Abstract
This paper proposes a class of minimum distance estimators for the underlying parameters in a Markovian parametric multiplicative error time series model. This class of estimators is based on the integrals of the square of a certain marked residual process. The paper derives the asymptotic distributions of the proposed estimators. In a finite sample comparison, some members of the proposed class of estimators dominate a generalized method of moments estimator in terms of the finite sample bias at a variety of chosen error distributions while neither dominate each other in terms of the mean squared error at these error distributions. A real data example is considered to illustrate the proposed estimation procedures.
Similar content being viewed by others
References
Bauwens, L. and Giot, P. (2000). The logarithmic ACD model: an application to the bid-ask quote process of three NYSE stocks. Ann. Econ. Stat. 60, 117–149.
Bauwens, L. and Veredas, D. (2004). The stochastic conditional duration model: a latent factor model for the analysis of financial durations. J. Econometrics119, 381–412.
Brownlees, C.T., Cipollini, F. and Gallo, G.M. (2012). Multiplicative Error Models. Wiley, New Jersey, Bauwens, L., Hafner, C. and Laurent, S (eds.), p. 225–247.
Chang, N.M. (1990). Weak convergence of a self-consistent estimator of a survival function with doubly censored data. Ann. Statist. 18, 391–404.
Chen, Y. -T. and Hsieh, C. -S. (2010). Generalized moment tests for autoregressive conditional duration models. J. Financ. Econom. 8, 345–391.
Chou, R.Y. (2005). Forecasting financial volatilities with extreme values: The conditional autoregressive range (CARR) model. J. Money, Credit Banking37, 561–582.
Cipollini, F., Engle, R.F. and Gallo, G.M. (2013). Semiparametric vector MEM. J. Appl. Econ. 28, 1067–86.
Donoho, D.L. and Liu, R.C. (1988a). The automatic robustness of minimum distance functionals. Ann. Statist. 16, 552–586.
Donoho, D.L. and Liu, R.C. (1988b). Pathologies of some minimum distance estimators. Ann. Statist. 16, 587–608.
Drost, F.C. and Werker, B.J.M. (2004). Semiparametric duration models. J. Bus. Econ. Stat. 22, 40–50.
Engle, R. (2002). New frontiers for ARCH models. J. Appl. Econ. 17, 425–446.
Engle, R.F. and Gallo, G.M. (2006). A multiple indicators model for volatility using intra-daily data. J. Econometrics 131, 3–27.
Engle, R.F. and Russell, J.R. (1998). Autoregressive conditional duration: a new model for irregularly spaced transaction data. Econometrica. J. Econometric Society 66, 1127–1162.
Fernandes, M. and Grammig, J. (2006). A family of autoregressive conditional duration models. J. Econometrics 130, 1–23.
Grammig, J. and Maurer, K.-O. (2000). Non-monotonic hazard functions and the autoregressive conditional duration model. Econometrics Journal 3, 16–38.
Grammig, J. and Wellner, M. (2002). Modeling the interdependence of volatility and inter-transaction duration processes. J. Econometrics 106, 369–400.
Gao, J., Kim, N.H. and Saart, P.W. (2015). A misspecification test for multiplicative error models of non-negative time series processes. J. Econometrics 189, 346–359.
Guo, B. and Li, S. (2018). Diagnostic checking of markov multiplicative error models. Econ. Lett. 170, 139–142.
Hall, P. and Heyde, C.C. (1980). Martingale Limit Theory and Its Applications. Academic Press, New York.
Hautsch, N. (2012). Econometrics of Financial High-Frequency Data. Springer, Heidelberg.
Koul, H.L. (1985a). Minimum distance estimation in multiple linear regression. Sankhya, Ser. A. 47 Part 1, 57–74.
Koul, H.L. (1985b). Minimum distance estimation in linear regression with unknown errors. Statist. Prob. Lett. 3, 1–8.
Koul, H.L. (1986). Minimum distance estimation and goodness-of-fit tests in first-order autoregression. Ann. Statist. 14, 1194–1213.
