🚀 CS224G-MACI-for-Stock-Prediction

[Sprint 4 - NEW UPDATE]

- Please review section How To Run The Code Part D for the app

- Code: cd MACI_stock_prediction

[Sprint 3 - UPDATE]

- Paper: MACI Framework

- Code: cd MACI_framework

🏗️ 7 Multi-Agent tools

• Our newest version MACI support Magentic multi-agent tool.
• Our previous MACI-framework could support multi-agent tools: LangGraph, AutoGen, Crewai, LangChain, and to be extended on LlamaIndex, Haystack.

🤝 20+ LLM Agents

• OpenAI LLMs: including gpt-4, gpt-4o, gpt-4o-mini, etc.
• Ollama: including Llama 3.3, DeepSeek-R1, Phi-4, Mistral, Gemma 2, etc.
• Anthropic: including Claud 3.7.
• Mistral:
• LiteLLM:
• or any other OpenAI schema-compatible model
• FinAgent(ours)
• Your Self-developed LLMs(ours)
• Access to financial database such as Alphavantage, and yahoo api. adapts to unexpected changes in real-time financial information.

🤖 Interactive Back-End App

⚠️ 1. LLM Limitations in Complex Planning

Large Language Models (LLMs) excel at pattern recognition but struggle with complex planning tasks that require:

• 🧠 Deliberate reasoning
• ⏳ Temporal awareness
• 📏 Constraint management

🔍 1.1 Key Limitations of Current LLM Models:

1. ❌ Lack of Self-Verification

   • LLMs cannot validate their own outputs, leading to errors.

2. 🎯 Attention Bias & Constraint Drift

   • Contextual focus shifts, ignoring earlier constraints.

3. 🏗️ Lack of Common Sense Integration

   • Omits real-world constraints (e.g., logistics delays).

---

🤖 2. MACI: Multi-Agent Collaborative Intelligence

MACI is designed to overcome these LLM limitations using a three-layer approach:

1. 🏗️ Meta-Planner (MP)

   • Constructs task-specific workflows, identifying roles and constraints.

2. 🤝 Common & Task-Specific Agents

   • Common Agents: Validate constraints & reasoning quality.
   • Task-Specific Agents: Optimize domain-specific tasks.

3. 📡 Run-Time Monitor

   • Adapts to unexpected changes in real-time.

---

📅 3. Project Plan

🔄 3.1 LLM Model Improvement on:

1. ✅ Lack of Self-Verification

   • Independent validation agents ensure correctness.

2. 🔍 Attention Bias

   • Task-specific agents with constrained context windows prevent bias.

3. 🌍 Lack of Common Sense

   • Integration agents enhance real-world feasibility.

🧪 3.2 LLM Research and Experiments:

• 📝 Spec completed, paper submitted to ICML.
• 🏆 Tested on Traveling Salesperson & Thanksgiving Dinner Planning, outperforming all LLMs, including DeepSeek.
• 📊 Stock Prediction application designed.

📆 3.3 General Timeline:

• 🏗 (Sprint 1) Implementation, Experiments, First Results →
• 🚀 (Sprint 2) First App Demo (Frond-End & Back-End) →
• 🔧 (Sprint 3) MACI Enhancement, Second App Demo →
• 🎤 (Demo Day) App Integration, Paper Submission for NeurIPS, Final App Demo →

🤖 3.4 Multi-Agent App Development:

• 🔍 (1) LLM Specializatio✅: Different LLMs tailored for Stock Prediction, Company Analysis, Personal Insights, and Job Market Trends.
• 🏗 (2) Overall Architecture✅: Using MACI Framework with website, Multi-Agents selection and collaboration, to generate holistic financial and career insights.
• 📈 (3) Agent 1 & 2✅: Stock Prediction & Company Analysis: Real-time data integration, LLMs for market trends and company analysis.
• 👥 (4) Agent 3 & 4✅: Investment Expert Analysis & Job Analysis: AI-driven investment, career suggestion, and matching based on market shifts.
• 🚀 (5) Milestones: Iterative development, testing, and user feedback loops.

