Abstract
We consider estimation for general power GARCH models under stable-Paretian innovations. Exploiting the simple structure of the conditional characteristic function of the observations driven by these models we propose minimum distance estimation based on the empirical characteristic function of corresponding residuals. Consistency of the estimators is proved, and the asymptotic distribution of the estimator is studied. Efficiency issues are explored and finite-sample results are presented as well as applications of the proposed procedures to real data from the financial markets. A multivariate extension is also considered.
Similar content being viewed by others
Notes
Because there exists no explicit form for this vector, it has been evaluated on the basis of a simulation of length 50, 000.
References
Adler RJ, Feldman RE, Taqqu MS (eds) (1998) A practical guide to heavy tails. Statistical techniques and applications. Birkhäuser, Boston
Andrews B, Calder M, Davis RA (2009) Maximum likelihood estimation for \(\alpha \)-stable autoregressive processes. Ann Stat 37:1946–1982
Billingsley P (1968) Convergence of probability measures. Wiley, New York
Billingsley P (1995) Probability and measure. Wiley, New York
Bonato M (2012) Modeling fat tails in stock returns: a multivariate stable-GARCH approach. Comput Stat 27:499–521
Bougerol P, Picard N (1992) Stationarity of GARCH processes and of some nonnegative time series. J Econom 52:115–127
Brockwell PJ, Davis RA (1992) Estimating the noise parameters from observations of a linear process with stable innovations. J Stat Plan Inference 33:175–186
Bryant JL, Paulson AS (1979) Some comments on characteristic function based estimators. Sankhyā 41:109–116
Davis RA (1996) Gauss–Newton and M-estimation for ARMA processes with infinite variance. Stoch Process Appl 63:75–95
Davis RA, Knight K, Liu J (1992) M-estimation for autoregression with infinite variance. Stoch Process Appl 40:145–180
Ding Z, Granger CW, Engle RF (1993) A long memory property of stock market returns and a new model. J Empir Finance 1:83–106
DuMouchel WH (1983) Estimating the stable index \(\alpha \) in order to measure tail thickness: a critique. Ann Stat 11:1019–1031
Fama E (1965) The behavior of stock market prices. J Bus 38:34–105
Feuerverger A (1990) An efficiency result for the empirical characteristic function in stationery time-series models. Can J Stat 18:155–161
Feuerverger A, McDunnough P (1981a) On the efficiency of empirical characteristic function procedures. J R Stat Soc Ser B 38:20–27
Feuerverger A, McDunnough P (1981b) On some Fourier methods for inference. J Am Stat Assoc 76:379–387
Feuerverger A, McDunnough P (1981c) On efficient inference procedures in symmetric stable laws and processes. In: Csörgő M et al (eds) Statistics and related topics, proceedings of the internal symposium on statistics and related topics. North-Holland, pp 109–122
Francq C, Zakoïan JM (2010) GARCH models: structure, statistical inference and financial applications. Wiley, New York
Ghosh S, Beran J (2006) On estimating the cumulant generating function of linear processes. Ann Inst Stat Math 58:53–71
Hamadeh T, Zakoïan JM (2011) Asymptotic properties of LS and QML estimators for a class of nonlinear GARCH processes. J Stat Plan Inference 141:488–507
Hannan EJ, Kanter M (1977) Autoregressive processes with infinite variance. J Appl Probab 14:411–415
Hansen LP (1982) Large sample properties of generalized method of moments estimators. Econometrica 50:1029–1054
Heathcote CR (1977) The integrated squared error estimation of parameters. Biometrika 64:255–264
Knight JL, Satchell SE, Yu J (2002) Estimation of the stochastic volatility model by the empirical characteristic function method. Aust NZ J Stat 44:319–335
Koutrouvelis IA, Meintanis SG (1999) Testing for stability based on the empirical characteristic function with applications to financial data. J Stat Comput Simul 64:275–300
Liu SM, Brorsen BW (1995a) Maximum likelihood estimation of a GARCH-stable model. J Appl Econom 10:273–285
