trading-signals

<quick_start> Confluence analysis (methodologies agree = high-probability setup):

score = sum(signal.strength * weights[signal.method] for signal in signals)
action = 'BUY' if score >= 0.7 else 'WAIT'

Score interpretation:

• 0.7-1.0: High conviction entry
• 0.4-0.7: Wait for more confluence
• 0.0-0.4: No trade

Cost-effective routing: DeepSeek-V3 for pattern detection → Claude Sonnet for critical decisions </quick_start>

<success_criteria> Analysis is successful when:

• Multiple methodologies provide signals (not just one)
• Regime identified (trending/ranging/volatile) before analysis
• Confluence score calculated with regime-weighted methodology fusion
• Cost-optimized: bulk processing on DeepSeek, critical decisions on Claude
• Clear action (BUY/SELL/WAIT) with supporting rationale
• NO OPENAI used in model routing </success_criteria>

<core_patterns> Standardized patterns for technical analysis across trading projects.

Quick Reference

Methodology	Purpose	When to Use
Elliott Wave	Wave position + targets	Trend structure, cycle timing
Turtle Trading	Breakout system	Trend following
Fibonacci	Support/resistance	Entry/exit zones, golden pocket
Wyckoff	Accumulation/distribution	Institutional activity
Markov Regime	Market state classification	Position sizing, strategy selection
Pattern Recognition	Candlestick + chart patterns	Entry confirmation
Swarm Consensus	Multi-LLM voting	High-conviction decisions

Confluence Detection

When methodologies agree = high-probability setup.

class ConfluenceAnalyzer:
    """Regime-weighted methodology fusion"""

    REGIME_WEIGHTS = {
        'trending_up':   {'elliott': 0.30, 'turtle': 0.30, 'fib': 0.20, 'wyckoff': 0.15},
        'trending_down': {'elliott': 0.30, 'turtle': 0.30, 'fib': 0.20, 'wyckoff': 0.15},
        'ranging':       {'fib': 0.35, 'wyckoff': 0.30, 'elliott': 0.20, 'turtle': 0.05},
        'volatile':      {'fib': 0.30, 'wyckoff': 0.30, 'elliott': 0.20, 'turtle': 0.10},
    }

    def analyze(self, df, regime: str) -> dict:
        weights = self.REGIME_WEIGHTS[regime]
        signals = self._collect_signals(df)

        score = sum(s.strength * weights[s.method] for s in signals)
        return {
            'score': score,  # 0-1.0
            'action': 'BUY' if score >= 0.7 else 'WAIT',
            'confluence': self._calc_agreement(signals)
        }

Score Interpretation:

• 0.7-1.0: High conviction entry
• 0.4-0.7: Wait for more confluence
• 0.0-0.4: No trade

File Structure

trading-project/
├── methodologies/
│   ├── elliott_wave.py     # Wave detection + halving cycle
│   ├── turtle_system.py    # Donchian breakouts
│   ├── fibonacci.py        # Levels + golden pocket
│   ├── wyckoff.py          # Phase detection + VSA
│   └── markov_regime.py    # State classification
├── patterns/
│   ├── candlestick.py      # Engulfing, hammer, doji
│   └── chart_patterns.py   # H&S, double bottom, triangles
├── aggregator.py           # Regime-weighted fusion
└── swarm/
    ├── consensus.py        # Multi-LLM voting
    └── adapters/           # Claude, DeepSeek, Gemini

Cost-Effective Model Routing

Task	Model	Cost
Pattern detection	DeepSeek-V3	$0.27/1M
Confluence scoring	Qwen-72B	$0.40/1M
Critical decisions	Claude Sonnet	$3.00/1M
Swarm consensus	Mixed tier	~$1.50/1M avg

Integration Notes

• Data Sources: yfinance, CCXT, Alpaca API
• Pairs with: runpod-deployment-skill (model serving)
• Projects: ThetaRoom, swaggy-stacks, alpha-lens

Reference Files

Core Methodologies:

• reference/elliott-wave.md - Wave rules, halving supercycle, targets
• reference/turtle-trading.md - Donchian channels, ATR sizing, pyramiding
• reference/fibonacci.md - Levels, golden pocket, on-chain enhanced
• reference/wyckoff.md - Phase state machines, VSA, composite operator
• reference/markov-regime.md - 7-state model, transition probabilities

Advanced Patterns:

• reference/pattern-recognition.md - Candlestick + chart patterns
• reference/swarm-consensus.md - Multi-LLM voting system
• reference/chinese-llm-stack.md - Cost-optimized Chinese LLMs for trading