A candy-themed AI agent stock market simulation with autonomous trading intelligence and future market prediction based on events
500 publicly traded companies rendered as candy storefronts in a Three.js 3D city. 10,000+ autonomous AI agents race between stores and compete for the most profitable trades in real time. A full Databricks medallion pipeline processes 1.5 M rows of real finance data into ML-scored trade signals with zero lookahead bias, validated against the 2008 financial crash.
- 3D Candy City — 500 real S&P stocks as candy storefronts, sized by golden score, colored by real brand, placed via Poisson disk sampling
- 10,000+ AI Agents — Autonomous crowd with spatial hash grid, door-fighting physics, and BUY / SHORT / CALL / PUT trade lanes
- 4 Competing Whale Funds — Wonka (Gemini AI), Slugworth (Momentum), Oompa (Value), Gobstopper (Contrarian)
- Golden Ticket System — 5-tier ML signal detection (Dip, Shock, Asymmetry, Dislocation, Convexity) + Platinum (Wonka Bar) rarity tier
- Databricks Medallion Pipeline — Bronze → Silver → Gold processing 602K rows with 50+ features, FinBERT sentiment, and HMM regime detection
- Time Travel — Scrub 5 years of historical data, present state, and Gemini-powered future predictions
| Layer | Technology |
|---|---|
| Frontend | React 19 + TypeScript + Vite 7 |
| 3D Engine | Three.js + @react-three/fiber + @react-three/drei |
| State | Zustand 5 |
| AI | Google Gemini 2.0 Flash (multi-agent hierarchy) + SphinxAI for monitoring |
| Data Pipeline | Databricks (PySpark + Delta Lake) |
| ML Models | FinBERT + Light GBM + BERTopic |
| Backend | FastAPI + WebSockets |
| Dataset | Yahoo Finance, Kaggle and FRED |
| Deployment | Vercel (frontend) |
ANALYZING (drifting, picking next target)
│
▼
RUSHING (pathfinding to store door, speed based on urgency)
│ distance < 2.5 units from door
▼
DOOR_FIGHTING (queued at entrance, 1 admitted per store per frame)
│ admitted
▼
INSIDE (trading at lane quadrant, idle sway animation, tinted to lane color)
│ timer expires
▼
ANALYZING (back to picking next target)
4.0-unit cell size grid for O(n) neighbor queries. Each agent hashed to cell. Collision checks only against agents in adjacent cells. Prevents O(n²) blowup with 10K agents.
Agents rendered as Oompa Loompas using 10 InstancedMesh instances (one per body part: overalls, shirt, head, hair, legs, arms, shoes). 10 draw calls total instead of 100,000+. Per-frame matrix updates via setMatrixAt().
4 competing hedge funds with different strategies. Each controls 25% of the agent population.
| Fund | Color | Strategy | Method |
|---|---|---|---|
| Wonka Fund | Gold #FFD700 | Gemini AI | Hierarchical 11 analysts + PM + risk desk |
| Slugworth Fund | Orange #FF4500 | Momentum | Top 8 stocks with RSI > 55 + MACD > 0 |
| Oompa Fund | Green #00FF7F | Value/Dip | Top 8 deep drawdowns with golden_score >= 1 |
| Gobstopper Fund | Purple #9370DB | Contrarian | Short RSI > 72, long RSI < 32 |
WhaleLeaderboard panel shows P&L, trade count, current allocations, and Gemini reasoning chain for Wonka Fund.
Two validation notebooks prove zero lookahead bias across the entire pipeline:
| Check | Result |
|---|---|
| Row count: Bronze 1,896,880 → Silver 1,896,259 (0.033% drop) | PASS |
| Ticker coverage: 1 missing (BAND) out of 621 | PASS |
| Core OHLCV columns: zero NULLs | PASS |
| GNN metadata: 12 sectors, 0 missing | PASS |
| Gold network_features → Silver referential integrity: 0 orphans | PASS |
| Feature completeness for GNN (daily_return, close): < 0.03% null | PASS |
| Check | Result |
|---|---|
| 2008 data coverage (424 tickers, 505 trading days) | PASS |
fwd_return_5d — 0 mismatched, 0 leaked future prices |
PASS |
fwd_return_20d — 0 mismatched, 0 leaked future prices |
PASS |
| SMA-20 backward-looking only: 0 mismatches in 49,072 crash rows | PASS |
| Drawdown uses only past peaks (deepest: -0.998) | PASS |
| Delta Lake Time Travel: v0 == current, 0 changed crash-era rows | PASS |
| Golden ticket drawdowns match silver layer exactly | PASS |
| Market regime: 69% Bear during Sep 2008 – Mar 2009 (84 Bear days) | PASS |
Conclusion: All signals (moving averages, drawdowns, volatility, golden tickets) are computed from data available at each point in time. Delta Lake Time Travel confirms zero retroactive contamination.
| Method | Endpoint | Description |
|---|---|---|
GET |
/health |
Health check + Databricks status + stocks_available flag |
GET |
/stocks |
Live stock payload from Databricks gold layer (5-min cache), falls back to static JSON |
POST |
/analyze/news |
Gemini + FinBERT sentiment analysis on news text/URL |
GET |
/market/regime |
Current market regime from HMM (Bull/Bear/Neutral) |
GET |
/stocks/correlations |
GNN correlation network edges |
GET |
/stocks/archetypes |
100 agent archetype definitions |
GET |
/scenarios/precomputed |
Pre-built scenario flows |
WebSocket |
/ws/agent-stream |
Bi-directional agent flow + breaking news streaming |
Built at Hacklytics 2026 by:
- Ibe Mohammed Ali — Data/ML Engineer — Georgia State University
- Ryan Varghese — Backend Engineer — Georgia Tech
- Poorav Rawat — UX/UI + Web Dev + Data Pipelines — Georgia State University
- Kashish Rikhi — ML Engineer — Georgia State University
