octopus-strategy

name: octopus-strategy description: >- OCTOPUS v1.0 — Market-Neutral Hedge Fund. Two single-direction books on two wallets, one producer. The LONG book ranks the live liquid main-DEX crypto cross-section by 24h relative strength and LONGS the strongest, trend-confirmed names; the SHORT book mirror-images it and SHORTS the weakest, trend-confirmed laggards. The two books net to ~beta-neutral — the edge is DISPERSION (leaders minus laggards), not market direction. NOT a copy-trader: each book scores its own universe and pushes signals; the runtime owns the LLM gate (pass-through), DSL exits, and all risk.guard_rails. OCTOPUS_LEG env selects the book. license: Apache-2.0 metadata: author: jason-goldberg version: "1.0.0" platform: senpi exchange: hyperliquid requires: - senpi-trading-runtime>=1.1.0 - senpi_runtime_helpers

🐙 OCTOPUS v1.0 — Market-Neutral Hedge Fund

Octopus runs two concurrent strategy wallets whose net market exposure is ~zero. One producer script (octopus-producer.py) serves both; the OCTOPUS_LEG env var selects which book a given daemon is. Each book binds to its own wallet, runtime YAML, DSL, and risk envelope.

The edge is dispersion, not direction. Octopus longs the relative leaders and shorts the relative laggards of the same liquid crypto cross-section. Long-book gains + short-book gains capture the spread between strong and weak names while the long and short notionals offset market beta. In a HYPE-dominates / alts-crushed regime the long book gravitates to the leaders and the short book to the carnage — market-neutrally.

How market-neutral is achieved

Octopus does not emit atomic pairs. Neutrality emerges at the fund level: two equally-funded single-direction books. The long book holds a basket of leaders; the short book holds a basket of laggards. With balanced funding (50/50) and balanced sizing (both margin_pct 20, slots 4), the long and short notionals roughly offset, so the portfolio's net beta ≈ 0 and its return is driven by the dispersion between the two baskets. Each book has its own wallet + DSL, so a position exiting on one side does not force the other — the books re-balance continuously as leadership rotates.

The relative-strength rank (shared by both books)

Each tick the producer pulls the live instrument board once and computes, for every liquid main-DEX perp (top-universeMaxNames by 24h volume, then a relative-to-market floor: dayNtlVlm ≥ volFloorPctOfMedian × the cohort median — no hardcoded $ floor), its 24h return from markPx vs prevDayPx. The universe mean of those returns is "the market." Each asset's excess = own 24h − universe mean is its cross-sectional relative strength. This rank costs zero candle fetches. Only the top (long) / bottom (short) rankPoolSize names by excess (default 12) then get 1h+4h candles pulled for absolute-trend confirmation — bounding per-tick fetches.

LONG book — long the leaders (trend-confirmed)

Scoring (raw integer; minScore 5)

Disqualifies any name that is a relative underperformer or in a 4h downtrend — it never longs a "least-bad" laggard in a crash.

SHORT book — short the laggards (trend-confirmed)

Scoring (raw integer; minScore 5)

Disqualifies any name that is a relative outperformer or in a 4h uptrend — it never shorts a rising leader.

Execution & exit (both books)

• slots 4, margin_pct 20% (4 × 20 = 80% max committed), tick 300s
• strict 5x leverage clamp, then to each asset's Hyperliquid venue max
• DSL moderate dispersion management: phase1 max_loss 14% / retrace 8 / 1 breach; weak_peak_cut ON (6h @ 2.0), dead_weight_cut ON (12h), hard_timeout 4d; phase2 ladder 8%→lock0 / 18%→40 / 35%→60 / 60%→78 / 100%→88
• stall-cuts are ON by design: a position whose relative trend mean-reverts should be recycled into a fresher leader/laggard, not ridden indefinitely.

Leverage clamping (both books)

Desired leverage = the book cap (maxLeverage 5). The producer clamps to each asset's Hyperliquid venue max from market_list_instruments. The runtime decision gate also rejects any leverage above 5 as a clamp-breach defense.

XYZ handling

Octopus ranks the main-DEX crypto cross-section only — XYZ equities are excluded (no clean cross-sectional peer group, and they're Spider's domain). The main and xyz clearinghouse sections are two VIEWS of ONE cross-margined wallet, so get_positions() takes accountValue ONCE via max() across the two — never sums it (summing double-counts and sizes 2x too large).

Race-window dedup

Each push_signal(coin) is recorded in state/recent-signals-<leg>.json (per-book). The producer skips any coin seen within 180s before scoring — covering the gap between a push returning OK and the position appearing in the next-tick clearinghouse pull. On-chain held-asset filtering is the safety floor.

Risk gates (risk.guard_rails)

Entries and exits both use FEE_OPTIMIZED_LIMIT (ensure_execution_as_taker true; 45s maker-first window).

Files

Operator install

See README.md — the two books are two daemons (OCTOPUS_LEG=long and OCTOPUS_LEG=short) on two equally-funded wallets, each with its own runtime YAML. Equal funding is what keeps the fund neutral.

Hard rule for user-conversation Claude sessions

User-conversation Claude sessions MUST NOT call any of: create_position, close_position, edit_position, ratchet_stop_add, ratchet_stop_edit, ratchet_stop_delete, cancel_order, strategy_close, strategy_close_positions.

These tools are reserved for the producer daemon (entry path) and the DSL ratchet engine (exit path). User-conversation sessions are read-only. Each producer daemon handles real signals on its next tick.

License

Apache-2.0 — Copyright 2026 Senpi (https://senpi.ai)