quantile-regression-pdlp 0.4.0

Quantile regression via linear programming using Google OR-Tools PDLP, with a scikit-learn compatible API and statistical summaries.

quantile-regression-pdlp

Optimization-based quantile regression built on Google OR-Tools. Scikit-learn API, statsmodels-style summaries, and features that go beyond what either package offers.

What makes this different from sklearn or statsmodels?

• Fits multiple quantiles jointly with non-crossing constraints
• Multi-output regression in a single model
• SCAD, MCP, and elastic net penalties (not just L1)
• Analytical, bootstrap, kernel, and cluster-robust standard errors
• Conformalized quantile regression for calibrated prediction intervals
• Evaluation metrics: pinball loss, coverage, interval score, crossing diagnostics
• Calibration diagnostics: coverage by group/bin, nominal vs empirical, sharpness analysis
• Crossing detection and rearrangement for any quantile model's predictions
• Prediction intervals, quantile process plots, and pseudo R²
• Censored quantile regression for survival data
• Scipy sparse solver for large-scale problems
• Validated against sklearn, statsmodels, and R's quantreg

Feature	This package	sklearn	statsmodels
Multiple quantiles (joint)	Yes	No	No
Non-crossing constraints	Yes	No	No
Multi-output	Yes	No	No
Analytical SEs	Yes	No	Yes
Kernel (robust) SEs	Yes	No	Yes
Cluster-robust SEs	Yes	No	No
Bootstrap SEs	Yes	No	No
L1 / Elastic Net / SCAD / MCP	Yes	L1 only	No
Conformal calibration (CQR)	Yes	No	No
Evaluation metrics suite	Yes	Partial	No
Crossing detection + fix	Yes	No	No
Calibration diagnostics	Yes	No	No
Prediction intervals	Yes	No	No
Pseudo R²	Yes	No	Yes
Formula interface	Yes	No	Yes
Censored QR	Yes	No	No
Sklearn pipeline compatible	Yes	Yes	No

Installation

pip install quantile-regression-pdlp

Optional extras:

pip install quantile-regression-pdlp[all]   # formula interface + plots
pip install quantile-regression-pdlp[plot]   # matplotlib only
pip install quantile-regression-pdlp[formula] # patsy only

Quick Start

import numpy as np
from quantile_regression_pdlp import QuantileRegression

X = np.random.default_rng(0).normal(size=(200, 3))
y = X @ [2.0, -1.5, 0.8] + np.random.default_rng(1).normal(scale=0.5, size=200)

model = QuantileRegression(tau=[0.1, 0.5, 0.9], n_bootstrap=200, random_state=0)
model.fit(X, y)

# Summaries with coefficients, SEs, p-values, and 95% CIs
print(model.summary()[0.5]['y'])

# Prediction intervals
interval = model.predict_interval(X[:5], coverage=0.80)
print(interval['y']['lower'], interval['y']['upper'])

# Pseudo R²
print(model.pseudo_r_squared_)

Features at a Glance

Regularization

# L1 (Lasso)
QuantileRegression(tau=0.5, regularization='l1', alpha=0.1)

# Elastic net
QuantileRegression(tau=0.5, regularization='elasticnet', alpha=0.1, l1_ratio=0.5)

# SCAD (less bias on large coefficients)
QuantileRegression(tau=0.5, regularization='scad', alpha=0.3)

# MCP
QuantileRegression(tau=0.5, regularization='mcp', alpha=0.3)

Inference Options

# Fast analytical SEs (no bootstrapping needed)
model = QuantileRegression(tau=0.5, se_method='analytical')
model.fit(X, y)

# Heteroscedasticity-robust kernel sandwich SEs
model = QuantileRegression(tau=0.5, se_method='kernel')
model.fit(X, y)

# Cluster-robust SEs
model = QuantileRegression(tau=0.5, se_method='analytical')
model.fit(X, y, clusters=group_labels)

Quantile Process Plot

model = QuantileRegression(
    tau=[0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95],
    se_method='analytical'
)
model.fit(X, y)
model.plot_quantile_process(feature='X1')

Formula Interface

model = QuantileRegression(tau=0.5, se_method='analytical')
model.fit_formula('y ~ x1 + x2 + C(region)', data=df)

Censored Quantile Regression

from quantile_regression_pdlp import CensoredQuantileRegression

model = CensoredQuantileRegression(tau=0.5, censoring='right', se_method='analytical')
model.fit(X, observed_time, event_indicator=delta)

Solver Options

# GLOP simplex (faster on small/medium problems)
QuantileRegression(tau=0.5, solver_backend='GLOP')

# Scipy sparse solver (memory-efficient for large datasets)
QuantileRegression(tau=0.5, use_sparse=True)

# Solver tuning
QuantileRegression(tau=0.5, solver_tol=1e-8, solver_time_limit=60.0)

Benchmarks

Validated against sklearn QuantileRegressor and statsmodels QuantReg across multiple dataset sizes and quantile configurations. All three solve the same LP, so pinball loss is equivalent — the difference is in what this package provides around the fit.

n	quantiles	Pinball loss (all equal)	Crossing rate	This package extras
500	3	0.2291	0% (all)	Joint fit, non-crossing guarantee, SEs, CQR
2000	5	0.2238	0% (all)	+ evaluation metrics, crossing tools
5000	5	0.2347	0% (all)	+ censored QR, SCAD/MCP/elastic net

Full benchmark results, methodology, and reproduction instructions: Benchmarks docs

# Reproduce benchmarks locally
pip install -e ".[benchmark]"
python benchmarks/run_linear_baselines.py
python benchmarks/report.py

Documentation

Full docs: joshvern.github.io/quantile_regression_pdlp

Why PDLP?

Quantile regression is naturally a linear program. OR-Tools' PDLP is a first-order solver designed for large-scale LPs, making it efficient for high-dimensional problems. For smaller problems, the package also supports GLOP (simplex) and scipy's HiGHS solver.

Dependencies

Required: ortools, numpy, pandas, scipy, tqdm, joblib, scikit-learn

Optional: matplotlib (plots), patsy (formulas)

Contributing

Contributions welcome! Open an issue or submit a pull request on GitHub.

License

MIT