tiger-strategy（老虎策略）

TIGER v2 — Multi-Scanner Goal-Based Trading

5 scanners. 1 goal. Configurable aggression. Mechanical exits.

Philosophy: WOLF hunts on instinct. TIGER calculates what it needs, then hunts exactly that much. Give it a budget, a target, and a deadline — it adjusts aggression automatically.

Architecture

┌──────────────────────────────────────────┐
│           10 OpenClaw Crons              │
│  Compress(5m) Corr(3m) Momentum(5m)     │
│  Reversion(5m) Funding(30m) OI(5m)      │
│  Goal(1h) Risk(5m) Exit(5m) DSL(30s)    │
├──────────────────────────────────────────┤
│           Python Scripts                  │
│  tiger_lib.py  tiger_config.py           │
│  5 scanners / goal-engine / risk /       │
│  exit / oi-tracker / dsl-v4              │
├──────────────────────────────────────────┤
│           Senpi MCP (via mcporter)        │
│  market_list_instruments                  │
│  market_get_asset_data / market_get_prices│
│  create_position / close_position         │
│  edit_position / cancel_order             │
│  strategy_get_clearinghouse_state         │
│  leaderboard_get_markets                  │
│  account_get_portfolio                    │
├──────────────────────────────────────────┤
│           State Files                     │
│  tiger-config.json → tiger_config.py      │
│  state/{instance}/*.json (atomic writes)  │
└──────────────────────────────────────────┘

State flow: OI Tracker samples all assets → Scanners score signals by confluence → Goal Engine sets aggression → Agent enters via create_position → DSL manages trailing stops → Risk Guardian enforces limits → Exit Checker handles pattern-specific exits.

Skill Attribution

When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.

Quick Start

1. Ensure Senpi MCP is connected (mcporter list shows senpi)
2. Create a custom strategy: strategy_create_custom_strategy
3. Fund the wallet: strategy_top_up
4. Run setup: python3 scripts/tiger-setup.py --wallet 0x... --strategy-id UUID \ --budget 1000 --target 2000 --deadline-days 7 --chat-id 12345
5. Create 10 OpenClaw crons from references/cron-templates.md

First hour: OI Tracker needs ~1h of history before compression/reversion scanners can use OI data. Goal engine and risk guardian work immediately.

5 Signal Patterns

1. Compression Breakout (Primary)

BB squeeze with OI accumulation → price breaks bands.

2. BTC/ETH Correlation Lag (v2 — Multi-Leader, Multi-Window)

Leader (BTC or ETH) moves significantly → high-corr alts haven't caught up.

Multi-window detection: Checks 1h, 4h, 12h, and 24h rolling windows with scaled thresholds (1×, 1×, 1.5×, 2× base threshold). Catches sustained moves that single-candle checks miss.

Multi-leader: BTC and ETH are both tracked as leaders. ETH uses 0.8× threshold multiplier (more volatile). Each leader has its own alt list:

• BTC alts: SOL, DOGE, AVAX, LINK, ADA, DOT, NEAR, ATOM, FIL, INJ, SEI, SUI, TIA, JUP, WIF, PEPE, RENDER, FET, TAO, AR
• ETH alts (ecosystem-specific): OP, ARB, MATIC, STRK, AAVE, UNI, LDO, SNX, MKR, ENS, PENDLE, EIGEN, ENA, ETHFI, DYDX, CRV, COMP, SUSHI, IMX, BLUR

Smart dedup: If BTC and ETH move in the same direction, only BTC's alts are scanned (avoids redundancy). Max 4 alts total per scan to stay within 55s timeout.

Window quality: STRONG (lag > 0.7), MODERATE (0.5-0.7), CLOSING (0.4-0.5).

3. Momentum Breakout

Strong price move with volume confirmation.

DSL note: Tighter Phase 1 retrace (0.012) — momentum reversals are fast.

4. Mean Reversion

Overextended asset with exhaustion signals → counter-trend.

5. Funding Rate Arb

Extreme funding → go opposite the crowd, collect income.

DSL note: Wider retrace tiers (0.02+) — edge is income, not price direction. Risk Guardian auto-exits if funding flips.

Goal Engine & Aggression

goal-engine.py runs hourly. Calculates required daily return and sets aggression:

DSL v4 — Trailing Stop System

Per-position DSL state file. Combined runner (dsl-v4.py) checks all active positions every 30s.

IMPORTANT: The DSL cron must first check activePositions in TIGER state. If no positions are open, output HEARTBEAT_OK immediately and do NOT invoke dsl-v4.py. This prevents unnecessary session spam when TIGER is idle.

Phase 1 (pre-Tier 1): Absolute floor. 3 consecutive breaches → close. Max duration: 90 minutes.

Phase 2 (Tier 1+): Trailing tiers.

Stagnation TP: ROE ≥ 8% + high-water stale 1h → auto-close.