Koul, H.L. (1996). Asymptotics of some estimators and sequential residual empiricals in non-linear time series. Ann. Statist. 24, 380–404.
Koul, H.L. (2002). Weighted Empirical Processes in Dynamic Nonlinear Models, 166, 2nd edn. Springer, USA,.
Koul, H.L. (2019). A Uniform Convergence Result for Weighted Residual Empirical Processes RM#719. Department of Statistics and Probability, MSU.
Koul, H.L. and Stute, W. (1999). Nonparametric model checks for time series. Ann. Statist. 27, 204–236.
Koul, H.L. and Perera, I. (2021). A minimum distance lack-of-fit test in a markovian multiplicative error model. J. Stat. Theory Pract. 15, 34.
Koul, H.L., Perera, I. and Silvapulle, M.J. (2012). Lack-of-fit testing of the conditional mean function in a class of Markov multiplicative error models. Econometric Theory 28, 1283–1312.
Ljung, G.M. and Box, G.E.P. (1978). On a measure of lack of fit in time series models. Biometrika 65, 297–303.
Lunde, A. (1999). A Generalized Gamma Autoregressive Conditional Duration Model. Discussion paper, Aalborg University.
Manganelli, S. (2005). Duration, volume and volatility impact of trades. J. Financ. Mark. 8, 377–399.
Meitz, M. and Teräsvirta, T. (2006). Evaluating models of autoregressive conditional duration. J. Bus. Econ. Stat. 24, 104–124.
Pacurar, M. (2008). Autoregressive conditional duration models in finance: a survey of the theoretical and empirical literature. J. of Economic Surveys 22, 711–751.
Perera, I., Hidalgo, J. and Silvapulle, M.J. (2016). A goodness-of-fit test for a class of autoregressive conditional duration models. Econ. Rev. 35, 1111–1141.
Perera, I. and Koul, H.L. (2017). Fitting a two phase threshold multiplicative error model. J. Econometrics 197, 348–367.
Perera, I. and Silvapulle, M.J. (2021). Bootstrap based probability forecasting in multiplicative error models. J. Econometrics 221, 1–24.
Stute, W. (1976). On a generalization of the Glivenko-Cantelli theorem. Z. Wahrscheinlichkeitstheorie und Verw. Gebiete 35, 167–175.
Stute, W. (1997). Nonparametric model checks for regression. Ann. Statist. 25, 613–641.
Acknowledgements
Authors are grateful to the referee and the editor for their useful comments and suggestions on the earlier draft of the paper.
Funding
N. Balakrishna received partial financial suppport from SERB of India under the MATRICS scheme MTR/2018/000195.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of Interest
The authors have no conflicts of interest to declare that are relevant to the content of this article.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix A: Main Proofs
Appendix A: Main Proofs
1.1 A.1 Proofs of Theorems 4.1 and 4.2
Proof 1 (Proof of Theorem 4.1).
The proof of (4.4) is given later. Arguing as in the proof of Theorem 5.4.1 of Koul (2002), one verifies that (4.1), (C.3) and (4.4) imply Eq. 4.5 and (4.6)(a).
To prove (4.6)(b), by the positive definiteness of \({\mathcal G}_{1}\) we clearly have \(\widetilde t_{n1}= {\mathcal G}_{1}^{-1}V_{n1}\). Moreover, Vn1 is a vector of the sums of martingale difference arrays satisfying (4.3). By the martingale central limit theorem (CLT), see Hall and Heyde (1980), \(V_{n1}\to _{D} N(0, \sigma ^{2}{\Sigma }_{1}) \) and \(\widetilde t_{n1}={\mathcal G}_{1}^{-1}V_{n1}\to _{D} N(0, \sigma ^{2} {\mathcal G}_{1}^{-1}{\Sigma }_{1} {\mathcal G}_{1}^{-1}). \) This fact together with the Slutsky Theorem, (4.5) and (4.6)(a) imply (4.6)(b).
Proof of (4.4).
