How Our AI Model Predicts NBA Games

Every day during the NBA season, our model analyzes each matchup to find edges the market may have missed. Here's how it works.

The Model: XGBoost

We use XGBoost, a gradient-boosted decision tree algorithm that excels at structured data problems. The model was trained on thousands of historical NBA games, learning the patterns that separate home wins from away wins.

Key Features

The model considers several inputs for each game:

• Net Rating — The gold-standard measure of team quality (points scored minus points allowed per 100 possessions)
• Home Court Advantage — Still a real factor in the NBA, though it's shrunk over the years
• Rest Days — Back-to-back games measurably hurt performance, especially on the road
• Offensive and Defensive Ratings — Breaking net rating into its components reveals matchup-specific edges

Finding the Edge

The model outputs a win probability for each team. We convert that into an implied spread and compare it to the Vegas line. When our spread differs from Vegas by more than 2 points, we flag it as a play.

A 2+ point edge doesn't mean we're always right — it means the market price is off enough to offer long-term value. That's how sharp bettors think.

What's Next

We're continuously improving the model by adding player-level injury data, pace-adjusted matchup stats, and larger training windows. Follow along as we refine the edge.