I have a question about MA(2) model $X_t = e_t + 0.5e_{t−1} + 0.4e_{t−2}$ with $e_t \sim IID(0, σ^2_e $).

I know by construction, MA process is (weakly) stationary, but can it be strictly stationary given IID?

To show it is strictly stationary, I get the $E(X_t)$ and $E(X_{t-5})$ and $V(X_t)$ and $V(X_{t-5})$ and found that both are the same ($\text{mean}=0$, and $\text{var}=1.41\sigma^2$).

So with that, is the process time-invariant? However, I know that the variance is finite so it implies WSS. I do not get a grab on this concept. I do confuse with IID that assumed the process to be strictly stationary.

  • $\begingroup$ Hi: stricly stationary means that if you look at observations at time t and calculate the joint distribution of say $n$ data points at that time and then look at the data at time $t+h$ and calculate the joint distribution of $n$ data points at that time, then they are the same. So, I think you are doing that sort of but you should check three data points say $x_1$, $x_2$, $x_3$ versus say $x_5$, $x_6$, $x_7$ since you have an MA(2). I say two because the MA(2) process is not correlated after lag 2 so there's really nothing to check except 3 data points. $\endgroup$
    – mlofton
    Commented Apr 7, 2021 at 15:25

1 Answer 1


There's a deeper and far more general result lurking here: a moving window operation on any strictly stationary process produces a strictly stationary process.

The demonstration is just a matter of applying definitions. I'll take you through it step by step, emphasizing the simplicity and--I hope--the obviousness of each one.

Let's begin with this "initial statement" of a trivial proposition. It will be applied repeatedly below.

Let the $p$-variate random variables $\mathbf X = (X_1,\ldots, X_p)$ and $\mathbf Y = (Y_1,\ldots, Y_p)$ have the same distribution and let $f:\mathbb{R}^p\to \mathbb R^q$ be any (measurable) function. Then $f(\mathbf X)$ and $f(\mathbf Y)$ are identically distributed.

Proof: the distributions of the results of applying $f$ depend only the the distributions of its arguments, which are identical, QED. Note that the resulting identical distributions are $q$-variate and we're talking about the full joint distribution, not just the marginal distributions of their components.

Suppose, then, that $\mathbf X = (\ldots, X_{-1}, X_0, X_1, \ldots)$ is a time series process. As a matter of notation, whenever $T$ is a finite sequence $(t_1,\ldots, t_p)$ of integers, define

  • $\mathbf{X}_T = (X_{t_1}, X_{t_2}, \ldots, X_{t_p})$ to be the sequence of components of the process picked out by the indexes in $T$ and

  • For any integer $h,$ let $T + h = (t_1+h, t_2+h, \ldots, t_p+h)$ be the translate of the indexes in $T.$

A process is strictly stationary when, for any finite sequence $T$ and any integer $h,$ $\mathbf{X}_T$ and $\mathbf{X}_{T+h}$ have the same distribution.

This is the definition. In words, it says all finite-dimensional marginal distributions of the process (namely, all $\mathbf{X}_T$) are time-invariant (namely, they do not change when translated by any lag $h$).

Take any $f$ as in the initial statement. For any finite sequence $T$ and integer $h,$ it follows that $f(\mathbf{X}_T)$ and $f(\mathbf{X}_{T+h})$ are identically distributed.

Let's apply these trivia to the situation in the question. Let $g:\mathbb{R}^r\to\mathbb{R}$ be a (measurable) function. Define a new time series process from the process $\mathbf X$ via

$$Y_t = g(X_{t-r+1}, X_{t-r+2}, \ldots, X_t).$$

This is the "moving window" application of $g$ to windows of length $r.$ We need to convince ourselves that when $\mathbf{X}$ is strictly stationary, so is $\mathbf Y.$

To show this, we must consider an arbitrary finite index set $S=(s_1,s_2,\ldots, s_q)$ and arbitrary integer $h$ and compare the distributions of $\mathbf{Y}_S$ and $\mathbf{Y}_{S+h}.$ Notice that, by construction,

$$\mathbf{Y}_S = (Y_{s_1}, Y_{s_2}, \ldots, Y_{s_q}) = (g(X_{s_1-r+1}, \ldots, X_{s_1}), \ldots, g(X_{s_q-r+1}, \ldots, X_{s_q})).$$

The right hand side can be expressed as the image of a function $f:\mathbb{R}^{rq}\to \mathbb{R}^{q}$ defined by

$$f(x_1, x_2, \ldots, x_{rq}) = (g(x_1,\ldots, x_q), g(x_{q+1}, \ldots, x_{2q}), \ldots, g(x_{(r-1)q+1}, \ldots, x_{rq})).$$

The initial statement (with $p=rq$) applies because this $f$ is (obviously) measurable. The relevant index set is

$$T = (s_1-r+1, s_1-r+2, \ldots s_1,\quad s_2-r+1,s_2-r+2,\ldots, s_2,\quad \ldots\\\quad s_q-r+1, s_q-r+2, \ldots, s_q).$$

(Many of these indexes might be the same. I never insisted the indexes must be unique!)

Because $\mathbf X$ is strictly stationary, $\mathbf{X}_T$ and $\mathbf{X}_{T+h}$ have the same $rq$-variate distributions. But--applying the initial statement once more--this means $\mathbf{Y}_S = f(\mathbf{X}_T)$ and $\mathbf{Y}_{S+h}= f(\mathbf{X}_{T+h})$ have the same $q$-variate distributions, proving $\mathbf{Y}$ is strictly stationary.

The question itself concerns the window function $g:\mathbb{R}^3\to\mathbb{R}^1,$ $g(x_1,x_2,x_3) = 0.4x_1 + 0.5x_2 + x_3,$ as applied to an iid sequence of variables $(e_t),$ which is clearly strictly stationary. Because this (linear) map $g$ is measurable, we're done.


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.