Asymptotic Overfitting Probabilities of some Model Selection Criteria in Large Samples

• 발행기관: 한국자료분석학회
• 저자명: 박찬근(Chan Keun Park) 조영석(Young-Seuk Cho)
• 간행물 정보: 『Journal of The Korean Data Analysis Society (JKDAS)』Vol.9 No.3, 1037~1048쪽, 전체 12쪽
• 주제분류: 자연과학 > 통계학
• 파일형태: PDF
• 발행일자: 2007.06.30

영문 초록

We look at overfitting in large samples by obtaining asymptotic probabilities of some model selection criteria, which are AIC, AICc, AICu, SIC, HQ, HQc, Cp, FPE, and FPEu. Asymptotic Overfitting probabilities refer to comparing one overfitted model of order k_{*} +L to the reduced model of order k_{*}. We make use of the following fact : As n rarrow INF and where k_{*} and L are fixed, {chi_{ n-k_{*} -L}^{2}} over { n-k_{*} -L} rarrow 1. It is clear that AIC, AICc, FPE and Cp are not constant criteria, since their probabilities of overfitting are not 0. But SIC, HQ, and HQc are consistent and thus do have 0 probability of overfitting.

목차

1. Introduction

2. Review of model selection criteria

3. Asymptotic Probabilities of Overfitting

4. Conclusions

References

키워드

Model Selection Criteria Asymptotic Overfitting Probabilities χ² distribution