Chapter 16: Key Takeaways - NBA Modeling

1. The Four Factors framework is the backbone of NBA modeling. Effective field goal percentage, turnover rate, offensive rebounding rate, and free throw rate on both sides of the ball capture the vast majority of what drives NBA team performance. Build your model around these metrics before adding complexity.

2. Everything must be pace-adjusted. Raw per-game statistics are misleading because NBA teams play at vastly different speeds. Always convert to per-100-possession metrics when comparing team or player efficiency. Predict game pace explicitly when forecasting totals.

3. Net rating is the best single predictor of NBA team quality. Offensive rating minus defensive rating, calculated on a per-100-possession basis, explains more future performance variance than win-loss record, margin of victory, or any single box score statistic.

4. Rest and schedule effects are real and measurable. Back-to-back games, long road trips, and travel across time zones all impact NBA performance by approximately 1-3 points. These effects are sometimes mispriced by the market, particularly early in the season and for less-visible mid-week games.

5. The NBA's large sample size is an advantage for modelers. With 82 regular-season games, team metrics stabilize relatively quickly (15-20 games for most key statistics). This allows models to rely more on current-season data and less on priors compared to shorter-season sports like the NFL.

6. Player-level modeling matters more in the NBA than in most sports. Individual NBA players have outsized impact on team performance. A single star player's absence can shift a point spread by 3-8 points. Use on/off court data to estimate player impact, but regress toward box-score-based priors to manage noise.

7. Three-point shooting variance drives game-level uncertainty. Teams relying heavily on three-point shooting have more volatile game-to-game performance. This is a key consideration for totals modeling and explains why high-volume three-point teams have less predictable ATS results.

8. Garbage time systematically affects large-favorite ATS results. When large favorites build big leads, garbage time allows opponents to close the gap with meaningless points. This creates a persistent bias against large favorites covering the spread that should be incorporated into any ATS prediction model.

9. Closing line value (CLV) is the best measure of betting skill. If your model consistently identifies edges that the market later confirms through line movement, you are likely making profitable predictions even if short-term results are noisy. Track CLV religiously.

10. Playoff and regular-season basketball are different. Playoff games feature slower pace, higher defensive intensity, and more half-court play. Totals models should adjust downward by 3-5 points for playoff games. Spread models should account for heightened home-court advantage and reduced variance from more deliberate play.

11. Mid-season roster changes create temporary inefficiencies. The market adjusts to trades and buyouts over 1-2 weeks. A model that can quickly estimate the impact of a roster change and compare it to the market's adjustment may find value in the immediate post-trade window.

12. Bayesian updating is ideal for the NBA's evolving information environment. Start with strong preseason priors, update with each game's data, and let the model smoothly transition from prior-heavy to data-heavy estimates over the first quarter of the season. This outperforms both pure-prior and pure-data approaches in early-season predictions.