Pinball

Fast, accurate quantile regression for Python.

---

Pinball brings the speed and statistical rigor of R's quantreg package to Python, wrapped in a familiar scikit-learn interface. It provides:

• High-performance Fortran solvers — the Barrodale-Roberts simplex and Frisch-Newton interior-point algorithms, compiled from the original quantreg Fortran code
• Preprocessing for massive datasets — the Portnoy-Koenker "globs" technique that reduces million-row problems to manageable size
• Full statistical inference — standard errors, confidence intervals, and three bootstrap methods (xy-pair, wild, MCMB)
• L1-penalised (lasso) quantile regression — for high-dimensional sparse models using the Belloni-Chernozhukov approach
• Nonparametric conditional quantiles — the Charlier-Paindaveine-Saracco quantization-based estimator for when linearity doesn't hold
• 100% scikit-learn compatible — passes check_estimator for both linear and nonparametric estimators

Quick Example

from pinball import QuantileRegressor, load_engel

X, y = load_engel(return_X_y=True)
model = QuantileRegressor(tau=[0.1, 0.25, 0.5, 0.75, 0.9], method="fn")
model.fit(X, y)
print(model.coef_)

Why "Pinball"?

The pinball loss (also called the check function or asymmetric absolute loss) is the objective function at the heart of quantile regression:

\[ \rho_\tau(u) = u \cdot (\tau - \mathbf{1}_{u < 0}) \]

It penalises underestimates by \(\tau\) and overestimates by \(1 - \tau\), making it the natural loss for estimating the \(\tau\)-th conditional quantile.

Project Lineage

Pinball is a faithful port of two R packages:

Component	R Source	Reference
Linear solvers & inference	quantreg	Koenker (2005), Quantile Regression
Preprocessing	quantreg (rq.fit.pfn)	Portnoy & Koenker (1997), Statistical Science
Nonparametric estimator	QuantifQuantile	Charlier, Paindaveine & Saracco (2015), JSPI