How We Achieved 74% Accuracy in Sports ML Predictions

The Challenge

Build ML models that predict football match outcomes better than random chance. The baseline for predicting draws is ~26%. Can we beat it significantly?

Feature Engineering > Model Complexity

We tried deep learning first. Neural networks with match embeddings, attention layers, the works. Accuracy: 38% on draws.

Then we switched to feature engineering. We created 62-65 features per match:

Team form over last 5/10/20 matches

Head-to-head statistics

Home/away performance deltas

Goal-scoring patterns

League-specific adjustments

With Scikit-learn’s ensemble methods on these engineered features: 44.7% accuracy on draws (+19% edge over baseline).

Three Models, Three Markets

**ODYSSEUS** — Draw predictions: 44.7% accuracy (vs 26% baseline)

**LEONIDAS** — BTTS (Both Teams To Score): 74.1% accuracy

**ACHILLES** — Over 2.5 Goals: 72.3% accuracy

Why Scikit-learn Won

Tabular data with engineered features favors gradient boosting

Training time: minutes vs hours

Interpretability: feature importance is transparent

Deployment: pickle files, no GPU needed

Key Takeaway

In ML, the quality of your features matters more than the complexity of your model. Spend 80% of your time on data and features, 20% on model selection.