1
$\begingroup$

According to "Time Series Analysis, Forecasting and Control" (Box, Jenkins), for any process represented by a linear filter $$ z_t = a_t + \psi_1 a_{t-1} + \psi_2 a_{t-2} ... = a_t + \sum_{J=1}^\infty \psi_j a_{t-j} $$ be a valid stationary process, it is necessary for the coefficients $\psi_j$ to be absolutely summable: $\sum_{J=1}^\infty |\psi_j| < \infty$.

For an ARIMA(0, 1, 1) process, $$ (1-B)Y_t = (1-\theta_1 B)a_t\\ Y_t = \frac{(1-\theta_1B)}{(1-B)}a_t = [1 + (1-\theta_1)B + (1-\theta_1)B^2 + (1-\theta_1)B^3 ...] \cdot a_t. $$

Thus $\psi_j = (1-\theta_1)$ for all $j$.

Since, I think, $\sum_{j=1} ^\infty |\psi_j|= \infty$, doesn't this mean that an ARIMA(0, 1, 1) process can't be stationary?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

You cannot manipulate the lag operator $B$ as you did when there is unit root. Box and Jenkins should tell you that. $\theta_1$ should be $\theta_1 B$, in your notation.

(The heuristic $$ ``\frac{1}{1-B} = \sum_{h \geq 0} B^h" $$ that your heuristic calculation is based on is incorrect. The precise reason---see, e.g. Box and Jenkins---is that, for a complex polynomial $f(z)$, $\frac{1}{f(z)}$ admit a power series representation $$ \frac{1}{f(z)} = \sum_{h \geq 0} \psi_h z^h $$ on an open neighborhood in the complex plane if and only if the roots of $f(z)$ lie strictly outside the unit circle.)

As to your question, yes, ARIMA(0,1,1) can be stationary. For example, take a white noise. Its first difference is an MA(1) series. Integrate this over-differenced MA(1) series gives the original white noise. (Indeed, if you take the first difference of any stationary series and integrate it, you recover the stationary series.)

Improve this answer
$\endgroup$
8
  • $\begingroup$ Thanks for catching the $B$ error. But I left $\theta_1$ arbitrary, so in the above situation, even if there is no unit root to the characteristic polynomial, wouldn't the infinite sum of $\psi_i$ still go to infinity for any $\theta_1$? $\endgroup$
    – Frank
    Commented Feb 15, 2020 at 6:33
  • 1
    $\begingroup$ Your algebra is still incorrect, Frank: every "$1-\theta_1$" in your final expression for $Y_t$ should be $1-\theta_1B.$ You will need to collect like powers of $B$ in order to determine the $\psi_j$ correctly. $\endgroup$
    – whuber ♦
    Commented Feb 15, 2020 at 15:31
  • $\begingroup$ @whuber Not sure what you mean : $\frac{1-\theta B}{1-B} = 1 + \frac{B-\theta B}{1-B} = 1 + \frac{(1-\theta)B}{1-B} = 1 + (1-\theta) B + \frac{(1-\theta)B - [(1-\theta)B (1-B)]}{1-B} = 1 + (1-\theta) B + \frac{(1-\theta)B^2}{1-B}...$ $\endgroup$
    – Frank
    Commented Feb 15, 2020 at 18:10
  • 1
    $\begingroup$ $$\frac{1-\theta B}{1-B} = (1-\theta B)(1+B+B^2+\cdots+B^n+\cdots) = 1+(1-\theta)B+(1-\theta)B^2+\cdots$$ implying $\psi_j=1-\theta$ for $j\ge1,$ whence $\sum_{j\ge 1}|\psi_j| = 1/\theta$ provided $0\lt \theta \lt 2.$ $\endgroup$
    – whuber ♦
    Commented Feb 15, 2020 at 19:22
  • $\begingroup$ @whuber, Thank you for the reply. Take $\theta = 1/2$, then $\sum_{j \geq 1} |1-\frac{1}{2}| = \sum_{j \geq 1} \frac{1}{2}$. You're saying the infinite sum of $\frac{1}{2} = \frac{1}{1/2} = 2$ ? For any $n$, $\sum_1^n \frac{1}{2} = \frac{n}{2}$. You're saying this converges to $2$? $\endgroup$
    – Frank
    Commented Feb 15, 2020 at 22:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.