Under the hood — algorithms

The model layer.

The algorithms that turn fixtures, form and market data into probabilities — from gradient-boosted trees to Poisson regressions and calibrated ensembles.

XGBoost

XGBoost

Gradient boosting

Fast gradient boosting with strong tabular performance, missing-value handling, and SHAP-based feature explanations.

LightGBM

LightGBM

Gradient boosting

High-speed, memory-efficient boosting with native categorical support, leaf-wise growth, and fast hyperparameter sweeps.

Random Forest

Random Forest

Bagging

Stable ensemble baseline with built-in out-of-bag scoring, permutation importance, and low variance on noisy data.

CatBoost

CatBoost

Gradient boosting

Handles categorical features natively with ordered boosting for overfitting resistance and built-in permutation importance.

Neural networks

Neural Networks / LSTM

Deep learning

Captures complex non-linear patterns in time-series data with embeddings, attention activations, and hybrid tree stacking.

Ensembles

Ensembles & Calibration

Stacking

Stacks multiple models and calibrates probabilities (Platt scaling, isotonic) for more reliable edge detection.

Ridge regression

Ridge Regression

Linear

Regularized linear model (L2) for spread and totals regression — stable under multicollinearity with interpretable coefficients.

Repo-specific models

Repo-specific Models

  • Logistic Regression — value probability modeling, manager/referee effect analysis (premier-league).
  • ELO + custom ratings — team/player form and matchup strength (all sports models).
  • Poisson/Score-based regressions — player prop and goal prediction pipelines (hockey, soccer).
  • Gradient Boosting variants — domain-specific feature-discrete handling (gridiron, tennis, golf).
  • Time-series models — sequential injury trend and live tracking models (nba-predictions).