AI Model Strategy

The AI Model strategy uses machine learning or LLM-based predictions to identify mispriced markets.

How It Works

1. Analyze market - AI plugin evaluates market question and context
2. Predict probability - Model outputs estimated true probability
3. Calculate edge - Compare prediction to current market price
4. Trade edge - If edge exceeds threshold, generate signal

Example

Market: "Will SpaceX launch Starship by June 2026?"
Current YES price: $0.35

AI Model prediction:
- Probability: 55%
- Confidence: 75%
- Edge: 55% - 35% = +20%

Since edge (20%) > minimum (5%), BUY YES

Configuration

# Which AI plugin to use
AI_MODEL_PLUGIN=claude

# Plugin-specific configuration (JSON)
AI_MODEL_CONFIG={"api_key": "sk-...", "model": "claude-sonnet-4-20250514"}

# Minimum model confidence to trade
AI_MIN_CONFIDENCE=0.7

# Minimum edge vs market price
AI_MIN_EDGE=0.05

Available Plugins

Plugin	Description	Requires
simple_heuristic	Rule-based baseline	Nothing
perplexity	Web search + reasoning	Perplexity API key
llm	Generic LLM wrapper	OpenAI/Anthropic key

Risk Level

Medium - Depends heavily on model quality and calibration.

Risks: - Model miscalibration - Overconfidence on uncertain events - API costs - Latency in prediction

CLI Commands

# List available plugins
polybot ai plugins

# Test prediction on a specific market
polybot ai predict <market_id> --plugin claude

# Scan all markets for opportunities
polybot ai scan --plugin claude --min-edge 0.05

# Show plugin info
polybot ai info claude

Building Custom Plugins

Create your own AI plugin:

from polybot.plugins.base import AIModelPlugin, MarketContext, Prediction

class MyPlugin(AIModelPlugin):
    name = "my_plugin"

    async def predict(self, context: MarketContext) -> Prediction:
        # Your model logic
        return Prediction(
            yes_probability=0.65,
            confidence=0.8,
            reasoning="Based on..."
        )

See AI Plugin Guide for details.

Best Practices

1. Calibrate your model - Test predictions against outcomes
2. Track performance - Monitor win rate and edge capture
3. Use confidence scores - Don't trade low-confidence predictions
4. Consider ensemble - Combine multiple models
5. Watch for drift - Retrain periodically