Let θnt := θ + n− 1/2t, \(t\in {\mathbb R}^{q}\). Define, for \(z\in {\mathbb R}_{+}^{p}, t\in {\mathbb R}^{q}\),
Use the model assumption εi = Yi/ψ(Zi− 1, θ) to obtain
Hence,
We shall shortly prove the following facts. For every \(0<b<\infty \),
Then by the Cauchy-Schwarz inequality, \( \sup _{\|t\|\le b}\big |T_{n12}(t)\big |^{2}\le \sup _{\|t\|\le b} T_{n11}\) \((t) \sup _{\|t\|\le b} Q_{n1}(t)\) = op(1). Hence the claim Eq. 4.4.
Proof of (A.3)(a).
Rewrite
Let dit := ψ(Zi− 1, θnt) − ψ(Zi− 1, θ) and \( \delta _{it}:= d_{it} -n^{-1/2}t^{\prime } \dot \psi (Z_{i-1}, \theta ). \) Then the above identity is equivalent to
In the sequel, the range of the vectors z, 𝜗 in \(\inf _{z,\vartheta }\) is over \({\mathbb R}_{+}^{p} \times {\Theta }\), unless mentioned otherwise. By (C.1), \(\inf _{z, \vartheta }\psi (z,\vartheta )\ge C>0\) and for every \(b<\infty \), \(\sup _{1\le i\le n, \|t\|\le b}n^{1/2}\big |\delta _{it}\big |=o_{p}(1)\). Moreover, because E(ε0) = 1, we have \(n^{-1}{\sum }_{i=1}^n \varepsilon _{i}=O_{p}(1)\). Hence
Next, recall that the stationarity of the process \(Y_{i}, i\in {\mathbb Z}\) and \(E\|\dot \psi (Z_{0},\theta )\|^{2}\) \(<\infty \) imply that \(n^{-1/2} \max \limits _{1\le i\le n}\|\dot \psi (Z_{i-1},\theta )\| =o_{p}(1)\), \(E\big (n^{-1}{\sum }_{i=1}^{n} \|\varphi \) \((Z_{i-1}) \| \varepsilon _{i}\big )= E\|\varphi (Z_{0})\|<\infty \), and by the Markov inequality, \(n^{-1}{\sum }_{i=1}^{n} \|\varphi \) (Zi− 1)∥εi = Op(1). Hence, (C.1) and (C.2) imply that
Next, consider Sn(z). Observe that Sn(z) is a vector of weighted empirical processes with the summands of each component being stationary and ergodic and ESn(z) ≡ 0. Using a Glivenko-Cantelli Lemma type argument one obtains that \(\sup _{z\in {\mathbb R}_{+}^{p}}\|S_{n}(z)\|=o_{p}(1)\). For details see Stute (1976) and Koul (2019). Similarly, \(\sup _{z\in {\mathbb R}_{+}^{p}}\big \|{\Psi }_{n}(z)-{\Psi }(z)\big \|^{2}=o_{p}(1)\). Upon combining these two facts with (A.4), (A.5) and (A.6) we obtain that for every \(0<b<\infty \),
This fact combined with the definition of Tn11 and L being a d.f. readily yields (A.3)(a).
Next, to prove (A.3)(b), note that
This also completes the proof of Theorem 4.1. □
Proof 2 (Proof of Theorem 4.2).
The proof is similar to that of Theorem 4.1, with some differences. Let
Then akin to (A.2),
Now, \( \sup _{\|t\|\le b} \tilde T_{n21}(t)\le \sup _{z\in {\mathbb R}_{+}^{p}, \|t\|\le b}\big (W_{n}(z,t)+t^{\prime }{\Psi }(z)\big )^{2}=o_{p}(1), \) by (A.7). We shall shortly prove that for every \(0<b<\infty \),
Argue as for (A.3)(b) to conclude that \(E\big (\sup _{\|t\|\le b} Q_{n2}(t)\big )=O(1). \) Hence \(\sup _{\|t\|\le b} \tilde Q_{n2}(t)=O_{p}(1)\) and \( \sup _{\|t\|\le b}\big |\tilde T_{n22}(t)\big |^{2}\le \sup _{\|t\|\le b} \tilde T_{n21}(t) \sup _{\|t\|\le b}\) \(\tilde Q_{n2}(t) \) = op(1). These facts together yield the claim Eq. 4.7.
