Abstract
The present work proposes a new stationary integer-valued bilinear time series model with dependent counting series. The model will enable one to tackle the presence of some correlation between underlying events. The various probabilistic and statistical properties of the model are discussed, unknown parameters are estimated by several methods. Moreover, the performance of the estimation methods is illustrated through a simulation study and an empirical application to two data sets.
Similar content being viewed by others
References
Basrak B, Davis RA, Mikosch T (1999) The sample ACF of a simple bilinear process. Stoch Process Appl 83:1–14
Bentarzi M, Bentarzi W (2017) Periodic integer-valued bilinear time series model. Commun Stat Theory Methods 46:1184–1201
Davis RA, Resnick SI (1996) Limit theory for bilinear processes with heavy-tailed noise. Ann Appl Probab 6:1191–1210
Drost FC, Akker VDR, Werker BJM (2008) Note on integer-valued bilinear time series. Stat Prob Lett 78:992–996
Doukhan P, Latour A, Oraichi D (2006) A Simple integer-valued bilinear time series model. Adv Appl Prob 38(2):559–578
Granger CWJ, Andersen AP (1978a) An introduction to bilinear time series models. Vandenhoeck and Ruprecht, Gottingen
Granger CWJ, Andersen AP (1978b) On the invertibility of time series models. Stoch Process Appl 8:87–92
Keenan DM (1985) A Tukey non-additivity type test for time series nonlinearity. Biometrika 72:39–44
Kim HY, Park Y (2008) A non-stationary integer-valued autoregressive model. Stat Papers 49:485–502
Liu M, Li Q, Zhu F (2019) Threshold negative binomial autoregressive model. Stat 53:1–25
Liu M, Li Q, Zhu F (2020) Self-excited hysteretic negative binomial autoregression. AStA Adv Stat Anal 104:385–415
Mohammadpour M, Bakouch HS, Ramzani S (2019) An integer-valued bilinear time series model via two random operators. Math Comput Modell Dyn Syst 25:429–446
Nastić AS, Ristić MM, Miletić Ilić AV (2017) A geometric time series model with an alternative dependent Bernoulli counting series. Commun Stat Theory Methods 46:770–785
Pascual L, Romo J, Ruiz E (2004) Bootstrap predictive inference for ARIMA processes. J Time Ser Anal 25:449–465
Qin J, Lawless J (1994) Empirical likelihood and general estimationg equations. Ann Stat 22:300–325
Ristić MM, Nastić AS, Miletić Ilić AV (2013) A geometric time series model with dependent Bernoulli counting series. J Time Ser Anal 34:423–516
Subba Rao T (1981) On the theory of bilinear time series model. J R Stat Soc Ser B 43:244–255
Turkman KF, Turkman MAA (1997) Extremes of bilinear time series models. J Time Ser Anal 18:305–319
Wang C, Liu H, Yao J, Davis RA, Li WK (2014) Self-excited threshold Poisson autoregression. J Amer Stat Assoc 109:776–787
Zhang Z, Tong H (2001) On some distributional properties of a first-order nonnegative bilinear time series model. J Appl Probab 38(3):659–671
Zhang H, Wang D, Zhu F (2011) Empirical likelihood inference for random coefficient INAR(p) process. J Time Ser Anal 32:195–203
Zhu F, Wang D (2008) Estimation of parameters in the NLAR(p) model. J Time Ser Anal 29:619–628
Zhu F (2011) A negative binomial integer-valued GARCH model. J Time Ser Anal 32:54–67
Zhu F, Wang D (2015) Empirical likelihood for linear and log-linear INGARCH models. J Korean Stat Soc 44:150–160
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A: Proof of Theorem 2.1
To investigate the stationarity of the precess, we follow a similar approach as that of Kim and Park (2008). Let \(\{X_{t}^{(n)}, t{\in \mathbb {Z}}\}\) be a sequence of random variables as
where sequence \(\{X_{s}^{(n)}\}\) is independent of et for s < t.
B1
The process \(\{X_{t}^{(n)}, t{\in \mathbb {Z}}\}\) is strictly stationary for any \(n \in \mathbb {N}\).
