Prediction Methodology

Ensemble Approach

Our system uses an ensemble of machine learning models instead of a single model. This increases reliability and prediction accuracy.

• LightGBM — fast gradient boosting for tabular data
• CatBoost — specialized in categorical features
• XGBoost — classic gradient boosting with high accuracy

The final prediction is formed by weighted averaging of all model predictions, where weights are determined based on each model's performance on validation data.

Feature Engineering

Prediction quality directly depends on the quality of features. Our system uses several dozen features grouped by category.

Main feature categories:

Model Validation

We use strict validation methodology to avoid overfitting and obtain realistic accuracy estimates.

Temporal Split

Data is split strictly by time, not randomly. This is critically important:

• Game meta changes with every patch — future data cannot be used during training
• Teams change rosters and strategies — past results aren't always representative
• Random splitting creates data leakage and inflates accuracy estimates

5-Fold Temporal Cross-Validation

We use 5-fold temporal cross-validation to evaluate model quality: data is split into 5 consecutive time periods, the model trains on previous folds and is evaluated on the next. The final score is averaged across all folds.

7-Day Buffer

Between training and test sets, a 7-day buffer is left — matches from this period are excluded entirely.

Quality Metrics

• Accuracy — share of correct winner predictions (primary metric)
• ROC AUC — ranking ability of the model (diagnostic, secondary)
• Calibration — alignment of predicted probabilities with actual frequencies
• Brier Score — quality of probabilistic predictions

Calibration and Overfitting Prevention

Even a well-trained model can be "overconfident" — giving probabilities close to 0% or 100% when real uncertainty is higher.

Calibration methods:

• Platt Scaling — logistic regression on top of model predictions
• Isotonic Regression — non-parametric calibration for complex cases
• Temperature Scaling — simple method for neural networks

Overfitting prevention:

• Early Stopping — stop training when validation performance degrades
• Regularization — L1/L2 penalization of complex models
• Feature Selection — removing noisy and correlated features
• Cross-Validation — checking stability across different time periods

Continuous Improvement

Models are retrained as new data accumulates — especially after patch changes, roster swaps, or significant meta shifts.

• Real-time performance monitoring via Accuracy Dashboard
• New features validated on temporal cross-validation before entering production
• Error analysis — which teams and scenarios the model systematically mispredicts
• Data leakage audit for every new feature added