The impact of unsuspected serial correlations on model selection in linear regression
We obtain the asymptotic distributions of the linear regression models selected by the Akaike Information Criterion (AIC) and the Schwarz Information Criterion (BIC), in the presence of unsuspected serial correlations. We assume that the models are fitted by ordinary least squares to a data set observed at equally spaced intervals of time, and that the errors are a stationary Gaussian process. We also assume that the true model is of finite dimension, and that the set of candidate models is held fixed as the sample size increases. We will show that, for AIC, the asymptotic probability of underfitting is zero, regardless of the error correlation structure, while the asymptotic probability of overfitting depends on the error correlations. Furthermore, we will show that BIC is consistent. Our results generalize those of Nishii (1984), who assumed an i.i.d. Gaussian error structure. Finally, we will present the results of a simulation study, which provides support for the applicability in moderate sample sizes of our asymptotic theorems.
Year of publication: |
1996
|
---|---|
Authors: | Hurvich, Clifford M. ; Tsai, Chih-Ling |
Published in: |
Statistics & Probability Letters. - Elsevier, ISSN 0167-7152. - Vol. 27.1996, 2, p. 115-126
|
Publisher: |
Elsevier |
Subject: | AIC BIC |
Saved in:
Online Resource
Saved in favorites
Similar items by person
-
Model selection for least absolute deviations regression in small samples
Hurvich, Clifford M., (1990)
-
Markov-Switching Model Selection Using Kullback-Leibler Divergence
Smith, Aaron, (2005)
-
Market uncertainty and market orders in futures markets
Chang, Matthew C., (2018)
- More ...