The strict stationarity of the process \(\{X_{t}^{(n)}, t{\in \mathbb {Z}}\}\) will be deduced by showing that the two vectors \((X_{1}^{(n)},...,X_{k}^{(n)})^{T}\) and \( (X_{1+h}^{(n)},...,X_{k+h}^{(n)})^{T}\) are identically distributed. It is clear that process \(\{X_{t}^{(n)}, t{\in \mathbb {Z}}\}\) is strictly stationary for n = 0. Now suppose that the process \(\{X_{t}^{(m)}, t{\in \mathbb {Z}}\}\) is strictly stationary for all 1 ≤ m ≤ n − 1. Hence for m = n we have
and
According to the induction hypothesis, the random vectors \( (X_{1}^{(n)},...,X_{k}^{(n)})^{T}\) and \((X_{1+h}^{(n)},...,X_{k+h}^{(n)})^{T}\) are identically distributed.
B2
The sequence \(\{X_{t}^{(n)}, t{\in \mathbb {Z}}\}\) belongs to the space \(\pounds ^{2}=\{X|EX^{2}<\infty \}\).
Let \(\mu _{n}=E(|X_{t}^{(n)}|)\). According to Eq. 12, we have
As \(E\left \vert X_{t}^{(n)}-e_{t}\right \vert =E|a\mathbf {\diamond _{\theta }} X_{t-1}^{(n-1)}+b\mathbf {\diamond _{\delta } }X_{t-1}^{(n-1)}e_{t-1}| \leqslant E|X_{t}^{(n)}|+\mu ,\) we have
Substitution of Eqs. 14 into 13 gives
Now we show that \(E|X_{t}^{(n)}|^{2}\leqslant \infty \).
On the other hand, after some calculations we have
and
where \(L_{1}(e)=\mu (3\mu _{e^{2}}+2\mu ^{2})+2\mu _{e^{2}}\mu +\mu _{e^{3}}\) and \(L_{2}(e)=\mu _{e^{2}}(3\mu _{e^{2}}+2\mu ^{2})+2\mu _{e^{3}}\mu +\mu _{e^{4}}\). By substituting these equations and Eqs. 14 into 15, we have
where \(K_{1}(e)=(|a \theta |+|b\delta |\mu _{e^{2}} )+2ab\mu \), \( K_{2}(e)=|a(1-\theta ) |+|b(1-\delta ) |\mu +2\mu _{e^{2}}\mu +\mu _{e^{3}}+4|a||b|(\mu _{e^{2}}+\mu ^{2})+2\mu (|a|+|b|\mu )\) and \( K_{3}(e)=|b(1-\delta ) |(\mu _{e^{2}}+\mu ^{2})+|b|^{2}L_{2}(e)+2|a||b|L_{1}(e)+2\mu |b|(\mu _{e^{2}}+{\mu _{e}^{2}})\).
Repeating (16), we have
Therefore under strictly stationary condition, it can be concluded that \( E|X_{t}^{(n)}|^{2}\leqslant \infty ,\) n ≥ 1. Hence \(\{X_{t}^{(n)}\}\in \pounds ^{2}\).
B3
The sequence \(\{X_{t}^{(n)}\}\) is Cauchy.
Let \(\psi (t,n,m)=|X_{t}^{(n)}-X_{t}^{(n-m)}|,\) m = 1, 2,.... Using the definition of \(\{X_{t}^{(n)}\}\), we have
As \(n\rightarrow \infty \), under strictly stationary condition, Eψ(t,n,m) converges to 0. Similarly, we have
As \(n\rightarrow \infty \), under given condition, it easily can be seen that Eψ2(t,n,m) converges to 0. So \(X_{t}^{\left (n\right ) }\) is a Cauchy sequence. Suppose \(\lim _{n\rightarrow \infty }X_{t}^{(n)}=X_{t}\), then Xt ∈£2.
B4
The process {Xt} satisfies (1)
Since \(X_{t}^{(n)}\rightarrow ^{\pounds ^{2}}X_{t}\), so
and
converge to zero. Hence process {Xt} satisfies (1).
B5
Uniqueness
Suppose that another solution \(X_{t}^{\ast }\) of Eq. 1 exists. By Minkowski inequality, we have
so \(X_{t}=X_{t}^{\ast } \) a.s.
B6
Strictly stationarity of the process {Xt}.
