Clifford M. Hurvich

The mean squared error of Geweke and Porter-Hudak's estimator of the memory parameter of a long-memory time series

We establish some asymptotic properties of a log-periodogram regression estimator for the memory parameter of a long-memory time series. We consider the estimator originally proposed by Geweke and Porter-Hudak (The estimation and application of long memory time series models. Journal of Time Ser. Anal. 4 (1983), 221–37). In particular, we do not omit any of the low frequency periodogram ordinates from the regression. We derive expressions for the estimators asymptotic bias, variance and mean squared error as functions of the number of periodogram ordinates, m, used in the regression. Consistency of the estimator is obtained as long as m∞ and n∞ with (m log m)/n 0, where n is the sample size. Under these and the additional conditions assumed in this paper, the optimal m, minimizing the mean squared error, is of order O(n4/5). We also establish the asymptotic normality of the estimator. In a simulation study, we assess the accuracy of our asymptotic theory on mean squared error for finite sample sizes. One finding is that the choice m = n1/2, originally suggested by Geweke and Porter-Hudak (1983), can lead to performance which is markedly inferior to that of the optimal choice, even in reasonably small samples.

Model Selection for Extended Quasi-Likelihood Models in Small Samples

We develop a small sample criterion (AICc) for the selection of extended quasi-likelihood models. In contrast to the Akaike information criterion (AIC). AICc provides a more nearly unbiased estimator for the expected Kullback-Leibler information. Consequently, it often selects better models than AIC in small samples. For the logistic regression model, Monte Carlo results show that AICc outperforms AIC, Pregibons (1979, Data Analytic Methods for Generalized Linear Models. Ph.D. thesis. University of Toronto) Cp*, and the Cp selection criteria of Hosmer et al. (1989, Biometrics 45, 1265-1270). Two examples are presented.

The Impact of Model Selection on Inference in Linear Regression

Abstract Model selection and inference are usually treated as separate stages of regression analysis, even though both tasks are performed on the same set of data. Once a model has been selected, one typically proceeds as though one has a fresh data set generated by the selected model. Here, we present Monte Carlo results on the coverage rates of confidence regions for the regression parameters, conditional on the selected model order. The conditional coverage rates are much smaller than the nominal coverage rates, obtained by assuming that the model was known in advance. Furthermore, the overall coverage rate is much smaller than the nominal value. A possible remedy based on data splitting is suggested.

Model selection for least absolute deviations regression in small samples

We develop a small sample criterion (L1cAIC) for the selection of least absolute deviations regression models. In contrast to AIC (Akaike, 1973), L1cAIC provides an exactly unbiased estimator for the expected Kullback--Leibler information, assuming that the errors have a double exponential distribution and the model is not underfitted. In a Monte Carlo study, L1cAIC is found to perform much better than AIC and AICR (Ronchetti, 1985). A small sample criterion developed for normal least squares regression (cAIC, Hurvich and Tsai, 1988) is found to perform as well as L1cAIC. Further, cAIC is less computationally intensive than L1cAIC.

Semiparametric Estimation of Multivariate Fractional Cointegration

We consider the semiparametric estimation of fractional cointegration in a multivariate process of cointegrating rank r > 0. We estimate the cointegrating relationships by the eigenvectors corresponding to the r smallest eigenvalues of an averaged periodogram matrix of tapered, differenced observations. The number of frequencies m used in the periodogram average is held fixed as the sample size grows. We first show that the averaged periodogram matrix converges in distribution to a singular matrix whose null eigenvectors span the space of cointegrating vectors. We then show that the angle between the estimated cointegrating vectors and the space of true cointegrating vectors is Op(ndu−d), where d and du are the memory parameters of the observations and cointegrating errors. The proposed estimator is invariant to the labeling of the component series and thus does not require that one of the variables be specified as a dependent variable. We determine the rate of convergence of the r smallest eigenvalues of the periodogram matrix and present a criterion that allows for consistent estimation of r. Finally, we apply our methodology to the analysis of fractional cointegration in interest rates.

