Specifying MA(q) Models And Their Finite Dependence

Why an MA(q) Process Has a Finite Innovation Window

A moving-average model describes an observation as a weighted combination of a current random shock and a fixed number of past shocks. The key feature is not that the observations themselves are moving averages of earlier observations. Rather, each observation uses a finite window of innovations. Once two such windows no longer overlap, the assumptions on the innovations determine the relationship between the corresponding observations.

This finite-window view is useful when interpreting an MA model and reasoning about dependence. It also prevents a common error: the cutoff statement says that observations more than \(q\) time steps apart are uncorrelated. It does not say that all observations within \(q\) time steps must be correlated.

Definition: An MA(\(q\)) process with mean \(\mu\) has the form

X_t=\mu+\varepsilon_t+\theta_1\varepsilon_{t-1}+\theta_2\varepsilon_{t-2}+\cdots+\theta_q\varepsilon_{t-q},

where \(q\) is a nonnegative integer, the coefficients \(\theta_1,\ldots,\theta_q\) are fixed, and \(\{\varepsilon_t\}\) is a white-noise sequence with \(E[\varepsilon_t]=0\), \(\operatorname{Var}(\varepsilon_t)=\sigma^2<\infty\), and \(\operatorname{Cov}(\varepsilon_t,\varepsilon_s)=0\) for \(t\ne s\). The coefficient on the current innovation is conventionally 1. The exact order is \(q\) when \(\theta_q\ne 0\); otherwise the representation can be written with a smaller order.

The coefficients are sometimes written as \(\theta_0=1,\theta_1,\ldots,\theta_q\), so the same model is a finite weighted sum from lag 0 through lag \(q\). For example, the coefficient \(\theta_2\) multiplies the innovation from two time steps earlier. It does not multiply \(X_{t-2}\).

White noise requires zero mean, a common finite variance, and zero covariance between distinct times. It does not, by itself, require normally distributed or independent innovations. Consequently, results based only on uncorrelatedness should not be strengthened to independence unless a stronger assumption is stated.

Worked Example: Expanding and Evaluating an MA(3)

A hypothetical quarterly claim-index series is modeled by \(X_t=20+\varepsilon_t+0.4\varepsilon_{t-1}-0.2\varepsilon_{t-2}+0.1\varepsilon_{t-3}\). The innovations are white noise with mean 0 and variance 9. For one quarter, the observed innovation values are \(\varepsilon_t=2\), \(\varepsilon_{t-1}=-1\), \(\varepsilon_{t-2}=3\), and \(\varepsilon_{t-3}=4\). Find the model value for that quarter and state which innovations enter it.

Method. Substitute the four relevant innovation values into the finite weighted sum, keeping each lag attached to its coefficient.

The model uses the current innovation and the three preceding innovations: \(\varepsilon_t,\varepsilon_{t-1},\varepsilon_{t-2},\varepsilon_{t-3}\). Substitution gives \(X_t=20+2+0.4(-1)-0.2(3)+0.1(4)\).

Evaluate the weighted contributions one at a time: \(0.4(-1)=-0.4\), \(-0.2(3)=-0.6\), and \(0.1(4)=0.4\). Therefore, \(X_t=20+2-0.4-0.6+0.4=21.4\).

Check. The innovation contribution is \(2-0.4-0.6+0.4=1.4\), so adding the mean 20 gives 21.4. No innovation older than \(t-3\) appears. The variance 9 is part of the innovation assumptions, but it is not needed to evaluate this particular realization.

Common mistake avoided. Treating the model as a weighted sum of past observations would substitute \(X_{t-1}\), \(X_{t-2}\), and \(X_{t-3}\). Those are not the terms in an MA model: the lagged terms here are innovations.

From Overlapping Windows to Zero Covariance

To see why the finite window matters, list the innovation indices used by an observation. \(X_t\) uses indices \(t,t-1,\ldots,t-q\). The observation \(X_{t+h}\) uses indices \(t+h,t+h-1,\ldots,t+h-q\). If \(h>q\), the earliest index in the latter window is \(t+h-q\), which is strictly later than \(t\). Thus the two windows have no index in common.

The covariance argument follows directly from the white-noise assumptions. Subtracting the means leaves finite sums of innovations. By bilinearity, the covariance of those sums is the sum of the pairwise covariances of their innovation terms. Every pair comes from distinct time indices when the windows do not overlap, and each such covariance is zero.

Result: Under the white-noise assumptions in the definition, \(X_t\) and \(X_{t+h}\) are uncorrelated whenever \(|h|>q\). This is a guaranteed cutoff for covariance, not a claim that covariance is nonzero at every lag from 1 through \(q\).

