ETH COIN-M Inverse-Perp Dual-Timeframe Momentum Confluence, Long-Short (ETHUSD_PERP.BINANCE_CM, 4H Primary + 1D Confirm, Flat on Disagreement, ATR-Trailing Exit, 3-Parameter) — Running the Factory's ONE Proven Survivor Mechanism on the Lowest-Fee Binance Venue, Not a Thin-Venue Clone
Outcome: Abandoned
EthCoinMDualTimeframeMomentumConfluenceLS
Outcome Summary
EthCoinMDualTimeframeMomentumConfluenceLS faithfully reproduced the factory's one proven survivor mechanism — 4H/1D momentum confluence, flat on disagreement, ATR-trailing exit — long-short on the ETH coin-margined inverse perp, wagering that the venue's ~0.06% round-trip fee would let the edge survive. The base backtest cleared the COIN-M viability floor but was thin (Sharpe 0.338, PF 1.119, CI straddling zero), so the analyst approved it for optimization. The 3-phase run then exposed it decisively: in-sample Sharpe 1.37 reversed to -0.746 out-of-sample, PBO 0.53 and deflated Sharpe 0.087 confirmed best-of-N overfitting, and drawdown ballooned to 38% against a 20% pre-registered cap. It was abandoned after one iteration, with the lower fee unable to rescue a family that had already died repeatedly on this venue.
A LONG-SHORT, single-instrument, pure-OHLCV directional momentum-confluence trend follower on ETHUSD_PERP.BINANCE_CM (the CURRENT coin-margined inverse perpetual, ETH-margined). The MECHANISM is deliberately identical to the three strategies that actually reached paper_stage (BtcDualTimeframeMomentumConfluenceLS Sharpe 1.99, AdaBinanceDualTimeframeMomentumConfluenceLS 1.60, EthVolumeConfirmedMomentumLS 1.05): take a directional position ONLY when a FAST 4H momentum signal AGREES with the SLOW 1D trend, stand flat……Show moreShow less
A LONG-SHORT, single-instrument, pure-OHLCV directional momentum-confluence trend follower on ETHUSD_PERP.BINANCE_CM (the CURRENT coin-margined inverse perpetual, ETH-margined). The MECHANISM is deliberately identical to the three strategies that actually reached paper_stage (BtcDualTimeframeMomentumConfluenceLS Sharpe 1.99, AdaBinanceDualTimeframeMomentumConfluenceLS 1.60, EthVolumeConfirmedMomentumLS 1.05): take a directional position ONLY when a FAST 4H momentum signal AGREES with the SLOW 1D trend, stand flat on disagreement, and trail winners with an ATR (chandelier) stop. The NOVELTY is not the signal but the VENUE and its economics. COIN-M is the most under-represented Binance venue in the corpus (2.7%, target >=5%) and, critically, it charges ~0.06% round-trip taker vs ~0.10% on USD-M — a 40% cut in the single cost line that kills 483 of our hypotheses. The proven survivor family is Binance-DEEP-LIQUIDITY-major-specific; ETH COIN-M perp is exactly that (top-liquidity Binance contract, tight book), so this is NOT the L94 anti-pattern of cloning the mechanism onto a thin Bybit/Hyperliquid venue where the swap itself is the reason it dies. This is also NOT a beta-neutral/residual/pairs reformulation (L95/L96): it keeps momentum OUTRIGHT and DIRECTIONAL, single-leg, no second fee-paying hedge. Positions are ETH-denominated (inverse contract), which gives the returns a different, portfolio-diversifying profile from the USD-M book we already hold.
