About NBA Analytics

Last updated: May 2026

What Is NBA Analytics?

NBA Analytics is an independent site that runs a machine-learning model on NBA games. Every day the model looks at each matchup, produces a win probability and an implied point spread, and compares that to the Vegas line. When our number differs meaningfully from Vegas, we flag the game as a LOCK.

The model is just math. No AI chatbot, no proprietary "smart money" feed. Just seven team-level features fed into a gradient-boosted decision tree, with the probabilities calibrated against historical games so the confidence numbers we publish actually match observed win rates.

How the Model Works (Short Version)

For each game, we compute seven inputs about the two teams and feed them into an XGBoost classifier trained on roughly 7,600 historical NBA games:

• ELO rating gap — each team's long-run strength rating using the FiveThirtyEight ELO system; usually the model's most important input
• 20-game point differential — rolling average of how much each team has beaten or lost to opponents recently
• 5-game and 10-game point differential — same idea over shorter windows; reacts faster to recent form
• Rest day gap — days since each team's last game
• Home back-to-back flag — whether the home team played the day before
• Visitor back-to-back flag — same for the visiting team (visitors on B2B are roughly 5-7% worse historically)

The model's output probability is then passed through an isotonic regression calibration layer fit on a held-out portion of games. This is what makes confidence numbers we publish actually match observed win rates in our backtest — see the Track Record page for the full reliability table.

For the long version with the reasoning behind each choice, see How It Works.

How Good Is It?

Across 7,146 historical regular-season games, the model picks the straight-up winner about 67.1% of the time. That lands right at the FiveThirtyEight ELO benchmark and roughly matches the Vegas closing line — the practical ceiling for any pre-game NBA model. On games where the model is most confident (predicted >65%), it hits about 75%.

It's wrong about a third of the time. No prediction service is right every time, and ours isn't either. The Track Record page shows every pick with monthly accuracy trends — go read it before trusting a LOCK.

What the Model Can't See

Calibration is honest about confidence; it doesn't make the model omniscient. A few things our seven features cannot capture:

• Late-breaking injuries or lineup changes (we surface injury context separately for reading but don't bake it into the model's feature vector)
• Player-level effects beyond team aggregates — coaching matchups, individual hot streaks
• Specific betting-market shifts after we generate the prediction
• Anything Vegas knows that we don't — sharps have access to information we don't

Important Disclaimers

This site is for entertainment and informational purposes only. Nothing here constitutes gambling advice. Sports betting involves real financial risk. Past performance does not guarantee future results. Never bet more than you can afford to lose.

Must be 21+ and in a jurisdiction where sports betting is legal to use any linked sportsbooks. Gambling problem? Call 1-800-GAMBLER.

Affiliate Disclosure

NBA Analytics participates in affiliate marketing programs with licensed US sportsbooks. When you click a sportsbook link on this site and create an account, we may receive a commission at no additional cost to you. This keeps the site free. Our picks and model outputs are never influenced by affiliate relationships — the model sees the same seven numbers regardless of which sportsbook is featured.

Contact

Questions, feedback, or press inquiries: ben.g.ballard@gmail.com

We read everything and aim to respond within 48 hours.

---