Fractional ARIMA with stable innovations (Q1909951)

If \(X\) is a random variable with a characteristic function \(Ee^{iaX} = \exp \{-\sigma^\alpha |a |^\alpha\}\), \(0 < \alpha \leq 2\), then \(X\) is called symmetric \(\alpha\)-stable \((S \alpha S)\). A stochastic process \(\{X_t\}\) is called \(S \alpha S\) if every linear combination \(\sum^n_{k = 1} a_k X_{t_k}\) is \(S \alpha S\). Since the covariance does not exist for \(S \alpha S\) variables with \(\alpha < 2\), codifference \(\tau (X_1, X_2)\) and covariation \([X_1, X_2]_\alpha\) are used instead. They are defined by \[ \tau (X_1, X_2) = \ln E \exp \bigl\{ i(X_1 - X_2) \bigr\} - \ln E \exp \{iX_1\} - \ln E \exp \{- iX_2\}, \] \[ \qquad \text{and} \quad [X_1, X_2]_\alpha = \int_{S_2} s_1 s_2^{\langle \alpha - 1 \rangle} \Gamma (ds), \] where \(a^{\langle q \rangle} = |a |^q \text{sign} (a)\) is the signed power and \(\Gamma\) is a Borel measure on the unit circle \(S_2 = \{(s_1, s_2) : s^2_1 + s^2_2 = 1\}\). If \((X_1, X_2)\) is a Gaussian vector \((\alpha = 2)\) then \(\tau (X_1, X_2) = \text{cov} (X_1, X_2)\) and \([X_1, X_2]_2 = 2^{-1} \text{cov} (X_1, X_2)\). The authors develop the theory of fractionally differenced ARIMA time series \(\{X_n\}\) defined by \[ \Phi_p (B) X_n = \Theta_q (B) (1 - B)^{-d} \varepsilon_n, \] where \(\Phi_p\) and \(\Theta_q\) are polynomials, \(B\) is the back-shift operator, \(\varepsilon_n\) are i.i.d. \(S \alpha S\) random variables with \(0 < \alpha < 2\), and \(d\) is fractional. Conditions for existence of \(X_n\) and invertibility are established. The asymptotic dependence structure is analyzed by means of the codifference and the covariation.