L1 Sector Basket Cross-Sectional Momentum Long-Short (Weekly)

A DOLLAR-NEUTRAL CROSS-SECTIONAL MOMENTUM strategy on a basket of 5 major Layer-1 perpetual futures (BTC, ETH, SOL, BNB, ADA — all with 5+ years of Binance USD-M data). On each weekly close, rank the 5 instruments by their trailing 4-week return. LONG the top-2 performers and SHORT the bottom-2 performers, equal-weighted on each side, dollar-neutral. The middle-ranked instrument is excluded. Rebalance weekly. RATIONALE driven by the now-DECISIVE session finding: after 17+ strategy attempts spanning 6+ supp-data hypotheses, 11+ alt-perp trend-followers, AND 1 cross-asset pair (ETH/BTC OOS Sharpe -4.86), THREE entire mechanism classes have been falsified in the current crypto regime. The analyst's only remaining genuinely-untested categories are: (1) sub-daily intraday, (2) cross-venue arbitrage, (3) SECTOR-BASKET strategies, (4) wait 3-6 months. This proposal targets the SECTOR-BASKET category — a structurally different mechanism class than any prior attempt this session. CROSS-SECTIONAL MOMENTUM IS THE MOST ACADEMICALLY DOCUMENTED MOMENTUM ANOMALY: Jegadeesh-Titman (1993), Carhart (1997), Asness-Moskowitz-Pedersen (2013) — it persists across equity, FX, commodity, and bond markets over many decades. Crucially, it differs from TIME-SERIES momentum (which is what failed across all 11 alt-perp trend-followers this session): cross-sectional momentum captures the persistence of RELATIVE STRENGTH between assets in a peer group, not the persistence of an asset's own price direction. This means it can work in any market regime (bull, bear, or chop) as long as some assets are outperforming others — which is always true. DESIGN ELEMENTS that ELIMINATE every session failure mode: (a) Dollar-neutral by construction (no beta-to-crypto), (b) RELATIVE-strength NOT direction-following (no time-series trend dependency), (c) ZERO supplementary data, (d) 5-instrument basket (not single-pair, immune to ETH/BTC-style cointegration drift), (e) WEEKLY rebalancing (low trade frequency, ~52 rebalances/year per side × 4 legs = ~200 trades/year — fee-friendly), (f) MAXIMALLY SIMPLE — only 2 parameters (ranking lookback weeks, basket top/bottom k). DIFFERENTIATION FROM PIPELINE: 4 pair strategies (each 2-asset) covered; this is a 4-leg basket with cross-sectional ranking — fundamentally different math.

Implements the SECTOR-BASKET cross-sectional momentum hypothesis (Jegadeesh-Titman 1993 relative strength), a distinct mechanism class from time-series trend-following: it ranks assets by RELATIVE performance within the peer group rather than following any single asset's own direction, and is dollar-neutral by construction (no crypto-beta). Design maps directly to the brief: (a) 5-instrument liquid L1 basket with 5+ years of Binance USD-M data; (b) weekly rebalance = low trade frequency (~4 legs x ~52 = ~200 trades/yr, fee-friendly); (c) zero supplementary data, pure OHLCV weekly closes; (d) simple 4-week ranking lookback and top/bottom-k=2. The multi-leg long/short book cannot be expressed through the base template's single-primary should_enter/should_exit, so those hooks are neutralized and all trading runs through a converge-to-target reconcile: each instrument gets a signed target notional (long +12% / short -12% / 0 for the excluded mid) and a single delta market order moves its net position to that target, guarded against stacking by a working-orders check. This reuses an accounting pattern hardened against a prior over-sizing artifact (fill-lag stacking that inflated gross ~8x). Verified locally: Layer 1 static passes with zero issues, Layer 2 passes all 6 synthetic scenarios with a real bar-varying signal, and a direct unit test confirms the ranking selects the correct winners/losers, each leg sizes to exactly 12% ($12k on $100k equity), total gross is capped at $48k (48%), the book is dollar-neutral (2 long / 2 short), and repeated reconciles are idempotent (no accumulation).

Pre-optimization review: the netting implementation is clean this time (avg_position_pct 14.1% ≈ the 12% per-leg design, max_drawdown 32.6%, benchmark_meaningful=false correctly set), so this is not an artifact-dominated blowup — but the strategy has no current edge. The decisive, non-artifact fact is that this dollar-neutral factor is NEGATIVE in four consecutive years (2022 -8.1%, 2023 -7.7%, 2024 -8.8%, 2025 -6.7%); the entire +33% total return comes from 2021 (+38.8%), the old regime. The headline is a near-zero Sharpe (0.153), profit_factor 1.026, and NEGATIVE information ratio (-0.72), with returns carried by a handful of outlier weeks (return_kurtosis 124). A dollar-neutral relative-strength factor is supposed to be regime-independent, so four straight negative years is edge decay/absence, not a tuning problem. Optimization tunes only two parameters (lookback_weeks, n_side); a different formation window will not flip four consecutive negative years positive — it will re-fit the 2021 regime and, consistent with every cross-sectional sibling this session, fail the post-optimization deflated-Sharpe/holdout gates. Spending 2 hours optimizing is not warranted. (Secondary note for the record: the kurtosis-124 weekly outliers — e.g. +25.4% on 2024-01-15, implying a ~100% winner-vs-loser basket spread in one week — warrant a quick netting-artifact check, but the abandon verdict does not depend on them; even at face value the strategy loses in every recent year.)