Strategies

ETC Perp 4H Trend, Gated by MAJOR-COMPLEX BREADTH — Long-Short Directional Momentum on ETCUSDT.BINANCE That Trades ONLY When a Majority of the Top-5 Majors Confirm the Same Market-Wide Regime (breadth is a signal-only overlay, no second fee-paying leg; ATR-chandelier exit, 3-parameter)

Outcome: Abandoned

EtcFourHourTrendMajorBreadthGatedLS

Outcome Summary

The strategy traded ETC perpetuals on 4H trends and only took positions when breadth across five majors confirmed a market-wide regime. The initial backtest cleared fees easily at 1.11% per trade, but its risk-adjusted edge was thin (Sharpe 0.28, PF 1.14) and it lost money in 2023-2025. The analyst still sent it to optimization, where it failed the holdout (Sharpe -0.16 on 52 trades) and the overfitting check (PBO 0.53). The analyst abandoned it after one iteration as no_edge, and separately flagged suspicious paired loss-then-gain days in the optimized daily series for engine review.

Strategy report

Hypothesis

A LONG-SHORT, DIRECTIONAL, single-traded-instrument 4H trend follower on ETCUSDT.BINANCE USD-M perpetual (listed 2020-01-16, 6.6 years of 4H history verified in the catalog, top-50 liquidity, NEVER traded in this factory's 3084-experiment corpus). The novelty is the confirmation variable: instead of confirming the trend with something measured on the SAME price series (volume surge, path efficiency, dual-timeframe, OI, taker flow, funding — all already spent here), the trade is gated by CROSS-SECTIONAL BREADTH of the major complex. Five majors (BTC, ETH, SOL, XRP, BNB, DAILY bars) are carried as SIGNAL-ONLY extra instruments — never traded, so no second round-trip is paid and lesson L95 does not apply: the position stays outright and directional. Breadth b = fraction of available majors whose daily close is above their own 50-day SMA. Premise: an idiosyncratic 4H trend on a mid-cap alt is mostly noise, but the same trend embedded in a market-wide regime (b ≥ 0.6 up, b ≤ 0.4 down) is a real risk-appetite move that persists for days; standing flat in the disagreement zone removes the chop that kills single-name alt trend followers. Deliberately 3 parameters (mom_lookback, atr_mult, breadth_threshold).

Implements the hypothesis exactly: ONE traded instrument (ETCUSDT perp, outright and directional), with the confirmation variable coming from a DIFFERENT cross-section rather than the same price series. The five majors are declared in extra_instruments/extra_bar_types purely for signal -- no order is ever submitted on them (only _submit_entry/_submit_exit on the primary), so no second round-trip fee is paid and the book carries no hedge leg. Breadth is recomputed in on_extra_bar() from each major's own 50-day SMA using only bars already delivered, so it is point-in-time; majors with fewer than breadth_sma_days daily bars are excluded and breadth is undefined (no trading) until at least min_majors=3 qualify. The daily extra legs are a slower timeframe than the 4H primary, so the base class's same-timeframe alignment barrier does not gate the primary bar. The three core parameters named in the hypothesis (mom_lookback, atr_mult, breadth_threshold) drive entry, exit and sizing; I added one small signal-quality parameter, entry_atr_thresh (default 0.5 ATR), because entering on a bare zero-crossing of the momentum score would churn 4H bars right at the anchor and hand the ~0.10% round-trip fee the whole edge -- the threshold plus the zero-crossing exit creates a hysteresis band. Expected firing rate: breadth in regime roughly half the time and |signal| > 0.5 ATR roughly a third of bars, so entries are frequent enough (hundreds of trades over 6.6 years of 4H history) while the chandelier holds winners for days; per-trade moves are multiple percent (4H ETC ATR is ~1-2% of price, the trail is 3x that), comfortably above the fee floor. Venue is futures because the strategy goes short; leverage is left at 1.0 and notional hard-capped at 90% of equity, with per-trade risk capped at 1.5% of equity by the ATR stop, to avoid the compounding-sizing blowups seen in past trend strategies.

