The estimation of misspecified long memory models
Introduction
The spectral density at low frequencies determines the long-run behavior of stationary time series. Let the covariance stationary and invertible process , , have a spectral density function , , defined by In practice, a finite realization, , may be the outcome of integer differencing or integrating or deterministic detrending of a nonstationary or non-invertible series. With meaning that , we assume that is regularly-varying at zero frequency, that is where and, for positive argument , the function is slowly-varying (in Karamata’s sense), being positive and measurable on some neighborhood , with
Detailed discussions of slowly-varying functions, and their applications in probability theory, are contained in Seneta (1976) and Bingham et al. (1987). A basic property is that as can diverge, or converge to zero, or converge to a positive constant, or oscillate, and for any ,
Therefore in (2) the power law dominates the slowly varying factor so that, for any , as still diverges for , and still for , while when diverges when as and when as .
The simplest example of such an is Others include (see Bingham et al. (1987, p. 16)) where and , , as well as powers and rational functions of the , (e.g. ), and
Let denote the -field of events generated by , , and define for . Then if as , is said to be -mixing. Suppose for the purposes of this paragraph that is Gaussian, in which case the coefficient of complete regularity decays at the same rate as , see Ibragimov and Rozanov (1978, pp. 111, 113). Thus from Ibragimov and Rozanov (1978, pp. 178) satisfying (2) cannot be -mixing when (because not every positive power of is integrable). The usual examples of Gaussian -mixing processes have bounded spectral density, e.g. a stationary and invertible autoregressive moving average (ARMA), and thus satisfy (2) with and constant , (5). However -mixing does not rule out all unbounded . From Ibragimov and Rozanov (1978, pp. 179, 180), for some implies is -mixing. The spectral density in (9) satisfies which corresponds to combining (6) for with (2) for . Incidentally under (9) decays very slowly, like (and thus does not satisfy conditions for central limit theory for statistics such as the sample mean of , ). From Ibragimov and Rozanov (1978, p. 180) a process with spectral density the reciprocal of the right side of (9) (which converges like as ) is also -mixing.
Under additional conditions to (2) (see Yong, 1974) the autovariance sequence satisfies The probability literature covers the asymptotic behavior of various simple statistics under (11), in particular linear and quadratic forms (see e.g. Taqqu (1975), Dobrushin and Major (1979), Fox and Taqqu, 1985, Fox and Taqqu, 1987). However, the frequency domain form (2) perhaps provides greater intuitive appeal. Early empirical support for the notion of a divergent spectral density at zero frequency was noted by Granger (1966). He reported nonparametric spectral density estimates for a number of economic time series, and while these are inevitably finite at zero frequency, they are strongly peaked there, and his Fig. 1 is suggestive of a spectral singularity at zero frequency. Of course such an outcome could also be consistent with nonstationarity (such as a unit root), and he did not present formulae such as (2), but clearly (2) with and any , or even with and diverging , is consistent with his “typical spectral shape”. The leading methods of semiparametric estimation of the memory parameter have also been frequency-domain. However, they have mainly focused on the simple power law form, with (5) assumed in (2), that is The leading fractional parametric models (which specify parametrically for all ), namely (Adenstedt (1974)) and its extension to fractionally-integrated ARMA (FARIMA) spectra are covered by (12). In (12) the knife-edge case describes short memory, when a FARIMA reduces to an ARMA, while the cases and respectively describe long memory and antipersistence. However, methods of estimating such parametric models are inconsistent when is misspecified, in particular high-frequency misspecification produces asymptotic bias even in estimates of the low-frequency parameter . This drawback is overcome (at cost of slower convergence, and of requiring choice of a smoothing number) by semiparametric methods, based on (12), in particular log-periodogram and local Whittle estimates of and , see e.g. Geweke and Porter-Hudak (1983), Kuensch (1987), Robinson, 1995a, Robinson, 1995b, where the latter two references established that both estimates are asymptotically normal for all , and with an asymptotic variance that is constant with respect to . Thus, standard large-sample inference using these estimates is very simple to implement. Extensions to estimates based on nonstationary processes have been developed by Velasco, 1999a, Velasco, 1999b and subsequent authors.
In principle, one could specify a particular in (2) up to an unknown scale factor as in (6), (7), (8), for example, and accordingly modify the estimates, and we would expect to achieve good statistical properties if is correctly chosen. One could also imagine specifying up to finitely many unknown parameters, e.g. for unknown , and extend the semiparametric methods to estimate , and the additional parameter vector. However in either case the prospect of correct specification of seems far-fetched, and of greater practical interest is the robustness of existing estimates to unknown, nonparametric, .
