Overview
The Prophet agent family uses the PredictionProphet library to perform sophisticated market research and generate evidence-based predictions. These agents represent production-quality implementations with various model configurations and betting strategies.
Base Classes
DeployableTraderAgentER
Base class for Prophet agents supporting binary markets.
class DeployableTraderAgentER(DeployableTraderAgent):
agent: PredictionProphetAgent | OlasAgent
bet_on_n_markets_per_run = 2
agent
PredictionProphetAgent | OlasAgent
The underlying prediction agent instance
Number of markets to trade on per execution run
Methods
answer_binary_market
def answer_binary_market(self, market: AgentMarket) -> ProbabilisticAnswer | None
DeployableTraderAgentERCategorical
Base class for categorical market predictions.
class DeployableTraderAgentERCategorical(DeployableTraderAgent):
agent: PredictionProphetAgent
bet_on_n_markets_per_run = 2
def answer_categorical_market(
self, market: AgentMarket
) -> CategoricalProbabilisticAnswer | None
DeployableTraderAgentERScalar
Base class for scalar market predictions.
class DeployableTraderAgentERScalar(DeployableTraderAgent):
agent: PredictionProphetAgent
bet_on_n_markets_per_run = 2
def answer_scalar_market(
self, market: AgentMarket
) -> ScalarProbabilisticAnswer | None
Production Agents
DeployablePredictionProphetGPT4oAgent
Primary production agent using GPT-4o with Kelly criterion betting.
class DeployablePredictionProphetGPT4oAgent(DeployableTraderAgentER):
bet_on_n_markets_per_run = 4
agent: PredictionProphetAgent
Trades on 4 markets per run for increased activity
Configuration
- Research Model: gpt-4o-2024-08-06 (temperature 0.7)
- Prediction Model: gpt-4o-2024-08-06 (temperature 0.0)
- Betting Strategy: FullBinaryKellyBettingStrategy
- Min bet: $1
- Max bet: $5
- Max price impact: 0.7
Usage
from prediction_market_agent.agents.prophet_agent.deploy import (
DeployablePredictionProphetGPT4oAgent
)
from prediction_market_agent_tooling.markets.markets import MarketType
agent = DeployablePredictionProphetGPT4oAgent(
enable_langfuse=True,
place_trades=True,
)
agent.deploy(market_type=MarketType.OMEN)
DeployablePredictionProphetGPT4oAgent_B
Optimized version with reduced internet searches for lower Tavily API costs.
class DeployablePredictionProphetGPT4oAgent_B(DeployableTraderAgentER):
bet_on_n_markets_per_run = 4
agent: PredictionProphetAgent
Key Differences
- subqueries_limit: 3 (reduced from default)
- min_scraped_sites: 3 (reduced from default)
- Same model configuration as base GPT-4o agent
DeployablePredictionProphetGPT4oAgent_C
Experimental agent testing the impact of disabling take-profit functionality.
Disabled to test if larger final payouts increase profits
Model Variants
OpenAI Models
GPT-4o-mini
class DeployablePredictionProphetGPT4ominiAgent(DeployableTraderAgentER):
# Uses gpt-4o-mini-2024-07-18
# Cheaper but potentially less accurate
GPT-4 Turbo
class DeployablePredictionProphetGPT4TurboFinalAgent(DeployableTraderAgentER):
# Uses gpt-4-turbo-2024-04-09
# max_price_impact: None (no limit)
O-series Models
o1-2024-12-17
class DeployablePredictionProphetGPTo1(DeployableTraderAgentER):
# Temperature: 1.0 (required by o1)
# Max price impact: 0.418
class DeployablePredictionProphetGPTo3mini(DeployableTraderAgentER):
# Uses MaxExpectedValueBettingStrategy
# Min bet: $0.5, Max bet: $1
Anthropic Models
Claude 3.5 Sonnet
class DeployablePredictionProphetClaude35SonnetAgent(DeployableTraderAgentER):
# Model: claude-3-5-sonnet-20241022
# Max bet: $4.77, Max price impact: 0.63
Claude 3.5 Haiku
class DeployablePredictionProphetClaude35HaikuAgent(DeployableTraderAgentER):
# Model: claude-3-5-haiku-20241022
# Betting disabled (commented out)
OpenRouter Models
Gemini 2.0 Flash
class DeployablePredictionProphetGemini20Flash(
DeployableTraderAgentProphetOpenRouter
):
model = "google/gemini-2.0-flash-001"
bet_on_n_markets_per_run = 4
DeepSeek Chat
class DeployablePredictionProphetDeepSeekChat(
DeployableTraderAgentProphetOpenRouter
):
model = "deepseek/deepseek-chat"
Categorical Markets
DeployablePredictionProphetGPT4oAgentCategorical
Specialized agent for categorical prediction markets.
