Octopus Strategy

🐙 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.

| Book | Style | Wallet env | Runtime | Scanner | |---|---|---|---|---| | long | Cross-sectional relative-strength LONG | OCTOPUS_LONG_WALLET | runtime-long.yaml | octopus_long_signals | | short | Cross-sectional relative-weakness SHORT | OCTOPUS_SHORT_WALLET | runtime-short.yaml | octopus_short_signals |

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)

| Component | Pts | Source | |---|---|---| | Relative strength (excess) | +3 (≥2×rsThresholdPct) / +2 (≥1×) / +1 (≥0) / disqualify ( rsiOverbought 80) | 1h RSI |

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)

| Component | Pts | Source | |---|---|---| | Relative weakness (excess) | +3 (≤−2×rsThresholdPct) / +2 (≤−1×) / +1 (≤0) / disqualify (>0) | 24h excess vs cross-section | | 4h trend structure | +2 BEARISH / disqualify BULLISH | 4h candles | | 1h trend confirmation | +1 BEARISH / −1 BULLISH | 1h candles | | Own absolute momentum | +1 (24h ≤ 0) / −1 (> 0) | instrument ctx | | RSI capitulation guard | −2 (RSI .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)

| Gate | long | short | |---|---|---| | dailylosslimitpct | 12 | 12 | | maxentriesperday | 6 | 6 | | maxconsecutivelosses | 4 | 4 | | cooldownminutes | 60 | 60 | | drawdownhaltpct | 20 | 20 | | perassetcooldownminutes | 180 | 180 | | dataretentionhours | 120 | 120 | | drawdownresetondayrollover | true | true |

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

Files

| File | Purpose | |---|---| | runtime-long.yaml | Long-book runtime spec (wallet, DSL, risk, LLM gate) | | runtime-short.yaml | Short-book runtime spec | | scripts/octopus-producer.py | Book-aware producer daemon (one script, both books) | | scripts/octopus_config.py | Leg resolution + SenpiClient wrapper + helpers | | config/octopus-long-config.json | Long-book tunables (universe floor, RS threshold) | | config/octopus-short-config.json | Short-book tunables |

Operator install

See [README.md](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)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Senpi-ai
• Source: Senpi-ai/senpi-skills
• License: MIT

Install and usage instructions live in the source repository linked above.