Next, to prove (A.8), note that the left hand side of (A.8) is bounded from the above by
By the Ergodic Theorem, \(\big \|{\int \limits } {\Psi } {\Psi }^{\prime }\big [dG_{n}-dG\big ]\big \|\to 0\), a.s. while by Lemmas A.1 and A.2 below, the first two terms tend to zero in probability. This completes the proof of (4.7). The proofs of the other two claims of this theorem are similar to those of (4.5) and (4.6)(a), (b) of Theorem 4.1. □
1.2 A.2 Proof of Lemma 4.1
Proof 3 (Proof of the claim about (C.3)).
Let h be a positive function on \({\mathbb R}_{+}^{p}\) with \(0<{\int \limits } h^{2} dL<\infty \) and let \({\mathbf {\infty }}\) (0) denote the vector of p infinities (zeros). Let \(\varphi (z):= {\int \limits }_{x\le z} h(x) dL(x)\). Note that φ is nondecreasing in each coordinate and \(\varphi ({\mathbf {\infty }})<\infty \). Moreover, \(\gamma (z):={\int \limits } I(z\le x) h(x) dL(x)={\int \limits } I(z\le x) d\varphi (x)\ge 0\), for all \(z\in {\mathbb R}_{+}^{p}\). Define
By the Cauchy-Schwarz inequality,
Because γ(z) ≥ 0 for all z, by (4.8), \(\forall e\in R^{q}, \|e\|=1, {\mathcal V}_{n1}(r e)\) is monotonic in r. Hence
Let
In view of (A.1),
Because \(\varphi ({\mathbf {\infty }})={\int \limits } h dL<\infty \), by (A.7),
Therefore,
Let \(\tau _{2}:={\int \limits } h^{2} dL\). Fix an 𝜖 > 0 and \(0<\eta <\infty \). By (A.9), there exists N𝜖, η such that
Moreover, by the Cauchy-Schwarz inequality, \( E{\mathcal V}_{n1}^{2}(0)=\sigma ^{2} E\gamma ^{2}(Z_{0}) \le \sigma ^{2}{\int \limits } h^{2} dL<\infty . \) Hence by the Markov inequality, for every 𝜖 > 0 there is a b𝜖 such that
Recall the elementary fact that for any real numbers a, b, ||a|−|b||≤|a ± b|. Also let \(K:= {\int \limits } \|{\Gamma }\| d \varphi \). Let b𝜖 be as in (A.10). Choose b ≥ (b𝜖 + (v2η)1/2)K− 1. Then, ∀n ≥ N𝜖, η,
thereby proving the claim about (C.3).
Proof of the claim about (C.4)
The proof of this claim is similar to that of the previous claim, but with some differences. In particular we need the following two preliminary lemmas.
Lemma A.1.
Under the above set up, Un converges weakly to a continuous Gaussian process \({\mathcal Z}(x), x\in {\mathbb R}_{+}^{p}\) in Skorokhod space \(D([0,\infty ]^{p})\) and uniform metric.
The proof of this lemma is similar to that of the main theorem in Stute (1976).
Lemma A.2.
Let \({\mathcal U}\) be a relatively compact subset of \(D[0,\infty ]^{p}\). Let μn, μ be a sequence of possibly random multivariate distribution functions on \([0,\infty )^{p}\) such that \(\sup _{x\in {\mathbb R}_{+}^{p}} \big |\mu _{n}(x)-\mu (x)\big |\to 0\), a.s. Then \( \sup _{y\in {\mathbb R}_{+}^{p}, \alpha \in {\mathcal U}}\Big | {\int \limits }_{x\le y} \alpha (x)\) [dμn(x) − dμ(x)]| →p0.
