Strategies

ETH COIN-M Inverse Perp 4H Sequential Higher-Low Accumulation Breakout Long-Only (BINANCE_CM, OHLCV-Only, 3-Parameter)

Outcome: Abandoned

EthCmInverseFourHourAccumulationHigherLowBreakoutLong

Outcome Summary

EthCmInverseFourHourAccumulationHigherLowBreakoutLong aimed to fill the under-represented COIN-M venue quota with a textbook Dow-Theory accumulation signal — three rising swing lows plus a breakout — long-only on the ETH inverse perp, carefully engineered to sidestep a long list of prior failure patterns. The backtest was cleanly implemented and cleared the CM fee floor (avg trade +0.994%), but delivered no risk-adjusted edge: Sharpe 0.172 with a CI straddling zero, information ratio -0.647, and a 0.84 correlation to ETH showing the +14.1% was captured beta carried by a few bull-year days. It was also thin (83 trades over 5.8 years, below its own frequency claim) and had decayed to negative in 2025 and 2026. As another net-edgeless member of the 0/213 OHLCV breakout family, the analyst abandoned it at backtest review rather than spend an optimization run that would only curve-fit its tail winners.

Strategy report

Hypothesis

A long-only, single-instrument, single-venue trend-initiation strategy on ETHUSD_PERP.BINANCE_CM (ETH-margined inverse perpetual) using 4-HOUR bars and OHLCV-only data. Mechanism: detects a classic Dow Theory accumulation phase via three consecutive higher swing lows, then enters long on a confirmed breakout above the recent N-bar high. This is a structurally DIFFERENT mechanism class from higher-high momentum breakouts already in the portfolio (which fire on rising tops, indicating momentum continuation) — higher-LOWS specifically indicate accumulation under resistance, which mechanically reduces float supply and often precedes trend transitions. STRATEGIC RATIONALE: (1) PORTFOLIO QUOTA FILL — BINANCE_CM is at 1.1% vs ≥5% target, the second-most under-represented venue after Options (0%). CM has only 12 experiments vs 915 on BINANCE USD-M futures. Inverse perpetuals have mechanically distinct P&L profiles (returns denominated in the base coin, not USD) producing different drawdown characteristics than linear perpetuals — they deserve dedicated coverage as a distinct risk sleeve. (2) FEE FRIENDLINESS — CM ~0.06% RT taker vs USD-M ~0.10% RT. This 40% fee reduction lowers the avg_trade_return viability floor from 0.15% to 0.10%, expanding the universe of mechanisms that can be net-positive. (3) MECHANISM PROVENANCE — sequential higher-low detection is a textbook accumulation signal from Dow Theory (1900s), Wyckoff (1930s), and Pring (1980s). It works because 3 consecutive higher lows require sustained buying absorption that reduces float; combined with breakout confirmation, the conjunction has higher win-rate than naked breakouts. (4) AVOIDS ALL RECENT FAILURE PATTERNS — OHLCV-only (no liquidations/funding/HL-supplementary data risk that killed BTC Cross-Venue Funding, ETH Liquidation Cluster, BTC CM Funding MR); single instrument single venue (no cross-venue execution coordination complexity); 4H bars proven sleeve (NOT 1H deprecated, NOT daily with WF-warmup deficit); only 3 hypothesis-declared tunable parameters with strict scope discipline (avoiding the 12-parameter bloat that killed ETH Pullback and AVAX Trend-Pullback); symmetric ATR trailing stop on entry (no asymmetric TP<SL math that killed ETH Liquidation); natural trade frequency 25-40/yr on 4H bars (avoiding trade-starvation that killed Multi-Week TFL with 9 trades/year); NOT BNB (avoiding post-2024 moratorium); NOT AVAX (skewness-regime mechanism class deprecated); NOT pair-MR (which has 1H deprecation and methodology contradictions); NOT funding-rate based (avoiding the HL funding data wiring issue that killed BTC Cross-Venue Funding Differential); NOT skewness (mechanism class empirically refuted on daily crypto alts). (5) DISTINCT FROM JUST-LANDED SolHyperliquidFourHourDonchianBreakoutLS — that strategy uses pure rolling-channel Donchian on SOL HL bidirectional; this uses Dow-Theory higher-lows + breakout confirmation on ETHUSD_PERP CM long-only. Different mechanism, different asset, different venue, different direction.

