BTC Directional Momentum, Gated by ALT-Breadth Confirmation — Long-Short, Trade BTC Only When the Broad Alt Complex Confirms the Move, Flat When BTC Moves Alone (BTCUSDT.BINANCE USD-M + 8-Alt Breadth Signal, 3-Parameter)

A LONG-SHORT directional strategy that trades ONLY BTCUSDT.BINANCE — the single asset whose directional momentum has actually survived here (the one paper-stage strategy is BTC momentum; every alt directional trend is 0/213) — but replaces the winner's SECOND-TIMEFRAME confirmation with a genuinely different, orthogonal filter: ALT-COMPLEX BREADTH. The core insight is that a BTC trend has staying power when the broad altcoin complex PARTICIPATES, and is a fragile, whipsaw-prone head-fake when BTC moves alone while alts diverge. So the strategy computes, from ~8 liquid alts (ETH, SOL, BNB, XRP, ADA, DOGE, AVAX, LINK), the FRACTION trading above their own trend, and takes a BTC position ONLY when BTC's own momentum AND alt-breadth agree; it stands FLAT when breadth is neutral or contradicts BTC. This is NOT another dead template: it is NOT a single-timeframe alt trend (traded asset is BTC, the survivor), NOT a mean-reversion fade (L53), NOT a cross-sectional rank rotation (L52 — nothing is ranked or rotated; breadth is a scalar regime filter and only BTC is traded), NOT a squeeze breakout (L54), NOT a non-price feed gate (L46 — breadth is PRICE-derived, and BTC momentum clears fees on its own; breadth only reduces whipsaw, exactly as the winner's 2nd timeframe does). It is distinct from the in-pipeline BTC dual-timeframe confluence (2nd TIMEFRAME confirm) and the ETH/BTC ratio momentum (a spread) by using a cross-sectional BREADTH regime filter on a single BTC directional book. Multi-instrument data (fills the multi bucket) and long-short (fills the under-weight direction bucket). Three parameters only.

Iteration 2 — addresses the three backtest-review items with the smallest possible change set; signal logic, breadth computation, imports and class structure are untouched, so Layers 1-3 cannot regress. (1) TRADE COUNT: loosened entry_z 0.8 → 0.2 and breadth_threshold 0.75 → 0.55. Re-measured on the full aligned daily catalog (2,143 bars, 2020-09 to 2026-08): 95 trades versus 44, distributed 13/19/9/19/20/15 across 2021-2026, so every walk-forward window and the 15-day holdout now sit inside populated periods. Sharpe is essentially unchanged (0.95 vs 0.96), drawdown 32%, and the gate still earns its keep (removing breadth entirely at these settings drops Sharpe to 0.85 and turns 2024 and 2026 negative). (2) SIZING: replaced the flat 0.6-of-equity notional with volatility targeting (1.5% daily vol target, 0.6 cap). This deliberately shrinks 2021 exposure, where realized vol was ~4-5%/day and the flat sizing was placing its biggest bets; 2021's contribution falls from +84% to +77% while 2022-2024 improve, and daily-return kurtosis drops from 5.9 to 4.7 in my replication. (3) OUTLIER DEPENDENCE: after loosening, deleting each year's single best trade leaves 2021 +11.3%, 2022 +10.0%, 2023 +10.3% and 2024 +9.5% still positive — versus the previous version where 2022 (-2.7%) and 2023 (-4.0%) went negative without their one best trade. That is a real improvement in distribution, and I will not overclaim beyond it: 2025 (-8.4%) and 2026 (-13.0%) do NOT survive removal of their best trade, and one 2021 trade is still +66% on its own. On the reported '+122% in a single day (2021-02-06)': I could not reproduce any intraday move of that size from a ≤0.6× notional position, and I do not believe it is a sizing bug — the backtest's daily-return series attributes a position's ENTIRE realized P&L to its CLOSE date, so a multi-week +66% trend trade lands as one enormous 'day' and inflates return_kurtosis to 38. More trades plus volatility targeting shrink those lumps (which is why kurtosis falls), but the artifact is in the reporting layer, not in the strategy, and I flag it so the analyst does not spend another iteration hunting a bug that is not in this file.

Second iteration; my own trail shows no progress on the decisive concern. Iteration 1 asked to (a) raise trade count, (b) reduce outlier dependence / verify the edge is not carried by 1-2 trades, and (c) fix sizing so one trade cannot produce a >100% equity swing. The developer loosened the gates and added vol-targeting: trade count improved (44->94, still <100) but the outlier concentration REGRESSED — return_kurtosis rose from 38 to 44.7, and the single +134% day (2021-02-06) not only persists but grew, with 2021's entire +135.6% essentially that one day and every other big year carried by 1-2 fat-tail trades. The sizing defect is also unfixed and mathematically evident: a +134% single-day return is impossible from a 0.6x-equity BTC position, so the vol-target cap (min(0.6, target_vol/sigma)) is not constraining the concentration — the +512% / Sharpe 0.79 headline is inflated by a residual sizing/compounding bug, and the sharpe_ci_low 0.18 will not survive deflated-Sharpe. This is the L10/L41 outlier-mirage/overfit signature, and the near-identical BTC vol-regime momentum sibling just collapsed in the exact 3-phase run this would enter (IS Sharpe 7.6 -> OOS -2.1, holdout -0.99); a more outlier-dependent variant would fail the same way. Re-issuing the same outlier/sizing feedback is not warranted when a targeted attempt already failed to move it. Abandon rather than iterate a third time or spend the optimization budget.

Do NOT proceed to optimization yet — the sample is too thin and too outlier-dependent to survive the 3-phase run, and the holdout would be empty. Keep the mechanism on BTC (the survivor) and the alt-breadth gate concept, but fix three things first: (1) TRADE COUNT — the breadth gate produces only 44 trades / ~7 per year, which makes Sharpe/PF statistically unmeasured and guarantees a near-empty 15-day holdout. Loosen breadth_threshold (0.75 is very strict) and/or entry_z so the trade count rises toward >100, so the edge is measurable and the OOS/holdout windows actually contain trades. (2) OUTLIER DEPENDENCE — return_kurtosis is 38 and a single day (2021-02-06, +122%) roughly doubled the account; each positive year is carried by 1-2 trades, which directly contradicts the hypothesis's outlier-robustness claim. After loosening the gate, verify (zero out the top ~5 absolute-return days) that no single year and no 1-2 trades carry the result. (3) SIZING — investigate why a risk_frac=0.6 single BTC position produced a >100% equity move in a day; a 0.6-of-equity long cannot gain 120% of equity unless the position is compounding/concentrating or there is a sizing/compounding bug. Consider volatility-scaling or capping per-trade notional so one trend cannot dominate the whole path. Re-run the backtest and confirm >100 trades with a distributed (low-kurtosis) return profile before requesting optimization.