Liu SM, Brorsen BW (1995b) GARCH-stable as a model for future price movements. Rev Quant Financ Acc 5:155–167
Mandelbrot B (1963) The variation of certain speculative prices. J Bus 36:394–419
McCulloch JH (1996) Financial applications of stable distributions. In: Maddala, Rao (eds) Handbook of statistics, vol 14. North Holland, Amsterdam, pp 393–425
Mittnik S, Rachev ST (1993) Modelling asset returns with alternative stable models. Econom Rev 12:261–330
Mittnik S, Rachev ST, Doganoglu T, Chenyao D (1999) Maximum likelihood estimation of stable Paretian models. Math Comput Model 29:275–293
Mittnik S, Paolella MS, Rachev ST (2002) Stationarity of stable power–GARCH processes. J Econom 106:97–107
Newey WK, McFadden D (1994) Large sample estimation and hypothesis testing. In: Engle, McFadden (eds) Handbook of econometrics, vol IV. Elsevier, Amsterdam, pp 2111–2245
Nolan JP (1997) Numerical calculation of stable densitries and distribution functions. Commun Stat Stoch Models 13:759–774
Nolan JP (2012) Stable distributions—models for heavy tailed data. Birkhauser, Boston
Nolan JP (2013) Multivariate empirically contoured stable distributions: theory and estimation. Comput Stat 28:2067–2089
Paolella MS (2001) Testing the stable Paretian assumption. Math Comput Model 34:1095–1112
Rachev S (ed) (2003) Handbook of heavy tailed distributions in finance. Elsevier/North-Holland, Amsterdam
Rachev S, Mittnik S (2000) Stable Paretian models in finance. Wiley, New York
Rotnitzky A, Cox DR, Bottai M, Robins J (2000) Likelihood based inference with a singular information matrix. Bernoulli 6:243–284
Samorodnitsky G, Taqqu MS (1994) Stable non-Gaussian random processes: stochastic models with infinite variance. Chapman and Hall, New York
Székely G, Rizzo M (2009) Brownian distance covariance. Ann Appl Stat 3:1236–1265
Tauchen GE (1985) Diagnostic testing and evaluation of maximum likelihood models. J Econom 30:415–443
Taufer E, Leonenko N (2009) Characteristic function estimation of non-Gaussian Ornstein–Uhlenbeck processes. J Stat Plan Inference 139:3050–3063
Thornton JC, Paulson AS (1977) Asymptotic distribution of characteristic function-based estimators for the stable laws. Sankhyā 39:341–354
Tsionas E (2002) Likelihood-based comparison of stable Paretian and competing models: evidence from daily exchange rates. J Stat Comput Simul 72:341–353
Yu J (2004) Empirical characteristic function estimation and its applications. Econom Rev 23:93–123
Acknowledgments
The authors wish to thank the anonymous reviewers for the constructive suggestions and comments which led to improvement of the original version of the article. A part of this research was carried out at the Department of Economics and was supported by Grant Number 11699 of the Special Account for Research Grants (ELKE) of the National and Kapodistrian University of Athens. Special thanks for the hospitality and financial support also go to the University of Lille 3. This work was also partially supported by the Agence Nationale de la Recherche (ANR) and the Economic and Social Research Council (ESRC) through the ORA Program (Project PRAM ANR-10-ORAR-008-01).
Author information
Authors and Affiliations
Corresponding author
Additional information
Simos G. Meintanis: On sabbatical leave from the University of Athens.
Appendices
Appendix 1: Technical lemmas
Let K and \(\varrho \) be generic constants, whose values will be modified along the proofs, such that \(K>0\) and \(0<\varrho <1\). Let also
where \(\varphi _T(u,\varvec{\theta })=\varphi _T\left\{ u; \varepsilon _1(\varvec{\theta }),\ldots ,\varepsilon _T(\varvec{\theta })\right\} \) is the ECF in (2.3) computed from the standardized innovations \(\varepsilon _t(\varvec{\theta })=y_t/c_t(\varvec{\theta })\). Note that \({\Delta }_T(\varvec{\vartheta })\) is well defined under A3(0) because \(\sup _u\left| \varphi _T(u,\varvec{\theta })-\varphi (u,\varvec{\lambda })\right| ^2<\infty \).
The following lemma shows that the choice of the unknown initial values is asymptotically unimportant for the objective function.