---

⚙️ 4. Experiment Set-up

We plan to evaluate our multi-agent temporal planning framework on S&P 500 stocks (2018-2024), incorporating historical trading data, reports data, and other relevant financial data. Our focus will be on three major market sectors.

All experiments will utilize publicly available data from sources like:

• 📈 Yahoo Finance (Stock Prices)
• 📜 SEC EDGAR (Financial Reports)

⚖️ 4.1 Baselines & Comparisons:

1. 📊 Baseline Models:

   • Traditional machine learning methods (Logistic Regression, SVM, LSTM(best), LLM, MACI(best) networks).

2. 🆚 Comparison with MACI:

   • Evaluating prediction accuracy (MAE, MSE) and directional accuracy for stock movement prediction.

3. 🧩 Ablation Studies:

   • Testing different combinations of agents and their impact on performance.

4. 📊 Robustness & Scalability:

   • Assessing performance across different market conditions and unseen stocks (out-of-sample validation).

---

🎓 5. Contribution

1. 📄 Paper: Multi-Agent Collaborative Intelligence for Robust Temporal Planning – Edward Y. Chang
2. 📄 Paper: REALM-Bench: A Real-World Planning Benchmark for LLMs and Multi-Agent Systems – Longling Gloria Geng, Edward Y. Chang
3. 💻 GitHub Setup, App Development, and Experiments – Longling Gloria Geng

---

🚀 How to Run the Code

A. Base Experiments

1) (Optional) Create and Activate a Virtual Environment

It is recommended to use a virtual environment to manage dependencies:

python3 -m venv env
source env/bin/activate  # On macOS/Linux
env\Scripts\activate     # On Windows

2) Install Dependencies

Ensure you have all necessary dependencies installed:

pip install -r requirements.txt

Or install manually:

pip install pandas numpy matplotlib prophet

3) Download & Place the S&P 500 Stocks Data

The dataset is available on Kaggle:
🔗 S&P 500 Stocks Dataset

Extract and place the CSV file inside the sp500_stocks/ directory:

mkdir -p sp500_stocks
mv path/to/sp500_stocks.csv sp500_stocks/

4) Execute the Python Script

Run the stock prediction script:

python3 main.py

5) Wait for the Script to Finish

The script will:
✅ Predict stock prices for 2 years into the future
✅ Load and preprocess the stock data
✅ Train a Prophet forecasting model
✅ Generate & save plots showing historical vs. forecasted values

6) View Generated Plots

Once the script completes, you’ll find the forecasted plots in the project folder:

• AAPL_forecast.png → Forecast for Apple
• TSLA_forecast.png → Forecast for Tesla
• META_forecast.png → Forecast for Meta
• other plots etc.

B. Front-End Demo

cd checkco

then open index.html

C. Back-End Demo

cd checkco
python3 server.py

then open front.html

D. MACI for Stock Prediction Demo (final)

cd MACI_stock_prediction

then follow README file to set up your own multi-agent framework and pipeline.

✅ Step 1: Run FastAPI with Uvicorn Make sure you're in the same directory as main.py and then run:

export OPENAI_API_KEY=
export...
uvicorn main:app --reload

✅ Step 2: go to 127.0.0.1.8000/static/front.html

Check File Structure Your project should be organized like this:

MACI-Stock-Prediction/
│── main.py  # ✅ FastAPI app entry point
│── static/  # ✅ HTML, CSS, and JavaScript for UI
│   ├── index.html
│   ├── style.css
│   ├── script.js
│── templates/  # (Optional) Jinja2 templates
│── utils/  # ✅ Helper functions (e.g., API calls)
│   ├── indicators.py
│   ├── charts.py
│── .env  # ✅ API Keys
│── requirements.txt  # ✅ Python dependencies
│── README.md  # ✅ Project documentation

This README provides an overview of the CS224G MACI for Stock Prediction project, highlighting its motivations, project plan, methodologies, demo, and future directions. 🚀