Adaptive Modeling Procedure Selection by Data Perturbation

Yongli Zhang, Xiaotong Shen

Research output: Contribution to journalArticlepeer-review

1 Scopus citations


Many procedures have been developed to deal with the high-dimensional problem that is emerging in various business and economics areas. To evaluate and compare these procedures, modeling uncertainty caused by model selection and parameter estimation has to be assessed and integrated into a modeling process. To do this, a data perturbation method estimates the modeling uncertainty inherited in a selection process by perturbing the data. Critical to data perturbation is the size of perturbation, as the perturbed data should resemble the original dataset. To account for the modeling uncertainty, we derive the optimal size of perturbation, which adapts to the data, the model space, and other relevant factors in the context of linear regression. On this basis, we develop an adaptive data-perturbation method that, unlike its nonadaptive counterpart, performs well in different situations. This leads to a data-adaptive model selection method. Both theoretical and numerical analysis suggest that the data-adaptive model selection method adapts to distinct situations in that it yields consistent model selection and optimal prediction, without knowing which situation exists a priori. The proposed method is applied to real data from the commodity market and outperforms its competitors in terms of price forecasting accuracy.

Original languageEnglish (US)
Pages (from-to)541-551
Number of pages11
JournalJournal of Business and Economic Statistics
Issue number4
StatePublished - Oct 2 2015

Bibliographical note

Publisher Copyright:
© 2015 American Statistical Association.


  • Adaptive model selection
  • High-dimensional data analysis
  • Modeling uncertainty


Dive into the research topics of 'Adaptive Modeling Procedure Selection by Data Perturbation'. Together they form a unique fingerprint.

Cite this