Lemma 8.1
Under the assumptions of Theorem 3.1, we have
Proof
Using the elementary relation \(|\cos x- \cos y|\le |x-y|\), we have
By the arguments used to show (5.3) in HZ, it is easily shown that
The mean-value theorem and the fact that \(\inf _{\varvec{\theta }\in \Theta }\min \left\{ c_t(\varvec{\theta }),\widetilde{c}_t(\varvec{\theta })\right\} \ge \underline{\mu }^{1/\rho }>0\), then imply that for \(c_t^{**}\) between \(c_t^\rho (\varvec{\theta })\) and \(\widetilde{c}_t^\rho (\varvec{\theta })\),
Noting that \(\text{ Re }\) can be replaced by \(\text{ Im }\) in (8.1), we thus have
The strict stationarity in A2 entails that \(E|y_t|^{2s}<\infty \) and \(E\sup _{\varvec{\theta }\in \Theta }|c_t(\varvec{\theta })|^{2s/\rho }<\infty \) for some small \(s>0\) (see Proposition A.1 in HZ). By the same arguments, we also have \(E\sup _{\varvec{\theta }\in \Theta }|\tilde{c}_t(\varvec{\theta })|^{2s/\rho }<\infty \). By the Cauchy-Schwarz inequality, the supremum over \(\Theta \) of the sum of the right-hand side of the inequality (8.2) admits a moment of order s. Therefore this sum is almost surely finite, uniformly in \(\Theta \). It follows that
We then conclude by using A3(1). \(\square \)
Let \(\varepsilon \) and \(\sigma \) be two independent random variables, \(\sigma >0\) and \(\varepsilon \) SP distributed with parameter \(\varvec{\lambda }_0\). The following lemma shows that if \(\sigma \) admits moments of order 10, then \(\sigma \varepsilon \) cannot follow a SP distribution with parameter \(\varvec{\lambda }\ne \varvec{\lambda }_0\). Note that the result is not true without the moment assumption (see Remark 8.3 below).
Lemma 8.2
Let \(\varphi _{\varvec{\lambda }}(u)= \exp \left\{ - |u|^\alpha \{1-{\mathrm {i}}\beta {{\mathrm{sgn}}}(u) \tan (\pi \alpha /2)\}\right\} \) where \(\varvec{\lambda }=(\alpha ,\beta )\in (1,2)\times [-1,1]\). If for some probability measure \(\nu \) on \([0,+\infty )\) admitting a moment of order 10
then \(\varvec{\lambda }=\varvec{\lambda }_0\) and \(\nu \) is the Dirac measure at the point 1.
Proof
First order Taylor expansions lead to the following inequalities (see Equation (26.4) in Billingsley 1995)
It follows that
where \(|R_{\varvec{\lambda }}(u)|\le K (|u|^{2\alpha }+|u|^{3\alpha })\) for some constant K. Moreover, as \(u\rightarrow 0\) we have
Identifying the right-hand sides of (8.3) and (8.4) as \(u\rightarrow 0\), we obtain \(\alpha =\alpha _0\) and
The real and imaginary parts of both sides being equal, it follows that
from which we deduce that \(\beta =\beta _0\). This is not sufficient to conclude concerning the measure \(\nu \). Doing Taylor expansions of higher orders, we have
where \(|R^*_{\varvec{\lambda }}(u)|\le K (|u|^{3\alpha }+|u|^{4\alpha }+|u|^{5\alpha })\) for some constant K. Note that \(\int R^*_{\varvec{\lambda }}(xu)\nu (dx)=o(|u|^{2\alpha })\) as \(u\rightarrow 0\) because \(\int |x|^{5\alpha }\nu (dx)<\infty .\) Identifying the approximations of \(\varphi _{\varvec{\lambda }_0}(u)\) and \(\int _0^{\infty }\varphi _{\varvec{\lambda }_0}(xu)\nu (dx)\) as \(u\rightarrow 0\), we obtain \(\int _{0}^{+\infty }|x|^{2\alpha _0}\nu (dx)=\int _{0}^{+\infty }|x|^{\alpha _0}\nu (dx)=1\). It follows that \(\nu \) is the Dirac measure at 1, which completes the proof. \(\square \)
Remark 8.3
The need of moment assumptions on the probability measure \(\nu \) is already evident from representation (5.2) which suggests that a symmetric SP random variable can be obtained as a scale mixture of normal distributions, with mixing distribution a SP distribution concentrated on \([0,\infty )\). A more general result involving non-normal mixtures of SP distributions is proved by Samorodnitsky and Taqqu (1994, §1.3). By way of example we consider the random variable \(W=X^{1/2} Z\), where Z is standard normal with CF \(\varphi _{\varvec{\lambda }}(u)=\exp \{-(1/2) u^2\}\) and X follows the Lévy distribution, which is a totally skewed to the right SP distribution with tail index \(\alpha =1/2\) and density \(\nu (dx)=1/(\sqrt{2\pi })x^{-3/2}\exp \{-1/(2x)\}dx, \ x>0\). Denote by \(\varphi _W(u)\) the CF of W. Then it readily follows that
which shows that W follows the Cauchy distribution, and consequently that this distribution has a stochastic representation as a mixture of normal distributions with variance following the Lévy distribution.
