• 

Inspiration

Global sports markets are often treated as “efficient,” but in reality they are heavily influenced by human bias, incomplete data, and emotional narratives.

We noticed patterns:

• Popular teams are often overvalued
• Underdogs (especially from less-covered regions) are systematically underestimated
• Early World Cup odds are set with limited information and slow adjustment

This raised a key question:

What if we could combine historical data, machine learning, and live market odds to systematically detect where the market is wrong?

WorldCup-Oracle was built to answer that.

What it does

WorldCup-Oracle is an AI-powered prediction engine that identifies mispriced teams in World Cup betting markets.

It:

• Predicts team performance probabilities using machine learning
• Ingests real-time odds from prediction markets
• Compares model predictions with market-implied probabilities
• Highlights undervalued and overvalued teams

Users can:

• Explore predictions through an interactive dashboard
• Compare teams in-depth
• Identify value opportunities using a ranked “edge” system

How we built it

We built a full-stack system with four main layers:

1. Data Layer

• Historical World Cup data (2002–2022) from FBref
• Team metrics: xG, defense, possession, squad strength
• Live odds from Polymarket (GraphQL) and Kalshi (REST)

2. Modeling Layer

• Gradient Boosting ensemble (XGBoost/LightGBM)
• Leave-one-tournament-out cross-validation

• Outputs probabilities for:

  • Winning
  • Reaching semifinals
  • Group stage exit

3. Market Analysis Layer

• Converts odds into implied probabilities
• Removes market margin (vig normalization)
• Computes edge = model – market

4. Interface & API

• Streamlit dashboard (interactive charts, comparisons)

• FastAPI backend with endpoints:

  • /predictions
  • /mispricing
  • /backtest

Challenges we ran into

• Data inconsistency: Historical World Cup data across years required normalization and cleaning to maintain feature consistency.

• Market normalization: Removing bookmaker margins and aligning different market formats (Polymarket vs Kalshi) was non-trivial.

• Overfitting risk: With limited World Cup samples, ensuring the model generalized required careful validation (leave-one-out tournaments).

• Real-time integration: Handling live odds reliably while maintaining performance required caching and fallback strategies.

🏆 Accomplishments that we're proud of

• Achieved 83% performance improvement over baseline market predictions in backtesting
• Built a complete pipeline from data → model → market → insights → UI
• Identified real, quantifiable mispricings (e.g., Argentina +18% edge)

• Delivered both:

  • A user-friendly dashboard
  • A developer-ready API

Most importantly:

We proved that combining AI with market data creates measurable advantage.

🧠 What we learned

• Markets are strong—but predictably biased in certain scenarios
• Blending model predictions with market data is more powerful than either alone
• Validation strategy matters more than model complexity
• Good visualization is critical for turning data into decisions

🚀 What's next for WorldCup-Oracle

• Expand beyond the World Cup into:

  • Club football (Champions League, leagues)
  • Other sports (NBA, NFL, etc.)

• Improve the model with:

  • Player-level data
  • Injury tracking
  • Real-time form updates

• Add:

  • Automated alerts for new mispricings
  • Mobile-friendly interface
  • User personalization

• Explore integration with:

  • Live betting tools
  • Portfolio tracking systems

WorldCup-Oracle is more than a prediction tool it’s a system designed to find truth where markets are imperfect.

Built With

• apis
• python
• streamlit
• zerve