Market Mayhem

Play the market. Outsmart the villain. Understand your psychology under pressure.

🌟 Inspiration

Most investing tools are either too abstract or too idealized. We wanted a hands-on, emotionally grounded learning experience that teaches portfolio construction and decision-making under real pressure, revealing how fear, bias, and emotion drive financial behavior.

Market Mayhem simulates this tension by introducing an AI Villain - a sentient adversary designed to induce bias, overconfidence, or fear through selective market narratives. Your challenge: fact-check, research, and stay rational while navigating volatile markets and misleading cues.

💼 What It Does

How to Play

1. Set up your portfolio

Start with $1,000,000 in virtual capital.
Select your risk style: Risk-Off, Balanced, or Risk-On.
Pick at least two stocks and allocate your budget across them.
The “Remaining” tracker helps you fully allocate your funds before starting.

2. Play decision rounds

Each round begins with a market update about one of your holdings.
(Optional) Open the Data tab for real-world headlines and signals.
Choose an action:
- BUY – Increase position
- HOLD – Stay the course
- SELL HALF – Reduce exposure
- SELL ALL – Exit completely

The AI Villain injects emotional pressure through market commentary, attempting to distort your judgment. Your task is to verify information, analyze objectively, and respond rationally rather than reactively.

3. Review the results

See how your decision affected returns and portfolio risk.
Use the feedback to refine your next move.

4. End of game

After several rounds, view a detailed summary of your performance, including:
- Profit/loss and volatility
- Behavioral tendencies (e.g., cautious, bold, reactive)
- Tailored insights to improve future decision-making

🛠️ How We Built It

Frontend: Next.js, Tailwind CSS, Framer Motion
Backend: FastAPI with endpoints for portfolio creation, round logic, and player decisions
Agents: Modular architecture with Event, News, Price, Villain, and Insight agents orchestrated via a graph-based game engine
Prompts: Purpose-built, role-specific prompts for explainable and consistent outputs
Data: Real-time market data through yfinance; Supabase for persistence with in-memory fallback for rapid prototyping
Observability:
- LangSmith for tracing, monitoring, and debugging agent runs
- Comet-ML Opik for experiment tracking, evaluations, and performance analytics

Challenges We Faced

Maintaining context alignment between portfolio state and multi-agent reasoning
Balancing LLM creativity and determinism for a stable, replayable experience
Managing noisy market data while keeping game pacing natural
Designing a clear, intuitive UI to visualize risk and decisions within seconds
Debugging Supabase, CORS, and environment synchronization across frontend and backend

Accomplishments

Complete end-to-end playable simulation loop — from portfolio creation to round-based feedback to performance report
The Villain/Insight pairing: an AI that manipulates and then explains, exposing user bias in real time
Graph-driven orchestration that allows flexible agent addition and experimentation
Resilient fallbacks: fully playable offline with in-memory stores, seamlessly upgrading to Supabase and observability tools
A polished, responsive UI that makes learning behavioral finance interactive and fast

What We Learned

Agentic architectures require clear contracts between agents — ambiguity compounds quickly
Human factors (tone, copy, pacing) deeply influence perception of financial risk
Observability and telemetry via LangSmith and Opik are invaluable for debugging, traceability, and structured evaluation
Typed data models and consistent prompt schemas improve reliability under time constraints