The following lemma is similar to Lemma 1 in Tauchen (1985) and Lemma 2.4 in Newey and McFadden (1994), except that the assumption of iid observations is relaxed.
Lemma 8.4
Let \((z_t)\) be a stationary and ergodic process. Assume that \(\Theta \) is compact, that \(\theta \mapsto a(z,\theta )\) is continuous on \(\Theta \) for all \(z\in \Omega _1\) such that \(P(z_1\in \Omega _1)=1\), and that there exists d(z) such that \(\Vert a(z,\theta )\Vert \le d(z)\) for all \(\theta \in \Theta \) and \(Ed(z_1)<\infty .\) Then \(\theta \mapsto E a(z_1,\theta )\) is continuous and
Proof
Let \(V_m(\theta )\) be the open ball of center \(\theta \) and radius 1 / m. The dominated convergence theorem entails that for all \(\theta _k\in \Theta \) and all \(\epsilon >0\) there exists m such that the neighborhood \(V(\theta _k)=V_m(\theta _k)\) satisfies
By a compactness argument, there exist \(\theta _1,\ldots , \theta _K\) such that \(\cup _{k=1}^KV(\theta _k) \subseteq \Theta \) where \(V(\theta _k)\) satisfies (8.5). Now note that
The ergodic theorem and (8.5) entail that, as \(T\rightarrow \infty \), the almost sure limit of the first term of the right-hand side of the inequality is bounded by \(\epsilon \). The ergodic theorem also shows that the limit of the second term is zero. By (8.5), the last term is bounded by \(\epsilon \). Since \(\epsilon \) is arbitrarily small, the conclusion follows. \(\square \)
We now show that the choice of the unknown initial values is asymptotically unimportant for the derivative of the objective function.
Lemma 8.5
Under the assumptions of Theorem 3.3, we have
Proof
In view of (9.4) and (9.5), we have
By (3.2), (9.6) and \(E\left| \varepsilon _t\right| <\infty \), the random variable of the right hand side of the last inequality is uniformly integrable in a neighborhood of \(\varvec{\vartheta }_0\). By the ergodic theorem, it follows that, when \(\varvec{\vartheta }\) is sufficiently close to \(\varvec{\vartheta }_0\), the right hand side is a.s. bounded by a constant or by u times a constant. Note also that \(\left| T^{-1}\sum _{t=1}^T\overline{g_t(u,\varvec{\vartheta })}\right| \le 2\). The dominated convergence theorem and A3(0)–A3(1) thus show that one can take the derivative under the integral symbol to obtain
at least when \(\varvec{\vartheta }\) is sufficiently close to \(\varvec{\vartheta }_0\). A similar expression holds for \({\partial \widetilde{\Delta }_T(\varvec{\vartheta })}/{\partial \varvec{\vartheta }}\). It follows that
where, with the obvious notation \(\widetilde{g}_t(u,\varvec{\vartheta })=e^{\mathrm {i}u\widetilde{\varepsilon }_t}-\varphi (u,\varvec{\lambda })\),
By the argument used to show that the series in (8.2) is bounded, we obtain
for some neighborhood \(V(\varvec{\vartheta }_0)\). By already used arguments, we also have
It follows that the integrand in \(a_T\) is almost surely bounded by \(KT^{-1/2}(|u|+u^2)W(u)\), when \(V(\varvec{\vartheta }_0)\) is sufficiently small. By the dominated convergence theorem and \(\mathbf{A3}(1)\), \(\mathbf{A3}(2)\), almost surely \(a_T\rightarrow 0\) as \(T\rightarrow \infty .\) Similar arguments show that \(b_T\rightarrow 0\), and the conclusion follows. \(\square \)
Lemma 8.6
Let \(\varepsilon \) be a random variable with the SP distribution of parameter \(\varvec{\lambda }=(\alpha ,\beta )\in (1,2)\times [-1,1]\). For all \(\nu \in (-3,\alpha -1)\), there exists a constant K such that for all \(c>0\)
Proof
The density \(f_{\varvec{\lambda }}(x)\) of \(\varepsilon \) is bounded and satisfies
for some constant \(K=K_{\varvec{\lambda }}\) (see e.g. Theorem 1.12 in Nolan 2012). To show the existence of \(E|\varepsilon |^{2+\nu } e^{{\mathrm {i}}c\varepsilon }\), it is thus sufficient to show the existence of
An integration by parts shows that the first integral is equal to
Similarly, is can be seen that the second integral is also bounded by K / c. The conclusion follows. \(\square \)
Lemma 8.7
Under the assumptions of Theorem 3.3, for any sequence \(\varvec{\vartheta }_T\) tending almost surely to \(\varvec{\vartheta }_0\) as \(T\rightarrow \infty \), we have
where \(\Vert \varvec{R}_T(u)\Vert \le (1+|u|+u^2)r_T\) and \(r_T\) tends almost surely to 0 as \(T\rightarrow \infty \).
