Helping the Oracle: Vector Sign Constraints for Model Shrinkage Methodologies


METADATA ONLY
Loading...

Date

2023

Publication Type

Conference Paper

ETH Bibliography

yes

Citations

Altmetric
METADATA ONLY

Data

Rights / License

Abstract

Motivated by the need for obtaining econometric models with theory-conform signs of long-run multipliers or other groups of predictor variables for numerous purposes in economics and other scientific disciplines, we develop a vector sign constrained variant of existing model shrinkage methodologies, such as (Adaptive) Lasso and (Adaptive) Elastic Net. A battery of Monte Carlo experiments is used to illustrate that the addition of such constraints "helps the Oracle property" (the ability to identify the true model) for those methods that do not initially carry it, such as Lasso. For methods that possess it already (the adaptive variants), the constraints help increase efficiency in finite data samples. An application of the methods to empirical default rate data and their macro-financial drivers for 16 countries from Europe and the U.S. corroborates the avail of the sign constrained methodology in empirical applications for obtaining more robust economic models.

Publication status

published

Book title

Machine Learning, Optimization, and Data Science

Volume

13810

Pages / Article No.

444 - 458

Publisher

Springer

Event

8th International Conference on Machine Learning, Optimization and Data Science (LOD 2022)

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Machine learning; Model selection; Oracle property

Organisational unit

Notes

Funding

Related publications and datasets