DSL Tuning by Pattern

Risk Management

All percentage values are whole numbers (5 = 5%).

Anti-Patterns

1. NEVER enter in ABORT aggression. Goal engine set ABORT for a reason.
2. NEVER override DSL. DSL auto-closes. Don't re-enter after DSL exit.
3. NEVER hold FUNDING_ARB after funding flips. The thesis is dead.
4. NEVER chase momentum after 2h. If you missed the 1h move, wait for the next one.
5. NEVER enter reversion without 4h RSI extreme. That's the required filter, not optional.
6. NEVER run scanners without timeout wrapper. timeout 55 prevents overlap.
7. NEVER mix xyz: assets into main scanner groups. Prescreener isolates them into group_xyz; scanners hard-filter them out.

Resilience Architecture

xyz: Hard Isolation (3 layers)

1. Prescreener: Separates xyz: prefixed assets into group_xyz, never mixed with main groups a-f
2. Scanner filtering: Every scanner has names = {n for n in names if not n.startswith("xyz:")} guard
3. Per-asset try/except: Even if an xyz: asset leaks through, it's caught by the per-asset error handler

Per-Asset Error Isolation

All 4 scanners (compression, correlation, momentum, reversion) wrap their per-asset analysis loops in try/except. One bad asset (missing data, API error, unexpected format) logs the error and skips — the rest of the scan continues normally. This prevents cascade failures where one problematic asset kills an entire scan run.

API Dependencies

State Schema

See references/state-schema.md for full schema with field descriptions.

Key state files:

state/{instanceKey}/
├── tiger-state.json              # Positions, aggression, safety, daily stats
├── dsl-{ASSET}.json              # Per-position DSL trailing stop state
├── oi-history.json               # 24h OI time-series
├── trade-log.json                # All trades with outcomes
└── scan-history/                 # Scanner output history

All state files include version, active, instanceKey, createdAt, updatedAt. All writes use atomic_write().

Cron Setup

See references/cron-templates.md for ready-to-use OpenClaw cron payloads.

Silence Policy — CRITICAL: When a cron fires and the result is HEARTBEAT_OK, NO_POSITIONS, or no actionable signals:

• Do NOT notify Telegram
• Do NOT reply in chat
• Do NOT explain what the cron did
• Do NOT summarize the scan results
• Produce NO output of any kind — complete silence Only speak (chat or Telegram) when something actionable happens: trade opened, trade closed, aggression changed, risk halt triggered, or an error that needs attention. Idle cycles = total silence.

All crons run isolated. No main session crons. No narration during scan cycles. The user should never see HEARTBEAT_OK, scanner reasoning, or "systems nominal" messages.

Tier 1 (fast/cheap): threshold checks, data collection, DSL math. Runs isolated with delivery.mode: "none" and explicit model (the configured model). Tier 2 (capable): aggression decisions, risk judgment, exit evaluation. Runs isolated with delivery.mode: "announce" and explicit model (the configured model). OpenClaw auto-suppresses HEARTBEAT_OK — only real content gets delivered. DSL (Cron 10): Runs in main session (systemEvent) — needs position state context.

Scanners are staggered by 1-2 minutes to avoid mcporter rate limits (see cron-templates.md).

ROAR — Recursive Optimization & Adjustment Runtime

ROAR is TIGER's meta-optimizer. It runs every 8 hours (+ ad-hoc every 5th trade), analyzes TIGER's trade log, and tunes execution parameters within bounded ranges. User intent (budget, target, risk limits) is never touched.

What ROAR tunes (within hard min/max bounds):

• Per-pattern confluence thresholds (0.25–0.85)
• Scanner thresholds (BB squeeze percentile, BTC correlation move, funding annualized)
• DSL retrace thresholds per phase (0.008–0.03)
• Trailing lock percentages per aggression level

What ROAR never touches (protected): budget, target, deadline, max_slots, max_leverage, maxDrawdownPct, maxDailyLossPct, maxSingleLossPct.

Rules engine (6 rules):

1. Win rate < 40% over 10+ trades → raise pattern confluence threshold by 0.05
2. Win rate > 70% over 10+ trades → lower threshold by 0.03 to catch more signals
3. Avg DSL exit tier < 2 → loosen phase1 retrace by 0.002 (let positions run)
4. Avg DSL exit tier ≥ 4 → tighten phase1 retrace by 0.001 (lock gains)
5. No entries in 48h for a pattern with 5+ trades → lower threshold by 0.02
6. Negative expectancy over 20+ trades → disable pattern for 48h (auto-re-enables)

Safety: revert-if-worse checks every cycle. If both win rate AND avg PnL degraded since last adjustment, auto-reverts to previous config.

Scripts: roar-analyst.py (engine), roar_config.py (bounds, state, revert logic).