Implements the hypothesis exactly: the mechanism is the factory's one proven survivor recipe (fast 4H momentum must agree with the slow 1D trend, flat on disagreement, ATR trailing exit) and the novelty is the venue -- ETHUSD_PERP.BINANCE_CM, the current, actively-listed ETH coin-margined inverse perpetual with ~0.06% round-trip taker vs ~0.10% on USD-M. It stays OUTRIGHT and DIRECTIONAL: single leg, no hedge, no residual/beta-neutral reformulation, no second fee-paying instrument. Implementation details that……Show moreShow less
Implements the hypothesis exactly: the mechanism is the factory's one proven survivor recipe (fast 4H momentum must agree with the slow 1D trend, flat on disagreement, ATR trailing exit) and the novelty is the venue -- ETHUSD_PERP.BINANCE_CM, the current, actively-listed ETH coin-margined inverse perpetual with ~0.06% round-trip taker vs ~0.10% on USD-M. It stays OUTRIGHT and DIRECTIONAL: single leg, no hedge, no residual/beta-neutral reformulation, no second fee-paying instrument. Implementation details that matter for this venue: (1) COIN-M is INVERSE, so position_size() returns whole CONTRACTS via the base helper contracts_for_notional() rather than the linear equity*frac/price coin quantity, which would mis-size by orders of magnitude; the ~$29k typical notional at 1.3% risk is ~2,900 contracts, far above min notional, so the run cannot die on a below-face size. (2) The slow leg is the SAME instrument's 1-DAY bars, listed in both extra_instruments and extra_bar_types (the resolver pairs them by index, so an empty extra_instruments would drop the leg and stall the alignment barrier); the differing timeframe means no same-timeframe sync barrier, so the primary 4H bar processes immediately. (3) calculate_signal returns the raw ATR-normalised spread EVERY bar (a continuous, varying decision variable), with the entry decision thresholded in should_enter -- no gated constant-return pattern. (4) The entry gate is a 2-condition conjunction of two persistent trend states, not a rare compound filter: on 4H ETH data over the last ~2.5 years the confluence fires and exits dozens of times, comfortably above the 1-trade sandbox floor and giving a decisive sample. (5) All per-bar state is incremental (EMAs, deque-bounded ATR, deque-bounded daily closes), so the 300s Layer-3 wall clock is not at risk. (6) leverage is left at 1.0 -- the risk-first sizing tops out at 35% of equity, so no margin amplification is requested or relied on, avoiding the leverage-set-but-unused defect and the compounding-sizing blowups. Per-trade edge: trend legs held to an ATR trailing stop on 4H ETH average multi-percent moves, well clear of the ~0.06% COIN-M round trip.
Overfit single-name OHLCV momentum-confluence clone. Walk-forward IS 1.37 -> OOS -0.746 (is_overfitted=true, 2 of 3 windows negative), OOS Sharpe <= 0, and PBO 0.53 > 0.5 — three HARD validity gates fail, so promotion is impossible and the edge REVERSES out-of-sample rather than merely decaying. Confirmed quantitatively by failed deflated Sharpe: DSR=0.087, expected-max=1.06 over 201 trials, is_significant=false, does not survive programme FDR (candidate_p 0.91, 4 of 273 kept) — the selected Sharpe is best-of-N……Show moreShow less
Overfit single-name OHLCV momentum-confluence clone. Walk-forward IS 1.37 -> OOS -0.746 (is_overfitted=true, 2 of 3 windows negative), OOS Sharpe <= 0, and PBO 0.53 > 0.5 — three HARD validity gates fail, so promotion is impossible and the edge REVERSES out-of-sample rather than merely decaying. Confirmed quantitatively by failed deflated Sharpe: DSR=0.087, expected-max=1.06 over 201 trials, is_significant=false, does not survive programme FDR (candidate_p 0.91, 4 of 273 kept) — the selected Sharpe is best-of-N noise. daily_lookback is a cliff parameter with no robust region to iterate toward. The positive holdout (0.806/44 trades) is not enough against negative WF-OOS + overfit + high PBO. Per the decision framework, overfit -> abandon (not revise): reframing the same momentum mechanism onto another instrument re-overfits rather than fixing a premise flaw, and the COIN-M port of this exact dual-TF/TSMOM family has already died repeatedly (XRP CM fee_edge, BTC CM code_defect, ETH CM dual-EMA/ignition overfit/risk_reject). The lower CM fee cannot manufacture robustness the 3-phase deflation strips out.