Proof
First note that, similarly to (9.7), we have
Note that \(\partial ^2 g_s(u,\varvec{\vartheta })/\partial \varvec{\theta }\partial \varvec{\lambda }'=0\), that \(\partial ^2 g_s(u,\varvec{\vartheta })/\partial \varvec{\lambda }\partial \varvec{\lambda }'\) is a non random bounded function of u uniformly in \(\varvec{\vartheta }\), and
Conditioning by the past, using Lemma 8.6 with \(c=|u| c_t(\varvec{\theta }_0)/c_t(\varvec{\theta })\) and \(\nu =0,-1,-2\), and using (3.2), (9.6) and (8.7), it can be shown that there exist a neighborhood \(V(\varvec{\vartheta }_0)\) of \(\varvec{\vartheta }_0\) and a constant K independent of u such that
Using this result, Taylor expansions for the real and imaginary parts, and the ergodic theorem, we obtain
Now, in view of (9.4), we have
Using Lemma 8.6 and (9.6), we then obtain
Because the derivative of \(g_s(u,\varvec{\vartheta })\) with respect to \(\varvec{\lambda }\) is non-random, we also have
We thus have shown that, a.s.
The conclusion follows. \(\square \)
Lemma 8.8
Under the assumptions of Theorem 3.3, for any sequence \(\varvec{\vartheta }_T\) tending almost surely to \(\varvec{\vartheta }_0\) as \(T\rightarrow \infty \), we have
where
and \(r^{(i)}_T\) tends almost surely to 0 as \(T\rightarrow \infty \) for \(i=1,2\).
Proof
A second-order Taylor expansion yields
for some \(\varvec{\vartheta }^*_T\) between \(\varvec{\vartheta }_T\) and \(\varvec{\vartheta }_0\). A similar expansion holds when \(\text{ Re }(\cdot )\) is replaced by \(\text{ Im }(\cdot )\). The result will thus follow from Lemma 8.7, if we show that
satisfies the condition of the lemma. Given the arguments of the proof of Lemma 8.7, it will be sufficient to show that
In view of (8.8), the left-hand side of the inequality can be bounded by
and by other terms which can be handled similarly. Using the Hölder inequality, the ergodic theorem, (3.2), (9.6), and the consistency of \(\varvec{\theta }^*_T\), it can be shown that for any \(r>1\) and any sufficiently small neighborhood \(V(\varvec{\theta }_0)\) of \(\varvec{\theta }_0\), the average (8.10) is almost surely bounded by \(Ku^{2}\left\{ E\sup _{\varvec{\theta }\in V(\varvec{\theta }_0)}\left| \frac{c_1(\varvec{\theta }_0)-c_1(\varvec{\theta })}{c_1(\varvec{\theta })}\right| ^{r}\right\} ^{1/r}\). The conclusion follows. \(\square \)
Appendix 2: Proof of the main results
Proof of Theorem 3.1
We begin to show that the limiting criterion is minimal only at the true value, that is
First note that Lemma 8.1 shows that \(\widetilde{\Delta }_T\) can be replaced by \(\Delta _T\) in (9.1). The ergodic theorem shows that, almost surely,
The second convergence of (9.1) is thus a direct consequence of the dominated convergence theorem and A3(0). Using \(W(\cdot )>0\) and the continuity of the characteristic functions, the same arguments show that \(\lim _{T\rightarrow \infty } \Delta _T(\varvec{\vartheta })=0\) iff \(\varphi (u,\varvec{\lambda })\) is the characteristic function of
that is iff
Lemma 8.2 of “Appendix 1” shows that, under A5, this is only possible if
which is equivalent to \(\varvec{\vartheta }=\varvec{\vartheta }_0\) by A4 (see the proof of Theorem 3.1 in HZ). The proof of (9.1) is complete.
