Model contract / exact inference
Bayesian RAPM Methodology
The first Bayesian baseline is the exact probabilistic counterpart of canonical ridge RAPM. It keeps the selected linear predictor and makes its assumptions about noise, shrinkage, and uncertainty explicit.
Likelihood
For stint \(i\), let \(x_i\) contain +1 for each home player and -1 for
each away player. Let \(\widetilde{w}_i\) be the stint possession weight
divided by the training-window mean weight. The model is
Thus a stint with twice the possession exposure has half the conditional variance. The target remains home point differential per 100 average team possessions. The intercept \(\alpha\) represents home-court advantage.
Priors
The player prior is conditional Gaussian:
where \(n\) is the number of training stints and \(\lambda\) is the normalized ridge penalty selected by chronological validation. The intercept is flat and unpenalized. The residual variance uses the scale-invariant prior
This is a baseline prior, not the final prior design. In particular, its common scale does not use age, position, draft information, season history, or offensive and defensive roles.
Exact posterior
Write \(Z=[\mathbf{1},X]\), \(W=\operatorname{diag}(\widetilde{w})\), and
The posterior precision and location are
The player portion of \(m\) is the ridge solution. This equivalence is a required numerical invariant in the implementation, not an empirical coincidence.
After integrating out \(\sigma^2\), the parameter vector has a multivariate Student-\(t\) posterior with \(n-1\) degrees of freedom. SciPy performs one dense Cholesky factorization of the roughly 583-by-583 posterior precision matrix. No MCMC, convergence diagnostic, or variational approximation is needed.
Reported uncertainty
The model reports:
- marginal posterior standard deviations and 90% credible intervals;
- posterior probability that each coefficient is positive;
- joint eligible-rank intervals and top-25/top-50 probabilities;
- held-out posterior predictive intervals and coverage;
- the full posterior location and precision Cholesky factor for future lineup contrasts and counterfactuals.
Rank summaries use joint posterior draws. Ranking marginal coefficient draws independently would discard the strong covariance created by shared lineups.
Bayesian posterior versus game bootstrap
The two distributions answer different questions:
| Distribution | Varies | Conditions on |
|---|---|---|
| Bayesian posterior | latent coefficients and residual variance | observed design, lambda, likelihood, prior |
| Complete-game bootstrap | sampled games and refitted ridge estimates | estimator, lambda, allocation policy |
Regularization can make an estimator stable across game resamples while the latent coefficients remain weakly separated under the posterior. Conversely, the Bayesian posterior does not include uncertainty from changing lambda, possession allocation, or model form. Both views belong in model review.
Fit and reproduce
uv run nba-train-bayesian-rapm 2025-26
uv run --group docs nba-build-bayesian-rapm-case-study 2025-26
The trainer refuses a zero-lambda source because the all-player Gaussian prior would be improper along unidentified directions. It validates the source ridge manifest and the exact analytical stint hash before fitting.
Next Bayesian model
The next Bayesian extension should estimate prior scales and partially pool across seasons or player information. Separate offensive and defensive effects, age curves, draft priors, and time-varying coefficients would move the model beyond conjugacy. PyMC is the preferred framework at that point; this SciPy implementation remains the exact reference case.
References
- van Wieringen, W. N. (2015). Lecture Notes on Ridge Regression. Includes the Gaussian-prior interpretation of ridge regression.
- Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. (2013). Bayesian Data Analysis, third edition. Chapman & Hall/CRC.
- Hoerl, A. E., and Kennard, R. W. (1970). "Ridge Regression: Biased Estimation for Nonorthogonal Problems." Technometrics, 12(1), 55-67. doi:10.1080/00401706.1970.10488634