Aggregated expectile regression by exponential weighting

Yuwen Gu, Hui Zou

Research output: Contribution to journalArticle

Abstract

Various estimators have been proposed to estimate conditional expectiles, including those from multiple linear expectile regression, local polynomial expectile regression, boosted expectile regression, and so on. It is a common practice that several plausible candidate estimators are fitted and a final estimator is selected from the candidate list. In this article, we advocate the use of an exponential weighting scheme to adaptively aggregate the candidate estimators into a final estimator. We show oracle inequalities for the aggregated estimator. Simulations and data examples demonstrate that the aggregated estimator could have substantial gain in accuracy under both squared and asymmetric squared errors.

Original languageEnglish (US)
Pages (from-to)671-692
Number of pages22
JournalStatistica Sinica
Volume29
Issue number2
DOIs
StatePublished - Jan 1 2019

Fingerprint

Weighting
Regression
Estimator
Local Polynomial Regression
Oracle Inequalities
Multiple Linear Regression
Estimate
Demonstrate
Simulation

Keywords

  • Cross-validation
  • Expectile regression
  • Model aggregation
  • Oracle inequality

Cite this

Aggregated expectile regression by exponential weighting. / Gu, Yuwen; Zou, Hui.

In: Statistica Sinica, Vol. 29, No. 2, 01.01.2019, p. 671-692.

Research output: Contribution to journalArticle

Gu, Yuwen ; Zou, Hui. / Aggregated expectile regression by exponential weighting. In: Statistica Sinica. 2019 ; Vol. 29, No. 2. pp. 671-692.
@article{34cc9fe71a84466b825d9dd5109836ee,
title = "Aggregated expectile regression by exponential weighting",
abstract = "Various estimators have been proposed to estimate conditional expectiles, including those from multiple linear expectile regression, local polynomial expectile regression, boosted expectile regression, and so on. It is a common practice that several plausible candidate estimators are fitted and a final estimator is selected from the candidate list. In this article, we advocate the use of an exponential weighting scheme to adaptively aggregate the candidate estimators into a final estimator. We show oracle inequalities for the aggregated estimator. Simulations and data examples demonstrate that the aggregated estimator could have substantial gain in accuracy under both squared and asymmetric squared errors.",
keywords = "Cross-validation, Expectile regression, Model aggregation, Oracle inequality",
author = "Yuwen Gu and Hui Zou",
year = "2019",
month = "1",
day = "1",
doi = "10.5705/ss.202016.0285",
language = "English (US)",
volume = "29",
pages = "671--692",
journal = "Statistica Sinica",
issn = "1017-0405",
publisher = "Institute of Statistical Science",
number = "2",

}

TY - JOUR

T1 - Aggregated expectile regression by exponential weighting

AU - Gu, Yuwen

AU - Zou, Hui

PY - 2019/1/1

Y1 - 2019/1/1

N2 - Various estimators have been proposed to estimate conditional expectiles, including those from multiple linear expectile regression, local polynomial expectile regression, boosted expectile regression, and so on. It is a common practice that several plausible candidate estimators are fitted and a final estimator is selected from the candidate list. In this article, we advocate the use of an exponential weighting scheme to adaptively aggregate the candidate estimators into a final estimator. We show oracle inequalities for the aggregated estimator. Simulations and data examples demonstrate that the aggregated estimator could have substantial gain in accuracy under both squared and asymmetric squared errors.

AB - Various estimators have been proposed to estimate conditional expectiles, including those from multiple linear expectile regression, local polynomial expectile regression, boosted expectile regression, and so on. It is a common practice that several plausible candidate estimators are fitted and a final estimator is selected from the candidate list. In this article, we advocate the use of an exponential weighting scheme to adaptively aggregate the candidate estimators into a final estimator. We show oracle inequalities for the aggregated estimator. Simulations and data examples demonstrate that the aggregated estimator could have substantial gain in accuracy under both squared and asymmetric squared errors.

KW - Cross-validation

KW - Expectile regression

KW - Model aggregation

KW - Oracle inequality

UR - http://www.scopus.com/inward/record.url?scp=85072091306&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85072091306&partnerID=8YFLogxK

U2 - 10.5705/ss.202016.0285

DO - 10.5705/ss.202016.0285

M3 - Article

AN - SCOPUS:85072091306

VL - 29

SP - 671

EP - 692

JO - Statistica Sinica

JF - Statistica Sinica

SN - 1017-0405

IS - 2

ER -