Semiparametric Estimation of Fractional Cointegrating Subspaces

We consider a common components model for multivariate fractional cointegration, in which the s ¸ 1 components have different memory parameters. The cointegrating rank is allowed to exceed 1. The true cointegrating vectors can be decomposed into orthogonal fractional cointegrating subspaces suchthat vectors from distinct subspaces yield cointegrating errors with distinct memory parameters, denoted by dk, for k = 1; : : : ; s. We estimate each cointegrating subspace separately using appropriate sets ofeigenvectors of an averaged periodogram matrix of tapered, differenced observations. The averaging uses the first m Fourier frequencies, with m fixed. We will show that any vector in the k th estimatedcointegrating subspace is, with high probability, close to the k th true cointegrating subspace, in the sensethat the angle between the estimated cointegrating vector and the true cointegrating subspace convergesin probability to zero. This angle is Op(ni®k ), where n is the sample size and ®k is the shortest distance between the memory parameters corresponding to the given and adjacent subspaces. We show that the cointegrating residuals corresponding to an estimated cointegrating vector can be used to obtain a consistent and asymptotically normal estimate of the memory parameter for the given cointegrating subspace, using a univariate Gaussian semiparametric estimator with a bandwidth that tends to 1 more slowly than n. We also show how these memory parameter estimates can be used to test for fractional cointegration and to consistently identify the cointegrating subspaces.

Estimating Fractional Cointegration in the Presence of Polynomial Trends

We propose and derive the asymptotic distribution of a tapered narrow-band least squares estimator (NBLSE) of the cointegration parameter β in the framework of fractional cointegration. This tapered estimator is invariant to deterministic polynomial trends. In particular, we allow for arbitrary linear time trends that often occur in practice. Our simulations show that, in the case of no deterministic trends, the estimator is superior to ordinary least squares (OLS) and the nontapered NBLSE proposed by P.M. Robinson when the levels have a unit root and the cointegrating relationship between the series is weak. In terms of rate of convergence, our estimator converges faster under certain circumstances, and never slower, than either OLS or the nontapered NBLSE. In a data analysis of interest rates, we find stronger evidence of cointegration if the tapered NBLSE is used for the cointegration parameter than if OLS is used.

The FEXP estimator for potentially non-stationary linear time series

We consider semiparametric fractional exponential (FEXP) estimators of the memory parameter d for a potentially non-stationary linear long-memory time series with additive polynomial trend. We use differencing to annihilate the polynomial trend, followed by tapering to handle the potential non-invertibility of the differenced series. We propose a method of pooling the tapered periodogram which leads to more efficient estimators of d than existing pooled, tapered estimators. We establish asymptotic normality of the tapered FEXP estimator in the Gaussian case with or without pooling. We establish asymptotic normality of the estimator in the linear case if pooling is used. Finally, we consider minimax rate-optimality and feasible nearly rate-optimal estimators in the Gaussian case.

Linear Trend with Fractionally Integrated Errors

We consider the estimation of linear trend for a time series in the presence of additive long-memory noise with memory parameter d∈[0, 1.5). Although no parametric model is assumed for the noise, our assumptions include as special cases the random walk with drift as well as linear trend with stationary invertible autoregressive moving-average errors. Moreover, our assumptions include a wide variety of trend-stationary and difference-stationary situations. We consider three different trend estimators: the ordinary least squares estimator based on the original series, the sample mean of the first differences and a class of weighted (tapered) means of the first differences. We present expressions for the asymptotic variances of these estimators in the form of one-dimensional integrals. We also establish the asymptotic normality of the tapered means for d∈[0, 1.5) −{0.5} and of the ordinary least squares estimator for d∈ (0.5, 1.5). We point out connections with existing theory and present applications of the methodology.

The Averaged Periodogram Estimator for a Power Law in Coherency

We prove the consistency of the averaged periodogram estimator (APE) in two new cases. First, we prove that the APE is consistent for negative memory parameters, after suitable tapering. Second, we prove that the APE is consistent for a power law in the cross-spectrum and therefore for a power law in the coherency, provided that sufficiently many frequencies are used in estimation. Simulation evidence suggests that the lower bound on the number of frequencies is a necessary condition for consistency. For a Taylor series approximation to the estimator of the power law in the cross-spectrum, we consider the rate of convergence, and obtain a central limit theorem under suitable regularity conditions.