Expected Performance

Known Limitations

• OI history bootstrap. Scanners need ~1h of OI data before OI-dependent signals are reliable.
• mcporter latency. ~6s per call. Scanners limited to 8 assets per cycle.
• DSL is per-position. Each position needs its own DSL state file.
• Correlation scanner tracks BTC + ETH as leaders. Doesn't detect moves led by other assets (e.g., SOL-led rallies).
• Funding arb needs patience. Edge is income over time; DSL must be wide enough.
• Goal engine recalculates hourly. Aggression can shift mid-trade.

Optimization Levers

Gotchas

• maxSingleLossPct is a whole number: 5 = 5%.
• minConfluenceScore values are decimals (0.40 = 40%), NOT whole numbers — this is a weighted score 0-1.
• trailingLockPct values are decimals (0.60 = lock 60%).
• triggerPct in DSL tiers is ROE % (5 = 5% ROE), not price %.
• lockPct in DSL is % of high-water move to lock, not a retrace threshold.
• DSL reads DSL_STATE_FILE env var ONLY — positional args are silently ignored.
• timeout 55 on all scanner scripts to prevent cron overlap.
• Cron stagger offsets::00 compression,:01 momentum,:02 reversion,:03 OI,:04 risk+exit.

Lessons from Live Trading

Operational

• DSL state file active field: MUST include active: true or dsl-v4.py returns {"status": "inactive"} (line 22 check). This is the #1 gotcha when setting up new positions.
• DSL invocation syntax: DSL_STATE_FILE=/path/to/file.json python3 scripts/dsl-v4.py COIN
• API latency: market_get_asset_data ~4s/call, market_list_instruments ~6s. Max 8 assets per 55s scan window. This is NORMAL — design around it, don't blame it.
• Correlation scanner timeouts: Max 4 alts per scan (down from 6) to stay within 55s. Cron interval tightened to 3min for time-sensitive lag trades.
• Compression scanner signals: Requires breakout: true AND a direction to be actionable — a high compression score alone is not enough.
• Test exit scripts after EVERY edit: A crashing exit script is invisible and bleeds real money (lost ~$43 on PUMP from UnboundLocalError).
• Config changes mean nothing if code doesn't read them: Always verify the consuming code actually loads changed config keys.
• Exchange SLs FIRST, DSL state SECOND: Set exchange-level SL via edit_position immediately after entry. DSL is intelligence, exchange SL is safety net.
• Zombie processes from frequent cron spawns: Can accumulate under PID 1 and cause fork failures. Monitor and restart container if needed.
• mcp-remote stdio proxy is unreliable: Use direct streamable-http transport instead.

Trading

• Don't short compressed assets with building OI — compression often resolves upward.
• No duplicate positions: Skip signals for assets already in active_positions.
• Re-entry in opposite direction IS valid: When signals are strong, entering the same asset in the opposite direction works.
• DSL trailing stops >> fixed TP: Every winning trade ran past where a fixed TP would have closed. Let winners run.
• High-score signals (0.85+) justify overriding blacklists: If original loss was small and new direction differs, take the trade.
• create_position format: Requires orders array with coin, direction, leverage, marginAmount, orderType fields.
• ALO fee optimization: Use orderType: "FEE_OPTIMIZED_LIMIT" + ensureExecutionAsTaker: true for entries AND non-urgent exits (stagnation, time-stop, daily target). Saves ~8 bps per fill. Use MARKET ONLY for: DSL breach closures, risk guardian hard stops, daily loss limit hits, and exchange-level SLs. Exchange SL orderType in edit_position must always be MARKET.
• close_position syntax: mcporter call 'senpi.close_position(...)'
• CLOSE_NO_POSITION pattern: Position may already be closed on-chain before DSL's close call — handle gracefully (not an error).
• Position sizing: think in MARGIN first: per_slot_budget = 30% of balance as margin. Multiply by leverage for notional. Old formula (balance÷slots÷leverage) left $730 on the table on PUMP.
• Fees at 10x eat 1.8% of margin per round-trip: Track balance from clearinghouse accountValue, NOT price deltas.
• 24h correlation signals go stale when 4h diverges: Always check timeframe alignment. Lost ~$90 on SEI/OP/ARB LONG trades that reversed in 1 hour.
• Volume ratio < 1.2 = no conviction for momentum entries: XPL at 0.42 volume ratio stopped out immediately. Hard gate in config.
• RSI > 70 on entry = red flag: FARTCOIN entered at RSI 70.4, lost -13.7% ROE. Tightened RSI gates.
• LIMIT SL on exchange = unreliable: Always use orderType: "MARKET" for stop losses. LIMIT can fail on fast moves.
• Realistic targets beat moonshots: Mission 1 (18.5% in 2.5 days, ~3%/day) succeeded. Mission 3 (8%/day) ABORTed. Target 1.5-3%/day max.
• 2 slots > 3 slots at $3K: Concentration beats diversification at small capital. Each position gets meaningful margin.