We now show that the inequality in (9.1) holds locally uniformly, i.e. that there exists a neighborhood \(V(\varvec{\vartheta }^*)\) of \(\varvec{\vartheta }^*=(\varvec{\theta }^{* '},\varvec{\lambda }^{* '})'\) such that
In view of Lemma 8.1, we can replace \(\widetilde{\Delta }_T\) by \(\Delta _T\), which makes possible to apply the ergodic theorem. Let \(E\exp \{iu\varepsilon _1(\varvec{\theta })\}\) be the almost sure limit of \(\varphi _T(u,\varvec{\theta })\). Lemma 8.4 shows that the convergence is actually uniform:
Note that
with
Using (9.3), A3(0) and the bound \(\left| \varphi _T(u,\varvec{\theta })-E\exp \{\mathrm {i}u\varepsilon _1(\varvec{\theta })\}\right| \le 2\), we show that \(\sup _{\varvec{\theta }\in \Theta }a_T(\varvec{\theta })\rightarrow 0\; a.s.\) By the Cauchy-Schwarz inequality, we similarly obtain \(\sup _{\varvec{\vartheta }\in \Xi }d_T(\varvec{\vartheta })\rightarrow 0\; a.s.\) For any positive integer k, let \(V_k(\varvec{\vartheta }^*)\) be the open ball of center \(\varvec{\vartheta }^*\) and radius 1 / k. By Beppo Levi, as \(k\rightarrow \infty \)
where the last inequality comes from (9.1). It follows that there exists a neighborhood \(V(\varvec{\vartheta }^*)\) such that \(\inf _{\varvec{\vartheta }\in V(\varvec{\vartheta }^*)}b(\varvec{\vartheta })>0\). We then obtain (9.2) by noting that
The conclusion follows from (9.1), (9.2) and a standard compactness argument. \(\square \)
Proof of Lemma 3.2
We have
and
with
Since \(E\left| \varepsilon _t\right| <\infty \), and \(\varepsilon _t\) and \(c_t(\varvec{\theta })\) are independent, we have
In view of (5.20) in HZ and its extension Page 506, we have
Using also (3.2), the Hölder inequality then entails that for some neighborhood \(V(\varvec{\vartheta }_0)\) of \(\varvec{\vartheta }_0\), we have
where the norm of a complex vector denotes the sum of the norms of its real and imaginary parts. By Lebesgue’s dominated convergence theorem, we thus have
where
We then have
with
In view of (9.5) and (9.8), each component of \(\varvec{g}(u)\) is a bounded function of u (since \(\varphi (u,\varvec{\lambda }_0)\) tends to zero at an exponential rate as \(|u|\rightarrow \infty \)). The existence of \(\varvec{G}\) thus follows from A3(0). Since \(|g_t(u,\varvec{\vartheta }_0)|\le 2\), the existence of \(\varvec{V}\) also follows.
Let us show that \(\varvec{G}\) is singular. This is equivalent to proving that there exists \(\varvec{a}\ne \varvec{0}\) such that \(\varvec{a}'\varvec{g}(u)=0\) for all u (see Theorem 2 in Bryant and Paulson 1979). Letting \(\varvec{a}=(\varvec{a}_1',a_2,a_3)'\) with \(\varvec{a}_1\in {\mathbb {R}}^{p+q+2}\), and using (9.5) and (9.8), we have
Since \(|\varphi (u,\varvec{\lambda }_0)|>0\) and since the functions \(\tau _1(u)\), \(\tau _2(u)\) and \(\tau _3(u)\) are linearly independent, \(\varvec{a}'\varvec{g}(u)=0\) for all u if and only if
This can obviously be achieved by choosing \(\varvec{a}_1\ne \varvec{0}\). Note that the rank of \(\varvec{G}\) is equal to 3. The singularity of \(\varvec{V}\) is shown similarly. \(\square \)
Proof of Theorem 3.3
Assumption A6 and the consistency of \(\widehat{\varvec{\vartheta }}_T\) entail that \({\partial \widetilde{\Delta }_T(\widehat{\varvec{\vartheta }}_T)}/{\partial \varvec{\vartheta }}=0\), at least for T large enough. In view of Lemma 8.5, (8.6) and Lemma 8.7, we thus have, a.s.
