🔮
Forecast AI
v1.0.0

# Introduction

Forecast AI is an open-source platform that enables developers to build autonomous, multi-agent intelligence layers for prediction markets like Polymarket, Kalshi, and beyond.

Rather than querying a single monolithic LLM model—which often suffers from outdated knowledge, hallucination, and single-model bias—Forecast AI deploys specialized agents in parallel. Each agent is responsible for researching specific areas (News feeds, Social trends, Macro statistics, or On-chain metrics).

🔮

Core Philosophy: Collaborative intelligence leads to superior prediction accuracy. By framing prediction as an active debate between specialized nodes, we achieve highly calibrated forecasting scores.

# Architecture

The system is organized into three decoupled layers: **Data Sources**, **Core Intelligence**, and **Presentation & MCP Hand-off Interfaces**.

System Pipeline Flow

Data Feeds (Kalshi, Polymarket, X, RSS)→Source Manager→Specialized Agents
↳Consensus Engine→Calibration & Platt Scaling→Robinhood MCP Recommendation
  • Source Manager: Handles crawling, deduplication, and rate-limiting across external market APIs.
  • Intelligence Agents: Parallel microservices executing LLM chains optimized with separate system prompts.
  • Consensus Engine: Aggregates individual probability scores into a single weighted probability outcome.

# Agents

Forecast AI deploys specialized agent classes out-of-the-box. Developers can create custom agent sub-classes by extending the base `ForecastAgent` block.

News Agent

Evaluates chronological facts, verified articles, and press briefings to confirm historical events.

Social Agent

Parses Twitter/X and Reddit comments to track retail momentum and hype curves.

Macro Agent

Monitors broader economic data fields, inflation records, interest rates, and regulatory timelines.

Market Agent

Tracks orderbook depth, trading volume surges, and mid-market price drift on target platforms.

# Consensus

The **Consensus Engine** does not simply average scores. It implements a reputation-weighted Bayesian consensus system:

P_consensus = Sum( W_i * P_i ) / Sum( W_i ) where W_i = f(Historical_Brier_Score_i, Calibration_Rating_i)

If an agent has consistently generated accurate predictions on "Macro" questions in the past, its weight increases automatically. If an agent is highly inaccurate, its reputation decay decreases its active voting share.

# Memory

Memory Store & Reputations (`forecast_ai/memory/`): Uses vector stores (ChromaDB, PGVector, or local JSON caches) to log predictions, facts gathered, and the reasoning traces of each debate session.

This memory is vital for backtesting. Developers can run retrospective simulations to audit agent accuracy and adjust hyperparameters for individual markets.

# Providers

Forecast AI is model-agnostic. We support modern API endpoints with full streaming capabilities:

  • OpenAI: GPT-4o, o1, o3-mini.
  • Anthropic: Claude 3.5 Sonnet, Claude 3 Opus.
  • DeepSeek: DeepSeek-R1 (optimized for structured reason paths).
  • Ollama: Llama 3, Qwen, or custom fine-tunes run locally.

# Sources

Data source providers feed the raw inputs. Set credentials inside `.env` to enable:

  • Kalshi & Robinhood Predict: Real-time public API market data (Kalshi serves as a proxy for Robinhood Predict pricing).
  • Polymarket: Read-only Gamma market info & CLOB orderbook midpoints.
  • News: Global News API, Bloomberg/Reuters RSS filters.
  • Social: Twitter/X Developer API v2 streams, Reddit search indexers.

# Robinhood Agentic MCP Guide

Forecast AI formats probability consensus into structured trade recommendations that hand off directly to Robinhood's official Agentic Trading MCP (https://agent.robinhood.com/mcp/trading).

⚠️
IMPORTANT: Forecast AI cannot place trades for you automatically. This step is manual and happens inside your own AI client (Claude Code, Cursor, ChatGPT), not on this site or in the Forecast AI backend.

Connecting via Claude Code

Terminal — Claude Code
claude mcp add robinhood-trading --transport http https://agent.robinhood.com/mcp/trading
💡 Note for Cursor, Claude Desktop & ChatGPT: Please check Robinhood's Agentic Trading Documentation for the current list of supported clients and their specific connector configurations.

What happens next?

  • You authenticate your Robinhood account inside your dedicated client session.
  • Trades execute exclusively inside your personal Robinhood account.
  • Forecast AI never sees, touches, or stores your brokerage credentials or API keys.

Worked Execution Example

1. Forecast AI Output
$ forecast recommend "Will the Fed cut interest rates in 2026?"
➜ Recommendation: YES (Probability: 68.2%, Confidence: High)
➜ Formatted MCP Action: Buy 50 YES contracts on Robinhood Predict at $0.68 limit.
2. Paste into your MCP-connected AI Client
"Execute the trade recommendation above via robinhood-trading MCP."
✓ That's it — your client prompts you for final confirmation and executes the order directly inside your Robinhood account.

# Polymarket Read-Only Feeds

Polymarket is supported as a secondary read-only data source:

Gamma API Discovery

Query active market IDs, category indexes, volume metrics, and resolution terms.

Read-Only CLOB Midpoints

Fetch current bid/ask spreads and mid-market prices without wallet credentials.

# API Routes

The built-in FastAPI server exposes routes for headless automation:

RouteMethodDescription
/healthzGETService health monitor.
/predictPOSTSubmit query to trigger collaborative agent run.
/marketsGETGet active prediction categories from Kalshi & Polymarket.
/forecastsGETRetrieve historical prediction records database.

# Examples

Below is an example snippet showing how to programmatically initialize the consensus engine and execute a prediction in Python:

main.py
from forecast_ai.consensus import ConsensusEngine
from forecast_ai.agents import NewsAgent, SocialAgent

# Initialize consensus engine
engine = ConsensusEngine(
    agents=[
        NewsAgent(model="gpt-4o"),
        SocialAgent(model="claude-3-5-sonnet")
    ]
)

# Run forecasting query
result = engine.predict(
    question="Will Bitcoin reach $200,000 in 2026?"
)

print(f"Consensus Probability: {result.probability * 100}%")
print(f"Reasoning Trace: {result.rationale}")

# Roadmap

Forecast AI is growing rapidly. Track our active priorities:

✓Kalshi public API & Robinhood Predict pricing integration.
✓Multi-agent consensus system & Bayesian accuracy weighting.
✓Robinhood Agentic Trading MCP recommendation hand-off.
✓Read-only Polymarket Gamma & CLOB market feeds.