Deploy specialized agents to crawl news, extract sentiment, and analyze prediction markets (Kalshi & Polymarket). Aggregate output using our Bayesian Consensus Engine and hand off trades via Robinhood's Agentic Trading MCP.
Native design for Robinhood's prediction market product.
Collaborative consensus over single model bias.
Fully audit-ready, secure, and self-hosted.
Real-time market feeds from Kalshi & Polymarket.
Execution hand-off via Robinhood Agentic Trading MCP.
Forecast AI orchestrates isolated agent pipelines to extract evidence and continuously calibrate confidence scoring.
Deploys specialized agents (News, Social, Macro, Market) in parallel to evaluate forecasting questions.
Aggregates independent agent probabilities based on their dynamic, historical accuracy scores.
Evaluates information source credibility, relevance, and recency to filter noise.
Utilizes vector stores to maintain contextual persistence across long-term prediction markets.
Formats consensus forecasts into structured trade recommendations for Robinhood's official MCP endpoint.
Generates clear, structured rationales detailing each specialized agent's confidence and findings.
Flexible backend compatible with OpenAI, Anthropic, DeepSeek, or local Ollama instances.
Configure models, fetch midpoints, and execute live forecasts directly from the terminal.
Exposes lightweight FastAPI routes for microservice orchestration and status metrics.
Free, extensible, and fully self-hostable codebase under the Apache 2.0 License.
Watch how raw signals flow from public feeds (Kalshi, Polymarket), compile in our Multi-Agent mesh, and output structured trade recommendations via Robinhood's Agentic Trading MCP.
Pipeline Phase
Click other components in the visualizer to explore how Forecast AI builds consensus dynamically across diverse agents.
Single-prompt LLMs suffer from recency bias, lack local context, and fail to weigh contradictory evidence properly.
Forecast AI separates concern: dedicated agents collect, debate, and consensus-pool, achieving +24% Brier Score accuracy gains over vanilla GPT-4.
Rather than asking a model to recall facts, sources stream real-time feeds (News, Social sentiment, Kalshi & Polymarket orderbooks). Each stream is indexed by the Source Manager to guarantee zero-shot ground truths.
Isolated agents receive customized system prompts representing different personas (Macro Economist, Market Analyst, Social Sentiment Expert). They generate localized predictions and present evidence.
The Consensus Engine weighs agent inputs using historical Brier score accuracy ratings. The engine filters anomalies, scales results via Platt calibration, and yields a single probability matrix.
The consensus probability outputs to structured trade recommendations handed off to your own Robinhood Agentic Trading MCP session for user-controlled execution.
Forecast AI continuously monitors public market feeds from Kalshi (serving as a proxy for Robinhood Predict pricing) and Polymarket as read-only market data feeds.
Kalshi public API provides real-time market data matching Robinhood Predict probabilities.
Polymarket Gamma & read-only CLOB orderbooks supply secondary liquidity and probability signals.
Forecast AI formats probability forecasts into structured trade recommendations and hands them off for you to execute via Robinhood's official Agentic Trading MCP endpoint (https://agent.robinhood.com/mcp/trading) in your own AI client (Claude Code, ChatGPT, Cursor). Forecast AI never holds credentials or executes trades directly.
Get started today by cloning our repository, installing our CLI, and deploying prediction agents. Forecast AI is fully open-source.