Fast Algorithms for the Quantile Regression Process

TL;DR

This paper introduces two innovative algorithms that significantly accelerate the computation of quantile regression processes, enabling efficient estimation across many quantile levels with minimal loss of accuracy.

Contribution

The paper presents two novel algorithms that leverage previous estimates to drastically reduce computation time in quantile regression sequences.

Findings

Algorithms reduce computation time by up to 100 times.

Simulations confirm minimal impact on estimate quality.

Applicable to bootstrap methods for inference.

Abstract

The widespread use of quantile regression methods depends crucially on the existence of fast algorithms. Despite numerous algorithmic improvements, the computation time is still non-negligible because researchers often estimate many quantile regressions and use the bootstrap for inference. We suggest two new fast algorithms for the estimation of a sequence of quantile regressions at many quantile indexes. The first algorithm applies the preprocessing idea of Portnoy and Koenker (1997) but exploits a previously estimated quantile regression to guess the sign of the residuals. This step allows for a reduction of the effective sample size. The second algorithm starts from a previously estimated quantile regression at a similar quantile index and updates it using a single Newton-Raphson iteration. The first algorithm is exact, while the second is only asymptotically equivalent to the traditional quantile regression estimator. We also apply the preprocessing idea to the bootstrap by using the sample estimates to guess the sign of the residuals in the bootstrap sample. Simulations show that our new algorithms provide very large improvements in computation time without significant (if any) cost in the quality of the estimates. For instance, we divide by 100 the time required to estimate 99 quantile regressions with 20 regressors and 50,000 observations.