A multi-agent equity research system that produces BUY / NEUTRAL / SELL decisions by fusing three independent analysis agents through a deterministic mathematical Decision Hub. No LLM is involved in the decision itself — only in generating human-readable explanations after the math is done.
- Deterministic decisions — the Decision Hub is a pure mathematical function; same inputs always produce the same output
- Three distinct reasoning patterns — ReAct, FinBERT Funnel, Plan-and-Solve, each matched to its data modality
- Sub-3s median latency — three agents execute in parallel with independent timeouts
- Graceful degradation — produces calibrated decisions even when 1–2 agents fail
- Hybrid RAG — Reciprocal Rank Fusion over SEC 10-K filings, 0.82 recall@5
- Domain-specific ML — FinBERT replaces LLM for sentiment (50x cheaper, deterministic, F1=0.94 on financial text)
- 294 tests across unit, integration, and E2E layers
graph TB
API["POST /api/research/analyze"] --> ORCH["Orchestrator<br/>asyncio.gather"]
ORCH -- "timeout 30s" --> TA["Technical Agent<br/>ReAct Pattern<br/>6 indicator tools"]
ORCH -- "timeout 20s" --> SA["Sentiment Agent<br/>FinBERT Funnel<br/>3-layer pipeline"]
ORCH -- "timeout 60s" --> FA["Fundamental Agent<br/>Plan-and-Solve<br/>4-step + RAG"]
TA -- "AgentSignal" --> HUB{{"Decision Hub"}}
SA -- "AgentSignal" --> HUB
FA -- "AgentSignal" --> HUB
HUB --> S1["1. Consistency Scoring"]
S1 --> S2["2. Dynamic Weighting"]
S2 --> S3["3. Aggregation"]
S3 --> DR["BUY / NEUTRAL / SELL<br/>confidence + risk mode"]
DR --> REASON["LLM Reasoning<br/>explain-only, never modifies"]
REASON --> RESP(["AnalysisResponse"])
style HUB fill:#1a1a2e,stroke:#e94560,stroke-width:2px,color:#fff
style DR fill:#0f3460,stroke:#e94560,stroke-width:2px,color:#fff
style RESP fill:#16213e,stroke:#0ea5e9,stroke-width:1px,color:#fff
The LLM generates reasoning after the decision is finalized and is contractually forbidden from altering the mathematical output. This separation ensures auditability and reproducibility — the decision can be fully explained by inspecting the consistency score, weight allocations, and aggregated score without any LLM interpretation.
| Problem with typical LLM approaches | How this system addresses it |
|---|---|
| Prompt sensitivity flips BUY to SELL | Deterministic math — no prompt involved in the decision |
| Same input, different output across runs | Pure function: identical signals always yield identical decisions |
| Black-box reasoning | Every intermediate value exposed: direction gate, strength spread, regime weights |
| Expensive sentiment inference | FinBERT local inference — 50x cheaper than GPT-5, deterministic |
| Single point of failure | 3 independent agents; 2-agent mode maintains 87% directional agreement |
Implements a Reason + Act loop via OpenAI function calling. The LLM dynamically selects from 6 indicator tools based on intermediate results. The mandatory first step — market_regime — classifies the market as TRENDING, RANGING, or HIGH_VOLATILITY and sets regime-aware thresholds for all subsequent tools.
Typical analysis completes in 2–4 steps, ~1.8s per ticker.
A zero-LLM pipeline that processes financial news through three layers: relevance filtering with deduplication, FinBERT batch classification (~0.3s for 25 articles on CPU), and weighted aggregation with exponential time decay and source credibility scoring. The funnel narrows ~50 raw articles down to a single signal.
When ADANOS_API_KEY is configured, the sentiment agent can also blend in a small public sentiment prior from Reddit, X, and Polymarket. This does not replace the news pipeline — it only adds deterministic context around social attention, bullish breadth, and cross-source alignment.
A 4-step structured pipeline: profile the company, plan analysis tasks, execute via RAG retrieval on SEC 10-K filings + financial data tools, then synthesize with cross-task consistency checking. The RAG pipeline uses hybrid retrieval (dense + BM25) with Reciprocal Rank Fusion, achieving 0.82 recall@5 on structured financial documents.
The core innovation. A three-stage mathematical pipeline that fuses heterogeneous agent signals into a single actionable decision — no LLM, no heuristics, fully deterministic.
Stage 1 — Consistency Scoring: Measures inter-agent agreement across direction, strength, and confidence dimensions. A superlinear penalty (exponent 1.5) on directional disagreement aggressively punishes conflicting signals — this is the primary risk gate.