Implements the hypothesis exactly: sequential higher-LOW accumulation structure (three strictly rising confirmed fractal swing lows) as the setup, N-bar high breakout as the trigger, long-only on the ETH COIN-M inverse perp at 4H, OHLCV only, single instrument/venue, symmetric ATR trailing stop, three declared tunables. calculate_signal returns a continuous ATR-normalised composite every bar — breakout distance (close - prior N-bar high)/ATR plus half the swing-low staircase slope (p3-p1)/ATR — so the signal varies in every regime (negative when lows are falling) while the hard structural test lives in should_enter. Inverse-contract handling per the COIN-M rules: contracts_for_notional (whole $10 contracts), never coin units from equity/price; the linear branch exists only for the non-inverse Layer-2 proxy instrument. Leverage is left at 1.0 (risk-based sizing already lands at ~30-45% notional with a ~4.5% ATR stop, so an unused leverage multiplier would be inert) and per-trade risk is capped at 1.5% of equity against the trailing-stop distance, avoiding compounding-sizing blowups. Dry-run on the real ETHUSD_PERP 4H catalog history (12,683 bars, 2020-08 to 2026-06) with the default parameters: 83 trades (~14/yr), avg per-trade return +1.36% net of 0.06% CM round-trip fees, 37% win rate with a long right tail (mean hold ~40 bars), avg notional ~$34k = ~3,400 contracts (far above min notional), and the expectancy stays positive across the whole declared parameter box — so the entry conjunction fires reliably inside the sandbox window and clears the 0.10% CM fee floor by an order of magnitude. Warmup is only 22 bars.

low_trade_count_no_edge_ohlcv_breakout_clone: ETH COIN-M 4H higher-low accumulation breakout is a single-instrument pure-OHLCV directional-breakout clone (0/213 survival, L9/L41/L63) with no measurable risk-adjusted edge. Only 83 trades in 5.8 years (~14/yr, below L10's <100 tripwire and below the hypothesis's own 25-40/yr claim), Sharpe 0.172 with CI [-0.45, 0.83] straddling zero. The 14.1% total return is captured ETH beta, not skill: benchmark_correlation 0.84, information_ratio -0.647, alpha -0.048, and the PnL is tail-driven (kurtosis 8.66) — a few bull-year days (2024-03-05 +12.9%, 2021-01-11 +11.1%) carry it. Regime-concentrated in 2021-2024 and negative in both 2025 (-3.9%) and 2026 (-4.2%) with rolling Sharpe deeply negative into 2026 — the mechanism has decayed. avg_trade_return_pct 0.994% clears the CM fee floor so fees were never the binding constraint; the edge itself is absent. A 225-trial / 3-param sweep on 83 trades would curve-fit the handful of bull-year tail winners and die at DSR/holdout. Not worth the 2-hour optimization run.

Implementation

Long-only Dow/Wyckoff accumulation-breakout on ETHUSD_PERP.BINANCE_CM (ETH-margined inverse perpetual), 4-HOUR bars, OHLCV only. Confirmed swing lows are detected with a symmetric fractal of half-width swing_strength; when the last three confirmed swing lows are strictly rising (accumulation: buyers defending at progressively higher levels) and the newest pivot is fresh (within breakout_lookback bars), a close strictly above the highest high of the prior breakout_lookback bars triggers a long entry. Exit is a symmetric ATR chandelier trail: highest close since entry minus atr_trail_mult * ATR(14), checked at every bar close (it also serves as the initial stop). Sizing is risk-based — the ATR stop distance is converted into a USD notional risking risk_pct of equity, capped at max_gross_frac of equity, then floored into WHOLE COIN-M contracts via contracts_for_notional ($10 per contract for ETHUSD_PERP). Three tunable parameters (swing_strength, breakout_lookback, atr_trail_mult) with declared _param_bounds matching the in-code clamps.

Verification Results

Route to backtest review; require full-history Sharpe CI to clear zero and check regime concentration (32% DD; end_unrealized -18.4% signals a large open-mark component in the headline return).

