AlphaAgents — Multi-Agent Equity Research

Problem

Equity research fuses heterogeneous, unstructured evidence — 100+ page SEC filings, noisy news flow, and price action. A single model struggles to weigh these fairly. This system decomposes the problem into specialist agents and reconciles their disagreements through a structured debate, mirroring how an investment team actually works.

Methods

• Fundamental agent: retrieval-augmented generation (RAG) over 10-K/10-Q filings with a deterministic keyword fallback
• Sentiment agent: news aggregation with LLM synthesis and a lexicon-based fast mode
• Valuation agent: technical indicators (moving averages, RSI, momentum)
• Collaborator: signal normalization, confidence weighting, and a debate triggered on signal spread
• Backtesting: state-machine positions with holding periods, stop-loss, transaction costs, and explicit leakage checks

Outputs

• Per-stock BUY/HOLD/SELL with confidence and rationale
• Portfolio equity curve, Sharpe, and max drawdown
• Benchmark comparison vs. buy-and-hold
• Debate transcripts for human review