Abstract
Since the early 1970s, some estimation-type identification procedures have been proposed. They are to choose the orders k and i minimizing
, where σ 2 k,i is an estimate of the white noise variance obtained by fitting the ARMA(k, i) model to the observations. Because σ 2 k,i decreases as the orders increase, it cannot be a good criterion to choose the orders minimizing it. If the orders increase, the bias of the estimated model will decrease while the variance increases. Therefore, we should compromise between them. For this purpose we add the penalty term, (k + i)C(T)/T, into the model selection criterion The penalty function identification methods are regarded as objective.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Preview
Unable to display preview. Download preview PDF.
Additional References
About choosing upper bounds of the AR and the MA orders, readers may refer to An, Chen, and Hannan (1982), Hannan and Rissanen (1982), Hannan and Kavalieris (1984b, 1986a), Poskitt (1987), Hannan and Deistler (1988), and the references therein.
The FPE procedure has been used for statistical modeling beyond AR order determination. McClave (1975) utilized Hocking and Leslies’ subset regression technique (1967) with the FPE for AR model identification. The FPE procedure was studied for vector AR processes by Akaike (1971), Reinsel (1980, 1983), and Jones (1976), and for more general stochastic processes by Baillie (1979b) and Toyooka (1982). Hsiao (1979) used it for Granger causality tests. Akaike (1969b, 1970b ), Gersch and Sharpe (1973), and Jones (1974) used the AR model and the FPE criterion to estimate spectral densities. Their numerical examples have shown that the MFPEE and the YW estimates of AR coefficients result in reasonable spectral density estimates. However, some examples in disagreement were presented by Marple (1980).
For more details about the asymptotic mean square error of a multi-step ahead predictor, readers may refer to Box and Jenkins (1976, p. 267), Bloomfield (1972), Bhansali (1978), Schmidt (1974), Janacek (1975), Yamamoto (1976, 1981), Baillie (1979a), Davies and Newbold (1980), Reinsel (1980), Shibata (1980), Ledolter and Abraham (1981), Puller and Hasza (1981), Newton and Pagano (1983), Fotopoulos and Ray (1983), and the references therein.
Readers who are interested in Sanov’s theorem may refer to Bahadur and Zabell (1979), Vincze (1982), Deuschel and Strook (1989), and the references therein.
The conditional probability characterization of the Kullback-Leibler information number has been discussed by Vasicek (1980), Csiszâr, Cover and Choi (1987), and Choi (1991b).
The AIC was used to select optimal models in many fields of statistics. Akaike (1971-1983), Gersch and Kitagawa (1983), and others utilized it to determine the orders of ARMA processes. Kitagawa (1981) applied the AIC to model fitting for nonstationary time series. Kozin and Nakajima (1980) used the AIC for time-varying AR models. Gabr and Subba Rao (1981) applied it to bilinear time series models. Jones (1974), Sakai (1981), Quinn (1980b, 1988), and Paulsen and Tjostheim (1985) proposed using the AIC for determining the orders of vector AR processes. It was also used in factor analysis by Akaike (1972b, 1975) and Tong (1975a), in regression analysis by Sawa (1978) and Shibata (1981a, 1984), in the analysis of Markov processes by Tong (1975b), in the analysis of distributed lag model by Tong (1976), in the analysis of covariance by Akaike (1977a), in signal processing analysis by Tong (1975a, 1977) and Findley (1984), and for determining the histogram width by Taylor (1987). Other possible applications have been suggested by Akaike (1973a, 1977a) and Sugiura (1978).
For more details of the asymptotic distribution of the MAICE, refer to Hannan and Deistler (1988, Section 5.6). Sakai (1981), Paulsen and Tjostheim (1985), and Quinn (1988) derived the asymptotic distribution of the MAICE for vector AR processes.
Some illustrative examples of the CAT procedure were given by Parzen (1979a, 1979b, 1980a) and Parzen and Pagano (1979). The CAT for vector AR processes has been proposed by Parzen (1977) and Parzen and Newton (1980).
For E. J. Hannan’s opinion about the AIC, refer to Hannan and Quinn (1979, p. 195) and Hannan (1980b, p. 1072 ); ( 1982, p. 411 ).
For the HQC method, readers may also refer to Heyde and Scott (1973) and Bai, Subramanyam, and Zhao (1988). Quinn (1980b) has generalized the MHQCE to vector AR models and has shown its strong consistency.
For more details about the consistency problem of the penalty function methods, refer to Hannan (1981), Hannan and Deistler ( 1988, Section 5.4), An and Chen (1986), and the references therein. Potscher (1990) has shown that if f is the estimate of r = max(p, q) having the first “local” minimum of the BIC under the assumption k = i.
Rissanen (1986b) has derived Rissanen’s lower bound using coding theory. Kabaila (1987) has shown that under some fairly strong restrictions it can be derived via the Cramer-Rao lower bound or the Fisher bound on asymptotic variances for the case of Gaussian AR processes. Also, refer to Hannan, McDougall, and Poskitt (1989).
There have been some recent advances in using cross-validation procedures in time series analysis. Some applications have been considered by Geisser and Eddy (1979), Bessler and Binkley (1980), and Hjorth and Holmqvist (1981). Hurvich and Beltrao (1990) have presented the cross-validated log-likelihood criterion, which can be viewed as a cross-validatory generalization of the AIC. Also, refer to Hurvich and Zeger (1990). Stoica, Eykhoff, Janssen, and Soderstrom (1986) have presented another cross-validation method, which yields asymptotically the same result as the BIC procedure. Also, refer to Jong (1988).
Tjostheim and Paulsen (1985) have applied the penalty function identification methods to a particular nonstationary AR process, where the variance of the innovation process depends on time.
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 1992 Applied Probability Trust
About this chapter
Cite this chapter
Choi, B. (1992). Penalty Function Methods. In: ARMA Model Identification. Springer Series in Statistics. Springer, New York, NY. https://doi.org/10.1007/978-1-4613-9745-8_3
Download citation
DOI: https://doi.org/10.1007/978-1-4613-9745-8_3
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4613-9747-2
Online ISBN: 978-1-4613-9745-8
eBook Packages: Springer Book Archive