12-Instrument Crypto Universe Cross-Sectional Momentum Long-Short (Weekly)

An EXPANDED DOLLAR-NEUTRAL CROSS-SECTIONAL MOMENTUM strategy on a 12-instrument crypto universe spanning multiple sectors (L1 majors, oracle infra, exchange tokens, payments, meme/retail). On each weekly close, rank all 12 instruments by trailing 4-week return; LONG the top-3 performers and SHORT the bottom-3 performers, equal-weighted, dollar-neutral. Rebalance weekly. RATIONALE driven by the now-DECISIVE session finding and the analyst's specific recommendation: the prior 5-coin L1 cross-sectional momentum strategy failed BECAUSE 'Adding more basket members would help (10+ alts with more dispersion) but exceeds the major L1 framing.' This proposal EXPLICITLY tests the expanded version — 12 instruments spanning multiple sectors to maximize return dispersion that drives cross-sectional momentum edge. KEY DIFFERENTIATION from the failed 5-coin L1 basket: (a) 12 instruments vs 5 — more than DOUBLES the universe; (b) Multi-sector spanning L1 (BTC/ETH/SOL/BNB/ADA/AVAX), oracle (LINK), payments (LTC, XRP), DeFi/meme (DOT/DOGE/ATOM) — much higher CROSS-SECTOR dispersion than within-L1 ranking; (c) Top-3 vs Top-2 — 3 legs per side means each instrument has more weight differentiation; (d) Same rebalance frequency (weekly) but with 6 legs (3 long + 3 short) instead of 4 — slightly more turnover but more diversification per side. Academic literature on cross-sectional momentum (Jegadeesh-Titman 1993; Asness-Moskowitz-Pedersen 2013) requires sufficient cross-sectional dispersion to identify persistent winners/losers. 12-instrument crypto universe has documented correlation range 0.4-0.9 (depending on regime), with non-L1 instruments often DECOUPLING from L1 majors during sector rotations — the source of edge. KEY STRUCTURAL DIFFERENCES from all 24+ failed strategies: (a) Not directional alt-perp trend-following; (b) Not cross-asset pair MR (multi-leg basket dilutes ratio-drift signature); (c) Not sub-daily MR; (d) Not vol-targeting overlay; (e) NEW: expanded universe addresses the SPECIFIC failure diagnosis of the L1 basket. MAXIMALLY SIMPLE — 2 parameters (ranking lookback weeks=4, top/bottom k=3). ZERO supplementary data dependency.

Iteration-3 fixes the liquidation / 552% exposure caused by the gross-cap being enforced against a per-leg COUNT times a FIXED leg_notional rather than realized open notional. Under Nautilus netting, repeated entries inflate ONE netted position's quantity, invisible to the old accounting. The rewritten _converge now: (1) computes realized gross as the ACTUAL Sum(|qty_i * price_i|) from real position quantities each bar; (2) treats a correct-side leg as 'at target' only when it is a single, non-oversized position (<= leg_notional_tol x leg_notional) and CLOSES any oversized/stacked/wrong-side leg, which unwinds a runaway netted position back toward target; (3) re-opens a leg ONLY when it is fully flat (no positions AND no in-flight orders), respecting the one-bar lag of positions_open after a market exit; (4) gates opens so Sum(|open notional|) <= gross_exposure x equity, with a hard-reset flatten if realized gross ever exceeds gross_hard_mult x cap and a halt-flatten when equity <= 0. Leverage remains 1.0 and is read from config into sizing, so single-week returns cannot exceed ~+/- gross_exposure x max single-leg move. Only _converge and two guard params changed; ranking/signal scaffolding that passed earlier layers is untouched.

Liquidated for the THIRD time (total_return -100%, max_drawdown 100%, liquidation_date 2023-12-25) with exposure_pct 524% against a 90% gross target — i.e. the book runs ~5.8x intended exposure despite the iteration-3 fix that was written specifically to solve this exact bug (realized summed-notional reconciliation, per-leg oversize close-out, hard gross cap, 1.25x-cap flatten). The targeted fix left the identical symptom, which means the multi-instrument gross-exposure control does not hold under Nautilus netting in this construction, and three iterations of engineering have not fixed it — continuing to iterate on the same liquidation/plumbing bug is diminishing returns. Independently, there is no edge to recover even if sizing were fixed: the pre-liquidation window is already unprofitable (profit_factor 0.88, expectancy -$304/trade, sharpe -0.27, 2022 -7.7%, 2023 -5.5%), and this is an explicit re-try of the already-falsified 5-coin L1 cross-sectional momentum basket expanded to 12 instruments — cross-sectional momentum on a highly-correlated crypto universe has failed repeatedly this session, and the expansion cannot even be assessed through a liquidated run (return_skew -18, kurtosis 523, single-period losses of -74% and -86%). Not worth optimizing and not worth a 4th iteration: the exposure control is unreliable AND the mechanism shows negative expectancy. If cross-sectional crypto momentum is pursued again it should be a fresh hypothesis on a spot/CASH or hard-capped-notional venue where >100% gross is impossible by construction, with the signal's edge demonstrated on a correctly-sized book before any expansion.

The account is LIQUIDATED and exposure_pct is 552% against an intended gross_exposure of 0.90 — the _converge state machine is NOT enforcing the dollar-neutral gross cap; positions are accumulating to ~6x the 6-leg target. Do NOT optimize until this is fixed. Specific fixes: (1) The close-then-settle-then-reopen cadence lets stale legs pile up across weekly bars while exits are still in flight — the held_legs/gross accounting is computed on target legs but real open notional balloons because old positions have not actually closed. Reconcile against the ACTUAL summed open notional Sum(|qty_i * price_i|) each bar, not against a per-leg count times a fixed leg_notional. (2) Enforce the gross cap on realized open notional: if Sum(|open notional|) already exceeds gross_exposure*equity, close excess before opening anything. (3) Do not open a new leg for an instrument that still has ANY open position or in-flight order — verify the flat/settled state before re-entry (positions_open can lag one bar after a market exit). (4) Verify leverage=1.0 is actually plumbed and that the two single-week -75%/-144% returns cannot recur — after the fix, no weekly return should be able to exceed roughly +/- gross_exposure * (max single-leg move). (5) Re-run the backtest and confirm exposure_pct sits near 90% and there is no liquidation BEFORE resubmitting. Once neutral gross is verified, the momentum edge can be judged fairly.