Failure code no_edge, taken from the first measured report (Sharpe 0.28, PF 1.14, losses in 2023-2025). It is not fee_edge: 1.11% per trade against a 0.15% floor. Optimization then failed two hard gates that cannot be waived: holdout Sharpe -0.16 on 52 trades, and PBO 0.53. Deflated Sharpe is 0.894, the Sharpe confidence interval includes zero, and the result fails programme-level FDR. The breadth gate did not rescue the single-name alt trend: the profit comes from 2021-2022 and the edge fades from 2023 on. Under L126, re-optimizing the same mechanism is not an allowed iterate. Separately, the paired loss-then-gain days in the optimized daily series need engine review; that does not change this verdict.

Implementation

Long/short 4H trend follower on ETCUSDT.BINANCE USD-M perpetual, gated by cross-sectional breadth of the major complex. The continuous signal is (close - SMA(mom_lookback)) / ATR(atr_period) -- distance from the momentum anchor in ATR units, computed every bar. Five majors (BTC, ETH, SOL, XRP, BNB) are carried as DAILY signal-only extra instruments and are never traded: breadth b = fraction of available majors whose daily close is above their own 50-day SMA. A long requires signal > entry_atr_thresh AND b >= breadth_threshold (0.6); a short requires signal < -entry_atr_thresh AND b <= 1 - breadth_threshold (0.4). Inside the disagreement zone the strategy stands flat. Exits on an ATR chandelier trail (atr_mult x ATR off the excursion extreme), a momentum sign flip, or breadth flipping into the opposite regime. After a trail stop the same side is locked out until momentum crosses back through zero, so a stop inside an intact regime cannot immediately re-enter. Sizing is risk-budget based (risk_pct of equity against the ATR stop distance), hard-capped at max_notional_frac of equity; no leverage.

Verification Results

Hypothesis is deliberately 3-parameter, but `_param_bounds` exposes NINE tunables to the optimizer — the anti-overfit intent is diluted. Suggest freezing all but mom_lookback / atr_mult / breadth_threshold.

`calculate_signal` returns 0.0 as its missing-data sentinel, and `should_exit` treats signal<=0 (long) / >=0 (short) as a momentum flip — so a flat-price stretch with ATR==0 unconditionally flattens the book and pays a round trip on a non-signal. Add a `_signal_valid` flag.

Chandelier `_hh`/`_ll` are seeded from the first POST-entry bar (position is opened after `calculate_signal` runs), so the entry bar's extreme never enters the trail anchor. Cosmetic — it can only loosen the stop, never fire early.

Sandbox (unoptimized) is profitable but thin in risk-adjusted terms: Sharpe 0.26 (CI -0.49 to 1.04), max DD 34.5% with a 1514-day underwater stretch, win rate 29.9% / PF 1.19 — a trend-follower payoff shape, so not disqualifying, but the return is carried by a fat right tail (largest win $31k vs avg win $2.9k, kurtosis 16.3) and the normal-vol tercile is already negative (-1.27%). Check first at backtest review: (1) ABLATION of the breadth gate (run with breadth_threshold=0.5, gate effectively off) — the entire novelty is the cross-sectional overlay, and if gated and ungated books are within noise, the stated edge is not what produces the return; (2) CAPACITY — impact_cost_pct 23.6% of gross with capacity_usd only $1.79M on a mid-cap alt, still under the 50% soft floor but the optimizer raising risk_pct/max_notional_frac could push it through. Also 9 optimizable parameters vs the hypothesis's stated 3 — watch PBO in phases 2/3.

Backtest Review

Sharpe: 0.28
Total return: 71.90%
Max drawdown: 34.13%
Trades: 469
Win rate: 28.4%
Profit factor: 1.14

Numerically viable sample: 469 trades over 6.6 years (2429 days) of 4H ETC data, metrics_reliable=true, 242 long / 227 short — a real two-sided book, not a degenerate or frozen signal.