As the process \(\{X_{t}^{(n)}\} \) is strictly stationary, so \( (X_{0}^{(n)},...,X_{k}^{(n)})^{T}\) and \((X_{h}^{(n)},...,X_{k+h}^{(n)})^{T}\) have the same distribution for each n,h and k. Since \( X_{t}^{(n)}\rightarrow ^{\pounds ^{2}}X_{t}\), it can be deduced \( X_{t}^{(n)}\rightarrow ^{P}X_{t}\) and hence \( {\sum }_{i=0}^{h}b_{i}X_{t+i}^{(n)}\rightarrow ^{P}{\sum }_{i=0}^{h}b_{i}X_{t+i}.\) Using Cramer-Wold device, it can be concluded that
and
Since \((X_{0}^{(n)},...,X_{k}^{(n)})\) and \((X_{h}^{(n)},...,X_{k+h}^{(n)})\) have the same distribution, hence their limits, (X0,...,Xk) and (Xh,...,Xk+h) have the same distribution. This completes the proof.
Appendix B: Proof of Proposition 2.1
Using the stationarity of the process and the properties of the operator, E(Xt) will be obtained. To obtain \(E({X^{2}_{t}})\), we have
where Bt− 1 = a ◇𝜃Xt− 1 + b ◇δXt− 1et− 1. Using the stationarity of the process and the properties of the operator, we have
and
where
After some simplifications, we have
where \(C =a(1-\theta )E(X_{t-1})+b\delta (E({e_{t}^{4}})+2E({e_{t}^{3}}) E(B_{t-1}))+b(1-\delta )(E(B_{t-1}) E(e_{t}) +E({e_{t}^{2}})) + 2ab(E({e_{t}^{3}})+2E({e_{t}^{2}}) E(B_{t-1}) )\). Hence substitution of \( E(B_{t-1}^{2})\) into \(E({X_{t}^{2}})\), \( E({X_{t}^{2}})\) can be obtained.
Appendix C: Proof of Proposition 2.2
For the autocovariance function of order one, we have
where \(E({X_{t}^{2}}e_{t})\) and \(E(B_{t-1}^{2})\) are obtained as
Substitution \(E(B_{t-1}^{2})\) into the second term of \(E({X_{t}^{2}}e_{t})\ \), after some calculations we have
and hence
Also
After some tedious computations, we obtain E(XtXt+ 1et+ 1) = σ2E(Xt) + μE(XtXt+ 1). By substituting E(XtXt+ 1et+ 1) in E(XtXt+ 2), we have
On the other hand, Eq. 2 implies that \(b\sigma ^{2}\mu _{X}+\mu \mu _{X}+[(a+b\mu )-1]{\mu _{X}^{2}}=0\). Hence
So by induction, we can conclude that
Appendix D: Proof of Proposition 2.3
The one step ahead conditional expectation can be obtained directly. Since
we have
Using the one step ahead conditional expectation E[Xt+ 1|t] = μ + (a + bet)Xt, we can be conclude that
and
Hence by induction, we conclude
Now we obtain an expression for \(E(X_{t+k}^{2}|t).\) One can easily find \( E(X_{t+1}^{2}|t) \). Also
where \(E(X_{t+1}^{2}e_{t+1}|t)\) and \(E(X_{t+1}^{2}e_{t+1}^{2}|t)\) are obtained as
Substituting the above equations in \(E(X_{t+2}^{2}|t)\) and after some calculations, we have
Also, we can find that
So by induction and properties of the i.i.d. process \(\{e_{t}, t{\in \mathbb {Z}}\}\), we can conclude that
where E(Bt+k−i|t) = E(Xt+k−i+ 1 − μ|t) for i = 0, 1,...,k.
Appendix E
Some expectations which are used in the Yule Walker method are calculated in this part.
Let Bt− 1 = a ◇𝜃Xt− 1 + b ◇δXt− 1et− 1. So Bt− 1 = Xt − et and we have
and
It is noted that under P(μ) distribution of innovations {et}, we have
Rights and permissions
About this article
Cite this article
Ramezani, S., Mohammadpour, M. Integer-valued Bilinear Model with Dependent Counting Series. Methodol Comput Appl Probab 24, 321–343 (2022). https://doi.org/10.1007/s11009-021-09853-x
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11009-021-09853-x