Building Prediction Models: Our Approach
After years of iteration, I want to share how we actually approach football prediction at OddsFlow. No magic—just careful data work and honest evaluation.
The Data Foundation
Everything starts with data quality. We aggregate from multiple sources:
Match-level data:
- Historical results (5+ years)
- xG and advanced metrics
- Lineup information
- In-match events
Market data:
- Multi-source odds snapshots
- Price movement history
- Market timing information
Contextual data:
- League standings and context
- Rest days and travel
- Competition phase importance
Feature Engineering: Where the Work Is
Raw data isn't useful. The real work is transforming it into predictive features.
Team strength features:
- Rolling xG averages (home/away specific)
- Elo-style power ratings
- Recent form indicators
Market-derived features:
- Implied probabilities from opening odds
- Opening-to-close movement
- Cross-market discrepancies
Contextual features:
- Match importance index
- Fatigue indicators
- Head-to-head adjustments
We've tested hundreds of features. Most don't add value. The discipline is in what you *don't* include.
Model Architecture
We use an ensemble approach—multiple models combined:
Base models:
- Gradient boosted trees (XGBoost) for tabular features
- Poisson models for goal expectations
- Market consensus baselines
Combination:
Weighted averaging based on out-of-sample performance. Weights adjust by league and market type.
We deliberately avoid overly complex architectures. Football is noisy. Simple, well-calibrated models often outperform complex ones.
What Actually Matters
After years of experimentation, here's what moves the needle:
- 1Data quality over quantity: Clean, consistent data beats more features
- 2Calibration over accuracy: Well-calibrated probabilities matter more than win rate
- 3Market awareness: Using odds as features is powerful but requires care
- 4Honest evaluation: Out-of-sample testing on recent data, not historical curves
Our Limitations
No model is perfect. Ours struggles with:
- Early season (small recent sample)
- Manager changes and squad upheaval
- Highly unusual match contexts
- Goalkeeper-dominated matches
We're transparent about uncertainty. When confidence is low, we say so.
📖 Related reading: Evaluating Prediction Models • Feature Engineering Deep Dive
*OddsFlow provides AI-powered sports analysis for educational and informational purposes.*