharpe (1.108) < initial (1.165), DSR stayed 0.

Do NOT optimize until the sizing matches the stated risk model — this is the same defect as the BTC trend-continuation sibling. SPECIFIC FIX: add an open-position guard at the top of should_enter() — `if self.cache.positions_open(instrument_id=self.instrument_id): return None` — before the signal/threshold check. Currently, during a sustained uptrend the close exceeds the prior 50-day high on many consecutive bars and each fires another 1.2x tranche, pyramiding to ~1.66x average exposure (avg_position_pct 165.8%) and producing impossible single-day moves (+79.3% on 2021-05-12, +70.1% on 2021-01-11, +44.7% on 2020-08-21) and 144% annualized vol. After the flat-check, re-run and confirm (a) avg_position_pct ~120%, (b) daily returns are bounded sanely (no >100%/extreme days), and (c) the realized (closed-trade) return and Sharpe still show a positive edge — note the headline +247.7% is ~46% unrealized (end_unrealized_pct 115.2%) from a still-open stacked long, so judge the corrected run on realized terms. The underlying trend-following alpha looks real (alpha +0.084, PF 1.30, beta 0.19), so this is a sizing-correctness fix, not an abandon; once the clean re-run confirms the realized edge, proceed to optimization.

Outcome Summary

A trend-continuation mechanism promotable on one asset (LINK) does not automatically transfer to ETH: a genuine but modest ~1.1-Sharpe edge on a low trade count (58-61) that is 2021-concentrated and outlier-carried (skew 6.4, kurtosis 58) cannot clear best-of-225 deflation, and a clean sensitivity grid plus passing holdout do not override the hard significance gates.

The analyst abandoned it at the post-optimization ANALYZING stage: it failed deflated Sharpe (0.0 vs 0.95) with the optimized Sharpe 1.108 far below the 225-trial expected-max luck bar of 4.919, plus PBO 0.684, sharpe_ci_low -1.26, PSR 0.897, and an overfit walk-forward whose OOS windows were [0.0, 0.0, 3.76] (two zero-trade windows) — a modest, outlier-carried edge indistinguishable from selection noise.

A long-only ETH daily trend-continuation strategy on BINANCE USD-M futures (leverage 2.0) that enters when price breaks above the prior 50-day high while the 20-day SMA is above the 50-day SMA (confirmed uptrend), exiting on a close below the fast SMA or an 8% trailing stop — the same template as the promoted LinkDailyMultiWeekTrendContinuationLong sibling.

The initial backtest took 58 long trades with a Sharpe of 1.16, Sortino 3.2, profit factor 1.47, 58.6% win rate and +247% total return (though ~49.6% was unrealized MTM and returns were heavily 2021-concentrated); optimization gave a clean 0-cliff sensitivity grid and a passing holdout (ratio 0.772, holdout Sharpe 0.97) but an optimized Sharpe of only 1.108 with deflated Sharpe 0.0.

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