A three-hierarchy Bayesian analysis of San Antonio's playoff and regular-season statistics against the Knicks — updated after the Spurs' Game 3 road win at Madison Square Garden.
by Charlie of Mr. Danoff's Teaching Laboratory · Last Updated June 10th, 2026
The Spurs' Game 3 victory (115-111 away) dramatically shifted their comeback prospects from near-zero to realistic, with the model showing:
The remaining schedule still runs two of the next three games in New York, so the away win probability is the critical number — and Game 3 showed they can get it.
Remaining schedule: G4 Away · G5 Home · G6 Away · G7 Home — NYK needs 2 more wins, SAS needs 3
| Scenario | P(win home) | P(win away) | P(comeback) | 90% CI |
|---|---|---|---|---|
| Finals avg (G1-G3) baseline | 14.6% | 17.3% | 4.9% | [0.0%, 27.4%] |
| Game 3 level last game | 66.6% | 70.1% | 62.1% | [2.8%, 99.8%] |
| Midpoint improvement scenario | 52.5% | 56.1% | 42.0% | [0.9%, 92.7%] |
| Other playoff win avg ceiling | 87.4% | 89.5% | 89.6% | [47.2%, 100.0%] |
Probabilities are posterior medians from the v3 model (3-hierarchy partial pooling, 10,000 MCMC draws). Run spurs_bayesian_model_v3.py after Game 4 for updated numbers.
Simulated from 5,000 browser-side draws using beta-distributed posteriors. The Game 3 and midpoint scenarios now overlap significantly — the Spurs' Game 3 performance was close to their historical winning profile, not an outlier.
Observed averages from SportRadar box scores — descriptive statistics, not model coefficients. The logistic regression assigns highest posterior confidence to assists and turnovers as predictors of winning.
Three-hierarchy Bayesian hierarchical logistic regression with partial pooling across groups: Group 0 = NYK Finals (n=3); Group 1 = Pre-Finals SAS vs NYK matchups — two regular season games and the NBA Cup Final (n=3); Group 2 = All other 2026 SAS playoff games (n=18).
Fit via adaptive Metropolis-Hastings, 10,000 post-warmup draws. Predictors: FG%, rebounds, assists, steals, blocks, turnovers, home/away indicator — all standardised. The third hierarchy adds matchup-specific information the earlier model lacked.