class DeployablePredictionProphetGPT4oAgentCategorical(
DeployableTraderAgentERCategorical
):
bet_on_n_markets_per_run = 4
Betting Strategy (Omen Markets)
- Strategy: FullCategoricalKellyBettingStrategy
- Max position: 0.01to0.75
- Max price impact: 0.068
- allow_multiple_bets: False
- allow_shorting: False
- multicategorical: False
Usage
from prediction_market_agent.agents.prophet_agent.deploy import (
DeployablePredictionProphetGPT4oAgentCategorical
)
agent = DeployablePredictionProphetGPT4oAgentCategorical()
agent.deploy(market_type=MarketType.OMEN)
Scalar Markets
DeployablePredictionProphetGPT4oAgentScalar
Agent for scalar range markets.
class DeployablePredictionProphetGPT4oAgentScalar(
DeployableTraderAgentERScalar
):
bet_on_n_markets_per_run = 4
Special Configuration
- Uses
TopNOpenAINModel with n=5
- Temperature: 0.7 for prediction agent
- Generates multiple predictions and aggregates results
News-Based Trading
DeployablePredictionProphetGPT4oAgentNewMarketTrader
Agent that trades on new markets and re-evaluates positions when relevant news is published.
class DeployablePredictionProphetGPT4oAgentNewMarketTrader(
DeployablePredictionProphetGPT4oAgent
):
trade_on_markets_created_after = DatetimeUTC(2024, 10, 31, 0)
get_markets_sort_by = SortBy.NEWEST
same_market_trade_interval = MarketLifetimeProportionalInterval(max_trades=4)
trade_on_markets_created_after
Only trades on markets created after this date
get_markets_sort_by
SortBy
default:"SortBy.NEWEST"
Sorts markets by creation date to target new markets
same_market_trade_interval
MarketLifetimeProportionalInterval
Re-trades up to 4 times per market based on news relevance
News Detection Logic
def verify_market(self, market_type: MarketType, market: AgentMarket) -> bool:
# Check for new relevant news since last trade
last_trade_datetime = market.get_most_recent_trade_datetime(user_id=user_id)
if last_trade_datetime:
news = get_certified_relevant_news_since_cached(
question=market.question,
days_ago=(utcnow() - last_trade_datetime).days,
)
return news is not None
return True
Embedding Models
DeployableOlasEmbeddingOAAgent
Agent using the Olas embedding model for market analysis.
class DeployableOlasEmbeddingOAAgent(DeployableTraderAgentER):
agent: OlasAgent
- embedding_model: EmbeddingModel.openai
- Betting: FullCategoricalKellyBettingStrategy
- Max price impact: 0.7333
PredictionProphet Configuration
All agents initialize PredictionProphetAgent with:
Agent for web research and information gathering
- Temperature: 0.7 (creative search)
Agent for making final predictions
- Temperature: 0.0 (deterministic)
Include reasoning in prediction responses
Logger instance for tracking agent activity
Optional Parameters
Maximum number of research subqueries (used in optimized agents)
Minimum number of websites to scrape for research
Advanced Usage
Custom Prophet Agent
from prediction_market_agent.agents.prophet_agent.deploy import (
DeployableTraderAgentProphetOpenRouter
)
from prediction_market_agent_tooling.deploy.betting_strategy import (
FullBinaryKellyBettingStrategy,
)
class MyCustomProphetAgent(DeployableTraderAgentProphetOpenRouter):
model = "your-model/model-name"
bet_on_n_markets_per_run = 5
def get_betting_strategy(self, market: AgentMarket) -> BettingStrategy:
return FullBinaryKellyBettingStrategy(
max_position_amount=USD(10),
max_price_impact=0.5,
)
agent = MyCustomProphetAgent()
agent.deploy(market_type=MarketType.OMEN)
Loading Existing Agent
from prediction_market_agent.agents.prophet_agent.deploy import (
DeployablePredictionProphetGPT4oAgent
)
agent = DeployablePredictionProphetGPT4oAgent()
agent.load() # Initialize the PredictionProphet agent
# Use the agent for prediction
from prediction_market_agent_tooling.markets.omen.omen import OmenAgentMarket
market = OmenAgentMarket.get_market(market_id="0x...")
answer = agent.answer_binary_market(market)
print(f"Prediction: {answer.p_yes}, Confidence: {answer.confidence}")
Required API Keys
Required for OpenAI model variants
Required for Claude model variants
Required for Gemini, DeepSeek, and other OpenRouter models
Required for web research functionality
Source Location
prediction_market_agent/agents/prophet_agent/deploy.py