NBA Game Prediction Models

Trained on 17,168 matches from 2009-present

Test Set Performance
ml_logistic_team12_player_2_051124 (Most recent model)
65.1%
Accuracy
65.3%
Precision
76.8%
Recall
69.2%
ROC AUC
Model Description

Our most recent NBA game prediction model, ml_logistic_team12_player_2_051124, uses a combination of advanced statistics and key game context to make its decisions. Here’s a breakdown of the features it considers:

  • Team Strength: The model compares the overall strength of the home and away teams, measured using a custom rating system. This strength difference accounts for how well teams have been performing recently.
  • Starting Lineup Strength: It takes into account the relative strength of each team’s starting lineup, helping to predict how the starting five players might impact the game's outcome.
  • Division Performance: We also factor in how well teams perform against others in their division, which can influence their expected performance.
  • Game Timing: The model understands that game timing matters. It knows when a game occurs in the season and adjusts expectations accordingly, differentiating between early-season games and those played during more intense, late-season matchups.
  • Recent Form Metrics: Shooting percentages over the last 10 games for each team, like field goal and free throw accuracy, are included to understand recent offensive performance.
  • Fatigue Considerations: The model accounts for whether a team is playing on back-to-back nights, as this can significantly impact player energy and game strategy.
  • Game Context: Features like the month of the game and how deep a game falls into the season provide additional context, helping the model to adapt based on how teams typically perform at different times of the year.