Per-trade edge well clear of costs: avg_trade_return_pct 1.11% vs the 0.15% BINANCE USD-M floor (7x); commissions only 4.85% of gross. Not a fee-fragility case.

Trades match the stated mechanism: short ETC through the March-2020 crash (majors below their 50d SMAs => breadth-down) and long into the Aug-2026 rally; 28.4% win rate with 2.9:1 payoff is the correct trend-follower shape and the 2d8h hold matches a 4H chandelier trail.

Entry diagnostics clean: 469 signals -> 469 submissions, 0 dropped on size/notional/cash, 0 bars waiting on extra legs — the 5 signal-only daily majors are wired correctly and cost no second round trip.

Positive through most stress windows (covid +7.1%, luna +9.2%, rate-shock-2022 +27.3%).

Risk-adjusted result is thin and not distinguishable from zero: Sharpe 0.279, bootstrap CI [-0.466, 1.052] straddles 0, PSR 0.718, PF 1.139.

Time-concentrated: 2021 (+47.7%) and 2022 (+28.6%) carry everything; 2023/2024/2025 all negative — 4 of 7 years positive (57%, under the 60% hard-gate line), with a 1514-day max-drawdown duration.

Regime attribution contradicts the premise: normal-vol tercile is -42.1% (Sharpe -0.33) while calm and stressed carry the result — the breadth gate is supposed to remove exactly that middle-regime chop.

22.2 of the 71.9-point headline (31%) is open-position unrealized at backtest end.

Tight capacity: impact already 30.1% of gross, capacity ~$1.1M on a mid-cap alt.

9 tunable parameters vs the hypothesis's advertised 3 — real PBO risk in phases 2/3.

Analysis

Fees are not the problem: 1.11% per trade on the first report and 2.04% after optimization, against the 0.15% floor.

No cliff parameters, and CPCV walk-forward is not flagged as overfit (in-sample 0.92, out-of-sample 0.545).

The book trades both directions (159 long / 166 short) and the breadth gate works as designed.

HARD: holdout Sharpe is -0.16 on 52 unseen trades (2024-04-07 to 2025-04-27). That is enough trades to count as a measured negative result.

HARD: PBO is 0.53, above 0.5. The chosen parameters were the in-sample winner in only 22% of splits.

Deflated Sharpe is 0.894 against a 0.95 bar. The Sharpe confidence interval (-0.16 to 1.38) includes zero, and the result does not survive programme-level FDR.

The first report is weak: Sharpe 0.28, PF 1.14, drawdown 34% against the pre-registered 25%, and 2023, 2024 and 2025 all lost money.

The edge decays after 2022: the optimized run lost 10.9% in 2025 and 45.9% so far in 2026.

The strategy's own pre-registered predictions: 2 of 6 met.

Suspected engine artifact, not verified: the optimized daily series has recurring loss-then-gain day pairs that grow over time (e.g. -35% then +52%). This should go to engine review.

Outcome Summary

A cross-sectional breadth gate did not rescue single-name alt trend-following: fees were not the issue at 1.11% per trade, but the profit came from 2021-2022 and the edge faded from 2023 on.

The analyst abandoned it as no_edge. Optimization failed two hard gates that cannot be waived: a negative holdout Sharpe (-0.16 on 52 trades) and PBO 0.53, above the 0.5 limit. The Sharpe confidence interval included zero and the result did not survive programme-level FDR.

A long-short 4H trend follower on the ETCUSDT.BINANCE perpetual. It entered only when breadth across five majors (BTC, ETH, SOL, XRP, BNB) confirmed the same regime: at least 0.6 of their daily closes above the 50-day SMA to go long, at most 0.4 to go short. The majors were signal-only and never traded. Exits used an ATR chandelier trail.

The initial backtest made +71.9% over 469 trades and 6.6 years, with Sharpe 0.28, PF 1.14, a 34.1% max drawdown and 1.11% average return per trade. 2023, 2024 and 2025 all lost money. After optimization the full-period Sharpe was 0.64, but the holdout Sharpe was -0.16 on 52 trades, PBO was 0.53 and the deflated Sharpe was 0.894.

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