Robinson (1994a) investigated asymptotic properties of the averaged periodogram statistic, and its functionals of interest, including a semiparametric estimate of , under (2) with unknown . Define the discrete Fourier transform
and the periodogram The averaged periodogram is defined as where here denotes the integer part and . For a user-chosen integer satisfying Robinson (1994a) showed that where For this purpose (2) was assumed but (like a good deal of the long memory literature) under the restriction (though there seems no reason why a similar result should not hold also for ), as well as regularity conditions. Further, Robinson (1994a) proposed the following averaged periodogram estimate of : where is chosen in the interval . He showed that under the same conditions as imposed for (17), Under somewhat stronger conditions he obtained a rate of convergence in (20), , for some . The property (20), like (17), holds for any slowly varying , which is unknown to the practitioner. Intuitively both properties might be anticipated due to (3) and the ratio forms on the left hand side of (17) and in . Robinson (1994b) discussed mean squared error and optimal choice of in this setting. The present paper addresses the above issues with respect to the log-periodogram estimate, which, like but unlike the local Whittle estimate, is defined in closed form, so relatively easily yields information on rates of convergence. Soulier (2010) established a lower bound for the rate of convergence of estimates of in (2), and proved it to be optimal, illustrating his results with the log periodogram estimate. Giraitis et al. (1997) had considered similar issues with respect to (12), but Soulier (2010) found that the presence of an unanticipated can produce much slower rates, and that unlike under (12), the log periodogram estimate is no less efficient than the local Whittle estimate, cf. Robinson, 1995a, Robinson, 1995b), where the asymptotic distributional results derived in the latter references may only hold alongside bandwidth choices that yield unacceptable imprecision.
The following section considers the consistency of the log-periodogram estimate. Section 3 evaluates the order of magnitude of the bias in several slowly varying examples, with some discussion of mean squared error and bandwidth choice. Section 4 provides some concluding remarks.
Section snippets
Consistency of log periodogram estimate
We employ the version of the log-periodogram estimate proposed by Robinson (1995a) (which is slightly simpler than Geweke and Porter-Hudak’s (1983)). For as described in the previous section, define and introduce the additional notation for integer . The log-periodogram estimate we consider is Define also
We introduce two assumptions.
Assumption 1 As The unprimitive Assumption 1 can hold
Examples and rates
The paragraph following Assumption 2 argues that the assumption does not much strengthen the slow variation property of , but it is nevertheless desirable to check it in several cases, and this will desirably indicate rates of convergence. Throughout the derivations it is understood that is chosen arbitrarily large and .
- 1.
, , .
This is actually a case of (12), and was assumed in the central limit theorem for of Robinson (1995a), because some refinement of (12) is
Final comments
We have considered the consistency of the semiparametric log-periodogram regression memory estimate in the presence of an unanticipated slowly-varying factor in the spectral density, under a general condition on the function, and verified this condition and calculated convergence rates in several examples. As implied by the results of Soulier (2010), these convergence rates are mostly slow, to the extent that unless the bandwidth grows extremely slowly the bias will be too large to allow the
Acknowledgements
I thank two referees for comments which have led to presentational improvements, and Phillipe Soulier for alerting me to the reference Soulier (2010). This research was supported by a Cátedra de Excelencia at Universidad Carlos III de Madrid, Spanish Nacional de I+d+I Grant SEJ2007-62908/ECON, and ESRC Grant ES/J007242/1.
References (22)
Non-stationary log-periodogram regression
Journal of Econometrics
(1999)On large-sample estimation for the mean of a stationary random sequence
Annals of Statistics
(1974)- et al.
Regular Variation
(1987) - et al.
Non-central limit theorems for nonlinear functionals of Gaussian fields
Zeitschrift fur Wahrscheinlichkeitstheorie und verwandte Gebiete
(1979) - et al.
Non-central limit theorems for quadratic forms in random variables having long-range dependence
Annals of Probability
(1985) - et al.
Central limit theorems for quadratic forms in random variables having long-range dependence
Probability Theory and Related Fields
(1987) - et al.
The estimation and application of long memory time series models
Journal of Time Series Analysis
(1983) - et al.
Rate optimal semiparametric estimation of the memory parameter of the Gaussian time series with long-range dependence
Journal of Time Series Analysis
(1997) The typical spectral shape of an economic variable
Econometrica
(1966)- et al.
The mean squared error of Geweke and Porter-Hudak’s estimator of the memory parameter in a long-memory time series
Journal of Time Series Analysis
(1998)
Gaussian Random Processes
Cited by (3)
Issues in the estimation of mis-specified models of fractionally integrated processes
2020, Journal of EconometricsHarmonically Weighted Processes
2020, Journal of Time Series AnalysisTesting for Change-Points in Long-Range Dependent Time Series by Means of a Self-Normalized Wilcoxon Test
2016, Journal of Time Series Analysis