NBA Model Health

Phase 3 validation — prediction accuracy, HCA/B2B calibration, FFI weight regression.2025-26 · last run 5/18/2026, 1:31:46 AM

Phase 3 — in-season evaluation only. Season-end ratings are used retroactively for prediction accuracy, so the accuracy figure is optimistically biased. True hold-out validation requires multi-season backfill (Phase 3B).

1. Prediction Accuracy

Straight-up picks using season-end adj_net ± configured HCA. Retroactive — not held-out.

Games evaluated

1,230

completed regular-season games

Straight-up accuracy

64.3%

correct picks

Brier score

0.222

lower = better (random = 0.25)

Log-loss

0.635

lower = better (random = 0.693)

2 & 3. Empirical HCA + B2B Calibration

OLS fit: home_margin ~ β₀ + β₁·adj_net_diff + β₂·home_b2b + β₃·away_b2b. 1230 games.

OLS R²

0.135

explanatory power of the model

Empirical HCA

+1.98 pts

configured: +2.10 pts/100

Model scale (β₁)

0.898

goal ≈ 1.0 (pts ≈ pts/100 poss)

Home B2B penalty

-2.49 pts

configured: −2.1

Away B2B (home benefit)

+1.06 pts

configured: +2.1 to home

4. FFI Weight Derivation

OLS: adj_net ~ w₀ + w₁·eFG_margin + w₂·TOV_edge + w₃·OREB_edge + w₄·FTA_margin across all 30 teams. R² = 0.757.

High R² — four factors explain adj_net well this season. Proposed weights below can be considered for NBA_FFI_WEIGHTS in nba/ratings_config.py after the season concludes.

Four Factor	Proposed (data-driven)	Current (configured)	Status
EFG margin	41.1%	48.0%	✓
TOV edge	43.5%	34.0%	◄ notable
OREB edge	11.3%	12.0%	✓
FTA margin	4.0%	6.0%	✓

Proposed weights are derived by running OLS on the current season's 30 teams and normalising positive coefficients to sum to 1. They reflect within-season patterns only and should not be adopted until multi-season validation is complete (Phase 3B).

Methodology Notes

Prediction accuracySeason-end adj_net ratings are used retroactively. Because these ratings already incorporate the outcomes being evaluated, accuracy is optimistically biased. All-else-equal, 69–72% is typical for mature rating systems.

OLS model scaleβ₁ converts adj_net advantage (pts/100 poss) to expected raw-point margin. An NBA game averages ~100 poss/team, so β₁ should approach 1.0 at season end.

Empirical HCADerived as the OLS intercept (β₀). Compared to home_court_adj in ratings_config.py, which is in pts/100 poss. At ~100 poss/game these are roughly the same scale.

FFI R²The four factors are mathematically related to scoring margin, so high R² is expected (≥0.95 at season end). R² here measures how well the specific four factor edges predict adj_net — not whether adj_net predicts outcomes.