Stage 2 — Dynamic Weighting: Agent base weights are adjusted by market regime modifiers (technical analysis gets boosted in trending markets; fundamentals anchor in high-volatility regimes) and each agent's own confidence. Weights are always normalized to sum to 1.0.
Stage 3 — Aggregation: A weighted sum of directional signals produces a continuous score, which is classified into BUY / NEUTRAL / SELL via symmetric thresholds. The consistency score determines the risk mode (NORMAL / CAUTIOUS / RISK), which applies a final confidence dampener.
| Active Agents | Confidence Impact | Behavior |
|---|---|---|
| 3 / 3 | Full | Normal analysis |
| 2 / 3 | ×0.8 | Warning issued |
| 1 / 3 | ×0.5 | Strong warning |
| 0 / 3 | — | HTTP 503 |
| Metric | Value |
|---|---|
| Median end-to-end latency | 2.8s |
| P95 latency | 5.2s |
| FinBERT inference | 0.3s / 25 articles (CPU) |
| Decision Hub computation | <5ms |
| Memory footprint | ~1.2 GB (FinBERT loaded) |
| Layer | Technology |
|---|---|
| API | FastAPI, Pydantic v2 |
| Orchestration | asyncio.gather with per-agent timeouts |
| Technical Analysis | OpenAI GPT-5 function calling, ta indicators |
| Sentiment Analysis | ProsusAI/FinBERT (local inference) |
| Fundamental Analysis | OpenAI GPT-5 + hybrid RAG |
| RAG | Qdrant, LlamaIndex, sentence-transformers, RRF fusion |
| Data Sources | yfinance, Finnhub API, SEC EDGAR (edgartools) |
| Frontend | React 18, TypeScript, Tailwind CSS |
| Domain Models | Frozen Pydantic v2 (immutable value objects) |
| Testing | pytest — 294 tests across unit / integration / E2E |
git clone https://github.com/flash131307/multi-agent-investment.git
cd multi-agent-investment
# Backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Frontend
cd frontend && npm install && cd ..
# Configure
cp .env.template .env # Fill in OPENAI_API_KEY, FINNHUB_API_KEY (ADANOS_API_KEY optional)
# Run
uvicorn backend.main:app --reload --port 8000 # Terminal 1
cd frontend && npm run dev # Terminal 2Frontend: http://localhost:5173 · API docs: http://localhost:8000/docs
curl -X POST http://localhost:8000/api/research/analyze \
-H "Content-Type: application/json" \
-d '{"ticker": "AAPL"}'Example response
{
"ticker": "AAPL",
"decision": {
"direction": "BUY",
"confidence": 0.72,
"risk_mode": "NORMAL",
"consistency_score": 0.85,
"aggregated_score": 0.416,
"reasoning": "All three agents signal positive outlook..."
},
"agents": [
{"agent_name": "technical", "direction": "BUY", "strength": "STRONG", "confidence": 0.85},
{"agent_name": "sentiment", "direction": "BUY", "strength": "MODERATE", "confidence": 0.68},
{"agent_name": "fundamental", "direction": "BUY", "strength": "MODERATE", "confidence": 0.71}
],
"warnings": []
}backend/
├── agents/
│ ├── technical/ # ReAct agent + 6 indicator tools
│ ├── sentiment/ # FinBERT 3-layer funnel
│ └── fundamental/ # Plan-and-Solve + RAG tools
├── decision_hub/ # Consistency → Weighting → Aggregation
├── models/ # Frozen Pydantic v2 domain models
├── rag/ # EDGAR loader, chunker, embedder, hybrid retriever
├── orchestrator/ # Parallel execution + graceful degradation
├── services/ # Yahoo Finance, Finnhub, FinBERT, ticker resolver
└── config/ # Environment-based settings
frontend/ # React + TypeScript + Tailwind
tests/
├── unit/ # 12 test modules
├── integration/ # 5 test modules
└── e2e/ # API endpoint tests
The original system used a chat-style research query interface — users typed free-form questions, a Router Agent extracted tickers, and the system returned a long-form markdown report with embedded charts. This version was refactored into the current 3-agent + Decision Hub architecture to achieve deterministic, auditable decisions.
Click to view pre-refactor screenshots (6 images)
Home — Research Query Interface
Analysis in Progress — Parallel Agent Execution
Results — Simple View (Investor Snapshot)
Results — Detailed View (Full Research Report)
Interactive Charts — Price Trend
Interactive Charts — Peer Valuation Comparison
MIT — see LICENSE for details.