Tim’s Picks

NHL player prediction dashboards for showcasing applied AI.

Active in-season. Off-season shows last finalized results.

Models

Heuristic Model

Rule-based picker using recent and season-to-date stats to estimate a player’s probability of scoring.

HeuristicBox score
Open results

Advanced Heuristic Model

Adds opponent strength, goalie quality, and contextual features to estimate the % chance of a player scoring.

Context featuresOpponent strength
Open results

Polynomial Regression

Polynomial regression mapping engineered features to calibrated scoring probabilities.

RegressionCalibrated
Open results

Neural Network

Multi-layer perceptron modeling nonlinear interactions to predict a player’s scoring probability. Regularized with dropout and early stopping.

MLPNonlinear
Open results

LightGBM

Gradient-boosted trees optimized for tabular data, outputting well-calibrated scoring probabilities.

LightGBMFeature importance
Open results

XGBoost

Robust boosted trees with strong regularization, trained to predict and calibrate the % chance of a player scoring.

XGBoostBoosted trees
Open results

Methodology

Building reliable scoring predictions

Models are trained on thousands of NHL games, incorporating both current and historical player data. Features include player form, usage patterns, opponent strength, recent performance trends, and team standings. This multi-layered approach enables accurate, context-aware scoring probabilities, achieving an overall winrate of about 70%. The bot on

  • Inputs: rolling player stats, season-to-date metrics, opponent performance, team standings.
  • Features: recent momentum, situational usage, opponent quality, goalie matchups, home/away splits.
  • Training: thousands of past games used for fitting and calibration.
  • Performance: ~70% winrate with proper validation and probability calibration.

Explore all models on GitHub .