A Taylor expansion shows that
for some \(\varvec{\vartheta }^*_T\) between \(\widehat{\varvec{\vartheta }}_T\) and \(\varvec{\vartheta }_0\). A similar expansion holds for the imaginary part. By Lemma 8.7, we thus have
Now, note that, by the central limit theorem
and by A3(0) and A3(4)
Theorem 4.1 in Billingsley (1968) thus entails
In view of (9.10) and (9.11), (9.9) implies
where
Note that
which tends almost surely to 0 as \(T\rightarrow \infty \), under A3(0) and A3(4). The conclusion follows from the central limit theorem. \(\square \)
Proof of Theorem 3.4
In view of (9.5) and (9.8), we have
Since the functions \(\tau _i(u)\), \(i=1,2,3\), are linearly independent, the matrix
is invertible (see Theorem 2 in Bryant and Paulson 1979), and we have
Letting \(\varvec{A}_T'=\varvec{B}^{-1}(\varvec{A}'\varvec{A})^{-1}\varvec{A}'\varvec{G}_T\), we thus have \(\varvec{A}_T\rightarrow \varvec{A}\) a.s. as \(T\rightarrow \infty \). By Theorem 3.3 we then obtain \(\sqrt{T}\varvec{A}_T'(\widehat{\varvec{\vartheta }}_T-\varvec{\vartheta }_0)\mathop {\rightarrow }\limits ^{d}\mathcal{N}\left( \varvec{0},\varvec{\Sigma }\right) \) with
Since
a computation similar to that of (17) in Thornton and Paulson (1977) gives \(\varvec{V}=\varvec{A}\varvec{U}\varvec{A}'\) with
The conclusion follows. \(\square \)
Proof of Theorem 3.5
From (9.9), Lemmas 8.7–8.8 and already employed arguments, we have
In view of (9.12), we thus have
Note that \({\partial ^2\overline{g_t(u,\varvec{\vartheta })}}/{\partial \varvec{\vartheta }\partial \varvec{\vartheta }'}\) is block-diagonal. Therefore \(\varvec{H}(u)\) and \(\varvec{R}^{(2)}_T(u) \) are also block-diagonal matrices. The block-diagonal terms of \(\varvec{H}(u)\) are
Those of \(\varvec{R}^{(2)}_T(u) \) are denoted by \(\varvec{R}^{(2)}_{\varvec{\theta }_0,T}(u) \) and \(\varvec{R}^{(2)}_{\varvec{\lambda }_0,T}(u)\). Since, under (3.3), Theorem 3.4 entails that \(\sqrt{T}(\widehat{\varvec{\lambda }}_T-\varvec{\lambda }_0)=O_P(1)\), we have
Noting also that \(\varvec{g}(u)=\varvec{A}\varvec{\tau }(u)\) where \(\varvec{A}\) has full rank, it follows from (9.17) that
Under A3(0) and A3(4), Lemma 8.8 entails
Similarly to (9.8), we obtain
with
and
In view of (9.20) and (9.21), we then have
where \(r_T=o_P(1)+o_P\left( T^{1/2}\left\| \widehat{\varvec{\theta }}_T-\varvec{\theta }_0)\right\| ^2\right) \) as \(T\rightarrow \infty \). In view of (3.6) this entails that
for \(i=1\) and \(i=2\). Since \(\varvec{H}^{(1)}_{\varvec{\theta }_0}\) is positive definite (see (ii) in the proof of Theorem 2.2 in HZ), the proof is complete. \(\square \)
Rights and permissions
About this article
Cite this article
Francq, C., Meintanis, S.G. Fourier-type estimation of the power GARCH model with stable-Paretian innovations. Metrika 79, 389–424 (2016). https://doi.org/10.1007/s00184-015-0560-x
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00184-015-0560-x