Implementation
Long-short, single-instrument, pure-OHLCV dual-timeframe momentum-confluence trend follower on ETHUSD_PERP.BINANCE_CM (the current ETH-margined INVERSE coin-M perpetual). The FAST leg is the primary 4H bar stream: an ATR-normalised EMA spread (EMA_fast - EMA_slow) / ATR, computed incrementally every bar (O(1), no history rescans) and returned as the continuous signal. The SLOW leg is the SAME instrument's 1-DAY bars consumed in on_extra_bar: daily_dir = sign(close_d[t]/close_d[t-daily_lookback] - 1). A position……Show moreShow less
Long-short, single-instrument, pure-OHLCV dual-timeframe momentum-confluence trend follower on ETHUSD_PERP.BINANCE_CM (the current ETH-margined INVERSE coin-M perpetual). The FAST leg is the primary 4H bar stream: an ATR-normalised EMA spread (EMA_fast - EMA_slow) / ATR, computed incrementally every bar (O(1), no history rescans) and returned as the continuous signal. The SLOW leg is the SAME instrument's 1-DAY bars consumed in on_extra_bar: daily_dir = sign(close_d[t]/close_d[t-daily_lookback] - 1). A position is taken ONLY when the two agree (spread>0 and daily_dir>0 -> LONG; spread<0 and daily_dir<0 -> SHORT); on any disagreement the strategy stands FLAT. Exits are (a) an ATR chandelier trailing stop -- close when price gives back trail_atr_mult x the entry ATR from the best close since entry, (b) the daily trend flipping against the position, (c) the 4H spread flipping against the position. Sizing is inverse-correct: USD notional is set risk-first (equity x 1.3% / trailing-stop distance, i.e. ~1.3% of equity risked per trade), capped at 35% of equity, then converted to an INTEGER COIN-M CONTRACT count via contracts_for_notional() ($10 USD face per ETHUSD_PERP contract) -- never coin units. Three tunables only: ema_ratio, daily_lookback, trail_atr_mult.
Verification Results
Confirm the base template does not double-gate warm-up; optionally rename to a private helper (e.g. _warmup_bars()) to avoid shadowing.
min_bars_required() overrides a framework method; used only internally as a warm-up gate in should_enter, which is harmless, but be aware the base template may also consult its own warm-up logic. Behaviour is consistent here (the override returns a larger, stricter value), so no functional defect.
On restart, prefer inferring side from the actual open position in cache rather than from signal sign.
should_exit() reconstructs self._side from the live signal sign if state was lost on restart. In a live crash-restart this could infer the wrong direction for an existing position for one bar until the daily/spread flip exits re-align it. Low impact in backtest (state never lost) but a minor live-parity risk.
Faithful implementation but headline economics are marginal: sandbox Sharpe 0.34 with CI [-0.44, 1.11] straddling zero, win_rate 34.6%, and the 'normal' vol tercile is net-negative (-7.74%, Sharpe -0.11). This is the single-name pure-OHLCV dual-timeframe momentum-confluence clone family (L82/L93) whose only survivor to date is the ADA Binance USD-M instance (Sharpe 1.60); most siblings abandon near breakeven. The COIN-M fee cut (~0.06% vs ~0.10% RT) helps per-trade viability but does not by itself create edge. Worth the analyst watching whether the full-history backtest confirms edge concentrated only in the calm regime.
Backtest Review
Sharpe
0.34
Total return
35.07%
Max drawdown
26.66%
Trades
460
Win rate
33.7%
Profit factor
1.12
Faithful implementation of the factory's ONE proven survivor mechanism (dual-TF 4H/1D momentum confluence, flat on disagreement, ATR-chandelier trail) — the same mechanism that reached paper on BTC/ADA/ETH USD-M. Trades match the hypothesis: 460 signaled = 460 submitted, balanced 227 long / 233 short, avg hold ~3.5d, single-leg directional (not a beta-neutral/pairs reformulation).
avg_trade_return_pct 0.118% clears the COIN-M low-fee floor (>0.10%) and sits at ~2x the ~0.06% round-trip cost — the venue thesis (lower fee line vs USD-M) is doing real work; commissions are only 4.88% of gross, so the edge is not fee-dominated.
Correct inverse-contract sizing for COIN-M; three tunables only (ema_ratio, daily_lookback, trail_atr_mult) keeps the multiple-testing luck bar low for the 3-phase run.
