STOCK ANALYST

Environment Setup

1. Create and activate conda environment:

conda create -n stock-analyst python=3.11 -y
conda activate stock-analyst

2. Install dependencies:

pip install -r requirements.txt

Or install individually:

pip install pyyaml python-slugify newspaper3k google-search-results lxml_html_clean

3. Set API keys:

export SERPAPI_API_KEY='your_api_key_here'
export OPENAI_API_KEY='your_openai_key_here'
export ANTHROPIC_API_KEY='your_anthropic_key_here'  # Optional: for Claude models

On Windows:

set SERPAPI_API_KEY=your_api_key_here
set OPENAI_API_KEY=your_openai_key_here
set ANTHROPIC_API_KEY=your_anthropic_key_here

Update of vynn_core

pip install --upgrade --force-reinstall git+https://github.com/Agentic-Analyst/vynn-core.git

4. Deploy to server:

docker buildx build --platform linux/amd64,linux/arm64 \
-t fuzanwenn/stock-analyst:latest --push .

docker pull fuzanwenn/stock-analyst:latest

Usage

1. Scrape articles:

python src/article_scraper.py --company "NVIDIA" --ticker NVDA --max 15

2. Filter articles:

python src/filter.py --ticker NVDA --min-score 5.0 --max-articles 8 --save-filtered --output-report

3. Screen and analyze:

python src/screener.py --ticker NVDA --min-confidence 0.7 --output-report --detailed-analysis

4. Scrape financial data:

python src/financial_scraper.py --ticker NVDA --statements modeling --save

5. Generate financial models:

python src/financial_model_generator.py --ticker NVDA --model comprehensive --save-excel

6. Complete pipeline with LLM selection:

# Default model
python main.py --ticker NVDA --company "NVIDIA" --email user@example.com --timestamp 20241003_120000

# Use specific model
python main.py --ticker NVDA --company "NVIDIA" --email user@example.com --timestamp 20241003_120000 --llm claude-3.5-sonnet

# List available models
python main.py --list-llms

Workflow Options

1. Supervisor Agent (LLM-Powered Agentic Workflow) 🧠 NEW

The Supervisor Agent provides intelligent, LLM-powered workflow orchestration that dynamically routes between analysis agents based on the current state. This is ideal for exploratory analysis where the system adapts to available data.

Key Features:

• 🧠 LLM-powered dynamic routing (non-sequential, intelligent decisions)
• ✅ Prerequisite validation (prevents invalid routing)
• 🔄 Deterministic fallback (when LLM fails)
• 📊 Complete observability (routing decisions, agent execution, results)

Quick Start:

# Basic supervisor workflow
python supervisor_main.py --ticker NVDA --company "NVIDIA" --email user@example.com

# Use Claude Sonnet for routing
python supervisor_main.py --ticker AAPL --company "Apple" --email user@example.com --llm claude-3.5-sonnet

# Run in Docker
./run_supervisor_docker.sh NVDA "NVIDIA" user@example.com

See README_SUPERVISOR.md for complete documentation.

2. Traditional Pipeline (Sequential Workflow)

The traditional main.py pipeline runs a fixed 7-step workflow in sequence. This is ideal for production scenarios where you want predictable, repeatable analysis.

Quick Start:

# Complete pipeline with LLM selection
python main.py --ticker NVDA --company "NVIDIA" --email user@example.com --timestamp 20241003_120000

# Use specific model
python main.py --ticker NVDA --company "NVIDIA" --email user@example.com --llm claude-3.5-sonnet

# List available models
python main.py --list-llms

Multi-LLM Support

Both workflows support multiple LLM providers:

Available Models:

• gpt-4o-mini (OpenAI) - Fast and cost-effective
• claude-3.5-sonnet (Anthropic) - Balanced quality/speed
• claude-3.5-haiku (Anthropic) - Fastest and cheapest
• claude-3-opus (Anthropic) - Most capable

Filter Logic

Balanced Weighting: Clear distribution of score components:

• Ticker mentions: up to 30%
• Relevance keywords: up to 40%
• Quality indicators: up to 20%
• Penalties: up to 30% deduction

Score Distribution:

• 0-3: Low relevance articles
• 3-5: Medium relevance articles
• 5-7: High relevance articles
• 7-10: Extremely relevant articles

Features

• Multi-LLM Support: Choose between OpenAI GPT-4o-mini and Anthropic Claude models
• Autonomous News Collection: Comprehensive multi-aspect news scraping with AI-powered query generation
• Smart Filtering: Advanced relevance scoring based on content analysis
• Investment Screening: Extracts growth catalysts, risks, and mitigation strategies
• Financial Data Collection: Comprehensive financial statements and market data scraping
• LLM-Powered Financial Modeling: Generates professional DCF models and valuation analysis
• Professional Analyst Reports: AI-generated comprehensive financial reports
• Excel/CSV Export: Professional financial models in Excel format ready for analysis
• Structured Reports: Generates comprehensive markdown reports and JSON data
• Complete Pipeline: End-to-end workflow from news analysis to financial modeling
• Configurable: Adjustable confidence thresholds, LLM selection, and output options