Multi-Agent Investment Research System

A research-oriented decision-support system that coordinates specialized AI agents to analyze U.S. equities, construct a portfolio, evaluate risk, and explain the resulting recommendation.

This repository contains my bachelor's thesis project. It combines LLM-based reasoning with machine learning, quantitative portfolio optimization, retrieval-augmented memory, and explicit policy constraints. It is an engineering and research prototype—not a proven investment strategy or financial advice.

Architecture

flowchart TD
    A[Investor request] --> B[Profile Agent]
    B --> C[Data Agent]
    C --> D[Scoring and Regime Agents]
    D --> E[Portfolio Agent]
    E --> F[Risk and Backtest]
    F --> G[Critic Agent]
    G -->|Revise, up to 3 cycles| E
    G -->|Approve| H[Explainability Agent]

Loading

The system uses two LangGraph workflows: one for portfolio construction and another for ongoing portfolio monitoring. Shared graph state carries profiles, market data, model outputs, risk reports, provenance, and execution traces between agents.

Specialized agents

• Profile Agent — converts a natural-language request into a structured investor profile.
• Data Agent — collects market, fundamental, macroeconomic, and news data.
• Scoring Agent — combines ML predictions with technical and fundamental signals.
• Regime Agent — classifies the current market environment.
• Portfolio Agent — selects assets and calculates portfolio weights.
• Risk Agent — evaluates volatility, drawdown, VaR, concentration, and policy constraints.
• Critic Agent — reviews the proposal and can trigger another construction cycle.
• Explainability Agent — produces a structured, human-readable rationale.
• Monitoring Agent — evaluates an existing portfolio and recommends whether action is required.

Engineering highlights

• LangGraph orchestration with conditional routing and up to three critic-driven revision cycles.
• A 100-asset universe with asset-class, sector, geography, and style metadata.
• XGBoost scoring based on technical and macroeconomic features.
• Portfolio optimization with PyPortfolioOpt and Ledoit-Wolf covariance estimation.
• Three retrieval mechanisms for investment knowledge, cached news, and previous decisions.
• FastAPI endpoints and a Streamlit interface for portfolio construction and monitoring.
• Request IDs, correlation IDs, provenance records, decision logs, and persistent execution traces.
• 58 unit and integration tests covering agents, graphs, API handlers, RAG, policies, risk, and UI helpers.

Technology stack

Python, LangGraph, LangChain, OpenAI, Pydantic, XGBoost, scikit-learn, PyPortfolioOpt, ChromaDB, Hugging Face embeddings, yfinance, FastAPI, Streamlit, Pandas, and NumPy.

Running locally

git clone https://github.com/Gregory-mcbit/diploma.git
cd diploma

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Create a .env file:

OPENAI_API_KEY=your_openai_api_key

Initialize the default knowledge base:

python -m app.rag.init_knowledge

Start the API:

uvicorn app.api.app:app --reload

Or launch the Streamlit interface:

streamlit run app/ui/streamlit_app.py

Run the test suite:

pytest

Disclaimer

This project is for research and educational purposes only. It does not provide financial advice or a recommendation to buy or sell securities.