Thin edge — PF 1.119 and Sharpe 0.338 with bootstrap CI [-0.44, 1.11] straddling zero; the point Sharpe is not yet distinguishable from noise.
Regime/year concentration (QA's concern borne out on full history): profit is in the calm (+26.4%) and stressed (+13.9%) terciles while the normal tercile is net-negative (-7.7%, Sharpe -0.11); profitable in only 3 of 7 calendar years (2021, 2022, 2026) and loses in 2020/2023/2024/2025 — below the 60% positive-years promotion floor.
This is the single-name pure-OHLCV dual-TF momentum-confluence clone family whose only paper survivor so far is the ADA USD-M instance; base metrics sit well below that survivor tier, so the 3-phase deflation is a real risk.
Analysis
Faithful implementation of the factory's proven survivor mechanism (dual-TF 4H/1D momentum confluence, flat on disagreement, ATR-chandelier trail); trades match the hypothesis (307 signaled = 307 submitted, balanced 151 long / 156 short).
Optimized per-trade return 1.04% clears the COIN-M fee floor comfortably, PF 1.59, impact negligible (0.16% of gross), capacity effectively unlimited.
Holdout Sharpe positive (0.806 on 44 trades) — the one hard holdout gate passed.
Decisively OVERFIT: walk-forward IS 1.37 collapses to OOS -0.746 (is_overfitted=true), with 2 of 3 windows negative (-0.46, -1.81, +0.03) — the edge reverses on unseen data, it does not merely decay.
OOS Sharpe -0.746 <= 0 and PBO 0.53 > 0.5 — two more HARD validity gates failed; the parameter selection is more likely than not overfitting.
Failed deflated Sharpe: DSR 0.087, expected-max 1.06 over 201 trials, is_significant=false, does not survive programme FDR (candidate_p 0.91, keeps 4 of 273) — the selected Sharpe is indistinguishable from best-of-N noise.
daily_lookback is a cliff parameter (Sharpe swings 0.18->0.56->0.37 across ±20%), so there is no robust region to tune toward.
Drawdown 38% is nearly 2x the pre-registered 20% cap; 'normal' vol regime is a persistent loser (Sharpe -0.11, -13% return).
The identical mechanism ported to COIN-M has died repeatedly: XRPUSD_PERP.BINANCE_CM (fee_edge), BTCUSD_PERP.BINANCE_CM (code_defect), ETH COIN-M dual-EMA/TSMOM/ignition variants (overfit/risk_reject) — the venue swap has never rescued this family.
Benjamini-Hochberg at q=0.10 over 273 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.
Outcome Summary
Moving an already-marginal momentum-confluence mechanism to a lower-fee venue does not manufacture robustness that optimization-stage deflation strips out — this COIN-M family has now repeatedly overfit, so re-porting the same mechanism to another instrument re-overfits rather than fixing a premise flaw.
The analyst abandoned it at the post-optimization ANALYZING stage: three HARD validity gates failed (is_overfitted=true, OOS Sharpe -0.746 <= 0, PBO 0.53 > 0.5), the deflated Sharpe was indistinguishable from best-of-N noise and did not survive programme FDR, so promotion was impossible and the edge reversed rather than merely decayed out-of-sample.
Run the factory's proven dual-timeframe momentum-confluence mechanism (4H fast signal must agree with the 1D trend, flat on disagreement, ATR-chandelier trailing exit) long-short on the ETHUSD_PERP.BINANCE_CM inverse perpetual, betting the venue's lower ~0.06% round-trip taker fee would preserve an edge that USD-M fees kill.
On a 6-year, 460-trade base backtest it was marginally viable (Sharpe 0.338 with CI [-0.44, 1.11], profit factor 1.119, avg trade return 0.118% of notional, max drawdown 26.7%), but after 3-phase optimization the walk-forward in-sample Sharpe of 1.37 collapsed to out-of-sample -0.746 (2 of 3 windows negative), with PBO 0.53, deflated Sharpe 0.087, and optimized drawdown 38%.
Backtest and paper results are hypothetical. Trading involves risk of loss.