🤖 Priced In [ARCHIVED]

Note: This repository has been archived. It served as an experimental AI-powered autonomous stock trading agent that ran from June to October 2025.

📊 Final Results

Initial Investment: $1,000.00
Final Portfolio Value: $1,389.80
Total Return: +38.98% ($389.80 gain)
Annualized Return (CAGR): 136.67%
Trading Period: June 24, 2025 - October 15, 2025 (113 days)
Total Trades Executed: 60+ trades

Final Holdings

Asset	Shares	Value
Cash	-	$0.03
NVDA (Nvidia)	4.82	$959.42
GOOGL (Google)	1	$290.10
XLI (Industrial Sector ETF)	0.106	$16.35
XLP (Consumer Staples ETF)	0.729	$55.72
XLU (Utilities Sector ETF)	0.380	$34.07
XLV (Health Care Sector ETF)	0.23	$33.87
XLY (Consumer Discretionary ETF)	0.001	$0.24

🎯 The Experiment

This project explored the capabilities and limitations of autonomous AI agents in financial decision-making by creating a fully automated trading bot that:

Operated independently - Made all trading decisions without human intervention
Ran on GitHub Actions - Executed trades every 6 hours via scheduled workflows
Used OpenAI’s Agents framework - Leveraged GPT-4 with function calling for decision-making
Had full market access - Could research stocks, check prices, and execute trades
Self-documented - Updated its own README with portfolio performance after each session

How It Worked

The agent followed a systematic approach each trading session:

Think - Used structured reasoning to plan its approach
Portfolio Review - Checked current holdings and performance
Market Research - Searched the web for market trends, news, and opportunities
Analysis - Evaluated potential trades based on momentum, news, and technical factors
Execution - Made buy/sell decisions within risk management constraints
Logging - Recorded all decisions and reasoning in agent.log

The system prompt instructed the agent to:

Focus on growth while managing risk
Never put all capital in a single position
Maintain some cash reserve for flexibility
Track performance against the initial $1,000 investment
Always think before acting using a dedicated “think” tool

🔧 Technical Implementation

Built with:

Node.js + TypeScript for the core agent logic
OpenAI Agents Framework (@openai/agents) for AI orchestration
GPT-4.1-mini for decision-making and web search
GitHub Actions for scheduled execution (every 6 hours)
Git automation for portfolio persistence and README updates

Key Features

Web Search Integration: Agent could research stocks and market conditions in real-time
Fractional Shares: Supported buying partial shares for precise capital allocation
Persistent Memory: Maintained conversation history across sessions via thread.json
Automated Logging: Every decision and trade was logged with timestamps
Self-Updating Dashboard: Automatically updated README with current portfolio state

Architecture

// Core tools available to the agent
- think: Structured reasoning before actions
- get_portfolio: Check current holdings
- get_net_worth: Calculate total value
- get_stock_price: Fetch current prices
- buy: Execute purchase orders
- sell: Execute sell orders
- web_search: Research markets and news

The agent operated with full autonomy within these constraints:

Started with $1,000 cash
Could only trade during market hours (via real-time price APIs)
Had to manage its own cash balance
Self-imposed risk management rules (no single-stock concentration)

📈 What We Learned

Successes

Technical Feasibility: Autonomous AI trading agents can be built and operate reliably
Positive Returns: The agent achieved ~39% returns over ~4 months
Consistent Execution: Successfully ran 100+ automated trading sessions without failures
Risk Management: Generally followed diversification principles
Momentum Capture: Effectively identified and traded on momentum in tech stocks (especially NVDA)

Limitations & Observations

Market Timing Dependency: Results heavily influenced by being active during a tech bull market
Heavy Tech Concentration: Final portfolio was ~69% NVDA, showing concentration risk
Short-Term Focus: Primarily made short-term momentum trades rather than long-term investments
Cost Considerations: Real-world trading costs (fees, spreads, taxes) not simulated
Limited Market Context: No access to fundamental analysis, financial statements, or complex technical indicators
Reasoning Quality: While “think” tool provided transparency, reasoning was sometimes superficial
No Backtesting: Performance not validated across different market conditions

Trading Patterns

Analyzing the 60+ trades:

Initial Phase (June): Aggressive buying of tech stocks (TSLA, NVDA, PLTR, GOOGL)
Mid Phase (July-Aug): Active rotation between sector ETFs (XLI, XLP, XLU, XLV, XLY)
Late Phase (Sep-Oct): Consolidation into core positions, mostly holding NVDA
Average Hold Time: Short-term (often 6-48 hours), typical of momentum trading
Win Rate: Not explicitly tracked, but frequent small trades suggest active management style

🎓 Key Takeaways

AI can trade, but context matters: The agent made reasonable decisions within its programmed constraints, but lacked deep market understanding
Transparency is crucial: The “think” tool requirement provided valuable insight into the agent’s reasoning
Bull markets help: Results likely wouldn’t replicate in bearish or volatile conditions
Automation works: GitHub Actions proved to be a reliable platform for scheduled agent execution
AI is a tool, not magic: The agent followed patterns similar to novice momentum traders

⚠️ Important Disclaimers

Educational Purpose Only: This was an experiment in AI capabilities, not investment advice
Past Performance ≠ Future Results: The positive returns occurred during favorable market conditions
Not Financial Advice: This project does not constitute financial, investment, or trading advice
Significant Risks: Real trading involves substantial risk of loss. Never invest money you can’t afford to lose
Simulated Environment: While using real price data, this didn’t execute real trades or account for real-world costs
No Endorsement: Results should not be interpreted as endorsement of autonomous trading systems

📚 Resources

Agent Source Code - The complete trading agent implementation
System Prompt - Instructions given to the AI agent
Trading Log - Complete log of all decisions and trades
Portfolio History - Final state of holdings and trade history
GitHub Actions Workflow - Automation configuration

🔍 For Researchers & Developers

If you’re interested in AI agents or algorithmic trading:

Study the logs: agent.log contains the complete decision-making history
Review the prompt: system-prompt.md shows how the agent was instructed
Examine the code: agent.ts demonstrates OpenAI Agents framework usage
Analyze trades: portfolio.json has the complete trade history with timestamps and prices

Running Locally (Historical)

While the live trading is archived, you can still run the code locally to study it:

git clone https://github.com/AnandChowdhary/priced-in.git
cd priced-in
npm install
export OPENAI_API_KEY="your-api-key-here"
npm start

Note: This will make real API calls and modify portfolio.json. Consider working on a fork or branch.

🙏 Acknowledgments

This experiment was made possible by:

OpenAI for the Agents framework and GPT-4 capabilities
GitHub Actions for reliable automation infrastructure
The open-source community for tools and inspiration

💭 Final Thoughts

This project demonstrated that autonomous AI trading agents are technically feasible and can operate successfully in favorable market conditions. However, it also highlighted important limitations:

AI lacks the nuanced judgment of experienced traders
Success was heavily dependent on market timing and conditions
Concentration risk emerged despite risk management instructions
Real-world complexity (costs, regulations, market microstructure) wasn’t fully captured

The future of AI in trading likely involves human-AI collaboration rather than full autonomy, with AI handling data analysis and pattern recognition while humans provide strategic oversight and risk management.

Thank you for following this experiment! 🚀

📄 License

Repository Status: Archived as of December 2025