Mechanism-family fragility (informational, not a code defect): pure-OHLCV directional breakout (close above prior-N high) plus a rising-swing-low accumulation filter — structurally a Donchian/high-breakout class that historically fails to generalize. Sandbox is positive (PF 1.49, avg_trade 1.32%, +14.1%) so it is NOT a no-edge death, but base Sharpe 0.17 with CI [-0.45, 0.83] straddling zero and IR -0.65 mean the risk-adjusted edge is indistinguishable from no-skill at this stage.

Flag for the optimizer/analyst to weight DSR and PBO heavily given the small trade count.

83 trades over ~5.8 years (~14/yr) is a modest sample — natural for a ~40-bar-hold 4H swing mechanism, not a zero-trades defect, but thin for fitting 3 params without overfitting.

Optional: guard the linear branch behind an explicit non-inverse instrument check.

The non-inverse linear fallback in position_size() exists only for the Layer-2 proxy. On the real inverse instrument contracts_for_notional returns >0 so it is never reached (confirmed by sandbox trades), but if it ever returned 0 live, sizing would silently fall to coin units.

Backtest Review

Sharpe: 0.17
Total return: 14.15%
Max drawdown: 32.09%
Trades: 83
Win rate: 36.1%
Profit factor: 1.38

avg_trade_return_pct 0.994% clears the CM 0.10% fee floor — not a fee death

Clean single-instrument single-venue OHLCV implementation; inverse-contract sizing handled correctly; 83 signaled = 83 submitted (no plumbing bug)

profit_factor 1.38 nominally positive

Only 83 trades over 5.8 years (~14/yr) — below L10's <100 tripwire and below the hypothesis's own 25-40/yr claim; Sharpe is unmeasured

No risk-adjusted edge: Sharpe 0.172 with CI [-0.45, 0.83] straddling zero

Closet-long ETH: benchmark_correlation 0.84, information_ratio -0.647, alpha -0.048 — the 14% is captured beta, not skill

Tail-driven (kurtosis 8.66) — a few bull-year days carry the whole result; strip them and the edge vanishes

Regime-concentrated: positive 2021-2024, negative 2025 (-3.9) and 2026 (-4.2); rolling Sharpe deeply negative into 2026 — mechanism has decayed

Single-instrument pure-OHLCV directional-breakout family (0/213 survival, L9/L41/L63); the accumulation/higher-lows relabeling adds no robustness the deflation can't strip

Outcome Summary

Relabeling a breakout as Dow/Wyckoff 'accumulation via higher lows' adds no robustness that deflation can't strip — a long-only OHLCV breakout on a high-beta alt captures beta, not alpha (correlation 0.84, negative information ratio), and ~14 trades/year over a tail-driven, regime-concentrated sample cannot support a measurable edge or survive optimization.

It was abandoned at the pre-optimization BACKTEST_REVIEW gate: fees were never the binding constraint, but the strategy is a single-instrument pure-OHLCV directional-breakout clone (0/213 survival) with no measurable edge — the +14.1% is captured ETH beta not skill, carried by a handful of bull-year days, and the mechanism has decayed (rolling Sharpe deeply negative into 2026). With only 83 trades, a 225-trial/3-param sweep would curve-fit the tail winners and die at DSR/holdout, so it was judged not worth the optimization run. Optimization, analysis, and risk review were never reached.

A long-only, single-instrument, single-venue trend-initiation strategy on the ETH COIN-M inverse perp (ETHUSD_PERP.BINANCE_CM, 4H, OHLCV-only) that detected a Dow-Theory/Wyckoff accumulation phase via three consecutive rising swing lows and entered long on a confirmed breakout above the recent N-bar high, exiting on a symmetric ATR chandelier trail.

Over 5.8 years it made only 83 trades (~14/yr, below its own 25-40/yr claim) with total return +14.1%, profit factor 1.38, and avg_trade_return_pct 0.994% (clearing the CM 0.10% fee floor), but no risk-adjusted edge: Sharpe 0.172 with CI [-0.45, 0.83] straddling zero, information ratio -0.647, alpha -0.048, benchmark correlation 0.84, and 32.1% max drawdown. The result was tail-driven (kurtosis 8.66) and regime-concentrated — positive 2021-2024 but negative in 2025 (-3.9%) and 2026 (-4.2%).

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