The proof of this lemma is similar to that of Lemma 3.1 of Chang (1990) and Lemma 4.2 of Koul and Stute (1999). Details are left out for the sake of brevity.
Now, let h be a positive function on \({\mathbb R}_{+}^{p}\) with \(0<{\int \limits } h^{2} dG<\infty \) and \({\int \limits } h dG=1\). Let
Note that βnk(0) = 0 = βk(0), \(0< \beta _{k}(z)\le \beta _{k}(\mathbf {\infty })={\int \limits } h^{k} dG<\infty , \) for all \(z\in {\mathbb R}_{+}^{p}\) and Eβnk(z) ≡ βk(z), k = 1, 2. By the Ergodic Theorem and a Glivenko-Cantelli type argument, see Stute (1976) and Koul (2019),
Thus for all sufficiently large n, βnk(z) > 0, for all \(z\in (0,\infty ]^{p}, k=1, 2\), Bn2 > 0 (a.s.). The arguments below are carried out on the event \(B_{n2}={\int \limits } h^{2} dG_{n}>0\).
Recall θnt = θ + n− 1/2t. By the Cauchy-Schwarz inequality,
Let \(\alpha (z):= {\int \limits }_{x\ge z} d\beta _{1}(x)\), \(\alpha _{n}(z):= {\int \limits }_{x\ge z} d\beta _{n1}(x)\), \(z\in {\mathbb R}_{+}^{p}\) and
Because αn(z) ≥ 0 for all \(z\in {\mathbb R}_{+}^{p}, n\ge 1\), w.p.1, by (4.8), \({\mathcal V}_{n2}(r e)\) is monotonic in r for all e ∈ Rq, ∥e∥ = 1. Hence
Let
By Lemma A.1, the process \(U_{n}(z), z\in [0,\infty ]^{p}\) is relatively compact with respect to the uniform metric. Hence (A.12) and Lemma A.2 applied with \(\mu _{n}=\beta _{n1}/\beta _{n1}({\mathbf {\infty }}), \mu =\beta _{1}\) yield
Here we have used the fact \(\beta _{1}({\mathbf {\infty }})={\int \limits } h dG=1\) and \(\beta _{n1}({\mathbf {\infty }})\to _{p} \beta _{1}({\mathbf {\infty }})=1\).
Let \(\bar {\mathcal V}_{n2}(t):={\mathcal V}_{n}^{*} - t^{\prime }{\int \limits } {\Psi } d\beta _{1}\). Because by the Ergodic Theorem, \({\int \limits } {\Psi } d\beta _{n1}\to _{p} {\int \limits } {\Psi } d\beta _{1}\) and
by (A.1), we obtain
Therefore, by (A.7), (A.12) and (A.13)
Also note that \(\sup _{\|t\|\le b} |\bar {\mathcal V}_{n2}(t)|=O_{p}(1)= \sup _{\|t\|\le b} | {\mathcal V}_{n2}(t)|\) and by (A.12),
This fact in turn implies that
Fix an 𝜖 > 0 and η > 0. By (A.14), there exists N𝜖, η such that
Moreover, \(E\big ({\mathcal V}_{n}^{*}\big )^{2}= \sigma ^{2} E\alpha ^{2}(Z_{0})\le \sigma ^{2} B_{2}<\infty \). Hence, by the Markov inequality, for any 𝜖 > 0 there exists b𝜖 such that
Let \(K:= {\int \limits } \|{\Gamma }\| d\beta \) and b𝜖 be as in (A.15). Choose b ≥ (b𝜖 + (B2η)1/2)K− 1 and argue as for Eq. A.11 to obtain that ∀n ≥ N𝜖, η,
thereby proving the claim about (C.4). This also completes the proof of Lemma 4.1. □
Rights and permissions
About this article
Cite this article
Koul, H.L., Perera, I. & Balakrishna, N. A class of Minimum Distance Estimators in Markovian Multiplicative Error Models. Sankhya B 85 (Suppl 1), 87–115 (2023). https://doi.org/10.1007/s13571-021-00274-x
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13571-021-00274-x