Let $\ MA(q)$ be a moving average process of order $\ q$, and $\hat{y} = \mu + a_t$ be a constant forecast model whose level is represented by $\mu $ and the random term at time $\ t$ by $\ a_t$.

Now, suppose $\mu \equiv \hat{a_1}$ so one can write $\hat{y} = \ a_1 + a_t$.

If $\hat{a}_1$ is calculated by a $\ MA(q)$, what would the variance of $\hat{y}$ be?

What if, then, $\hat{a}_1$ is calculated by a $\ MA(2q)$ or a $\ MA(q^2)$ - how would that affect the variance of $\hat{y}$?

My goal is understand how the variance of my forecasted values will vary as I vary$\ q$.

Bob Stine, from UPenn, made some actual calculations on page 9 of this document, but the way he calculates the variance for 3 steps ahead is something I'm yet to understand. More precisely, I don't get why he keeps getting rid of his $ a_n$ terms.

Thank you.


Edit: I'll try to show some of my thoughts on this matter here:

On page 86, Montgomery, Johnson,and Gardiner (1990) write that:

$\large var(M_T) = \large var[\frac{1}{N}\sum_{k=0}^{N-1} x_{T-k}] = \large \frac{\sigma^2_\epsilon}{N}$

where $\ M_T$ is the $\ MA$ process, $\ N$ is the number of the most recent observations taken into account, and $\sigma^2_\epsilon$ is the variance of the noise (random component).

However, if I write the $\ MA$ process as $\phi_p. z_t = \theta_q . a_t$, for $\phi_p$ equals 1, I have that my time series $z_t$ is equal to $\theta_q . a_t$, where $\theta_q$ is the lag operator of order $q$ and $a_t$ is my random component.

Thus, I can write $z_t$ as $z_t = (1 + B + B^2 + ... + B^q).a_t$, and then $z_t = a_t + a_{t-1} + a_{t-2} + ... + a_{t-q+1}$.

If then I take the variance of $z_t$, $var(z_t)$, I take the variance of $a_t + a_{t-1} + a_{t-2$} + ... + a_{t-q+1}$, which is the same as:

$var(z_t)$ = $\sum_{k=1}^{t-q+1} var(a_t)$. Now, if $var(a_t) = \sigma^2_\epsilon$, the variance of z will be equal to $N.\sigma^2_\epsilon$, where $N$ is defined as above.

My result clearly doesn't match that of Montgomery et al.

Assuming it's quite fair to say I made a mistake, where is it?


1 Answer 1


In order to keep it simple, I use a MA(q)-process where $\mu=0$ and $E[a_t] = 0$.

  1. Take the random shock model and express the MA(q) process as a linear filter, $y_t = \Psi(B) \cdot a_t = \sum_{j=0}^\infty \psi_j \cdot a_{t-j}$. Therefore, \begin{align} y_{t+\ell} &= \sum_{j=0}^\infty \psi_j \cdot a_{t + \ell -j} \\ &= \sum_{j=0}^{\ell-1} ... + \sum_{j=\ell}^{\infty} ... \end{align} Looking at this eq. from the time $t$ (in the forward direction), the first sum contains only future terms, while the second contains only past terms.
  2. Suppose that the "best" forcast (smallest RMS) at time $t$ and lead time $\ell$ is given by $$\hat{y}_t(\ell) = \sum_{j=0}^\infty \psi^\star_{j+\ell} a_{t-j}$$ where the $\psi^\star_{j+\ell}$ need to be determined. Note, that the forecast contains only past terms. Hence, using the eq. from (1) we know that $$E[y_{t+\ell} - \hat{y}_t(\ell)]^2 = \sum_{j=0}^{\ell-1} \psi^2_j \cdot \sigma^2_a + \sum_{j=\ell}^{\infty} (\psi_{\ell+j} - \psi^\star_{\ell + j} )^2 \sigma^2_a$$ where the first term contains the future terms and the second the past terms. Clearly this is minimized by setting $\psi^\star_{\ell + j} = \psi_{\ell+j}$.
  3. Put this result back to eq. (1) we identify the to sums, \begin{align} y_{t+\ell} &= \sum_{j=0}^{\ell-1} ... + \sum_{j=\ell}^{\infty} ... \\ &= e_t(\ell) + \hat{y}_t(\ell) \end{align} as the error of the forecast, $e_t(\ell)$, and the forcast, $\hat{y}_t(\ell)$, at time $t$ and lead time $\ell$, respectively.
  4. Now that we have an expression for the error of the forecast, we use the index of $a_t$ to obtain its variance, $$var[e_{t}(\ell)] = \sum_{j=0}^{\ell-1} \psi^2_j \sigma_a^2$$

Finally, note that this calc is valid for an ARMA(p, q)-process. For a MA(q) we simply use $\psi_j =0 $ for $j>q$.


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.