Here is the justification in full for \(h>q\). The centered observations are finite sums. When the sums are multiplied inside the covariance, only terms involving the same innovation time could have nonzero covariance. No such time occurs in both windows, so every term is zero:

\begin{aligned} \operatorname{Cov}(X_t,X_{t+h}) &=\operatorname{Cov}\left(\sum_{j=0}^{q}\theta_j\varepsilon_{t-j}, \sum_{k=0}^{q}\theta_k\varepsilon_{t+h-k}\right)\\ &=\sum_{j=0}^{q}\sum_{k=0}^{q} \theta_j\theta_k\operatorname{Cov}(\varepsilon_{t-j},\varepsilon_{t+h-k})\\ &=0,\qquad h>q, \end{aligned}

In this display, \(\theta_0=1\). For a term to have matching innovation indices, \(t-j=t+h-k\), which requires \(k-j=h\). But \(j\) and \(k\) both range from 0 to \(q\), so \(k-j\) cannot exceed \(q\). When \(h>q\), no matching pair exists. For negative lags, covariance symmetry gives the same conclusion for \(|h|>q\).

This argument proves a zero-covariance statement and does not require a detailed formula for the covariance at each shorter lag. It also shows why the threshold is strictly greater than \(q\): at lag \(q\), the innovation windows can still share an endpoint.

Worked Example: Checking the Boundary Lag in an MA(2)

Let \(X_t=5+\varepsilon_t+0.5\varepsilon_{t-1}-0.25\varepsilon_{t-2}\), where the innovations are white noise with variance 4. Determine whether \(X_t\) and \(X_{t+2}\) are uncorrelated, and then determine whether \(X_t\) and \(X_{t+3}\) are uncorrelated.

Method. Expand both observations and identify shared innovation indices. Only a shared index can contribute to covariance, because distinct innovations have zero covariance.

At lag 2, \(X_t=5+\varepsilon_t+0.5\varepsilon_{t-1}-0.25\varepsilon_{t-2}\), while \(X_{t+2}=5+\varepsilon_{t+2}+0.5\varepsilon_{t+1}-0.25\varepsilon_t\). The only shared innovation is \(\varepsilon_t\). Its coefficient is 1 in \(X_t\) and \(-0.25\) in \(X_{t+2}\). Its covariance contribution is \(1(-0.25)(4)=-1\). All terms involving different innovation times contribute zero. Therefore, \(\operatorname{Cov}(X_t,X_{t+2})=-1\), so the pair is not uncorrelated.

At lag 3, \(X_{t+3}=5+\varepsilon_{t+3}+0.5\varepsilon_{t+2}-0.25\varepsilon_{t+1}\). This expression contains none of \(\varepsilon_t,\varepsilon_{t-1},\varepsilon_{t-2}\), the innovations used by \(X_t\). The two windows are disjoint, so \(\operatorname{Cov}(X_t,X_{t+3})=0\).

Check. The MA order is \(q=2\). Lag 2 is not greater than \(q\), and the calculation finds nonzero covariance. Lag 3 is greater than \(q\), and the calculation gives zero, as the cutoff result requires.

Common mistake avoided. Applying the cutoff at lag \(q\) rather than at lags strictly greater than \(q\) would incorrectly classify the lag-2 pair as uncorrelated.

Uncorrelated Does Not Always Mean Independent

The words “finite dependence” need care. Under the white-noise assumptions alone, the conclusion established above is that observations separated by more than \(q\) are uncorrelated. Zero covariance does not generally imply independence.

If the innovations at distinct times are independent, a stronger conclusion follows. For \(h>q\), the two observations are functions of disjoint sets of independent innovations. Those sets are independent, so the observations are independent. In particular, independent identically distributed innovations provide the needed independence assumption. Normality is not required for this disjoint-set argument.

The distinction matters in applications. Covariance describes linear association, while independence rules out all forms of statistical dependence. A calculation showing zero covariance, on its own, supports the first statement only. Conversely, an MA model with independent innovations has no dependence between observations whose innovation windows are disjoint.

Worked Example: Shorter-Lag Covariance Can Cancel

Consider the hypothetical process \(X_t=3+\varepsilon_t-0.5\varepsilon_{t-1}+0.25\varepsilon_{t-2}+0.5\varepsilon_{t-3}\), where the innovations are white noise with variance 2. Compare the windows and covariance for \(X_t\) with \(X_{t+2}\), \(X_{t+3}\), and \(X_{t+4}\).

Method. Expand each later observation, retain only innovation indices shared with \(X_t\), and add their covariance contributions. The contribution from a shared innovation is the product of its two coefficients times its variance.

The window for \(X_t\) is \(\{\varepsilon_t,\varepsilon_{t-1},\varepsilon_{t-2},\varepsilon_{t-3}\}\), with respective coefficients \(1,-0.5,0.25,0.5\). For lag 2, \(X_{t+2}=3+\varepsilon_{t+2}-0.5\varepsilon_{t+1}+0.25\varepsilon_t+0.5\varepsilon_{t-1}\). The shared terms are \(\varepsilon_t\) and \(\varepsilon_{t-1}\). Their contributions are \(1(0.25)(2)=0.5\) and \((-0.5)(0.5)(2)=-0.5\). Thus \(\operatorname{Cov}(X_t,X_{t+2})=0.5-0.5=0\).

This zero covariance comes from cancellation, not from disjoint windows. For lag 3, \(X_{t+3}=3+\varepsilon_{t+3}-0.5\varepsilon_{t+2}+0.25\varepsilon_{t+1}+0.5\varepsilon_t\). The only shared term is \(\varepsilon_t\), giving covariance \(1(0.5)(2)=1\). For lag 4, \(X_{t+4}=3+\varepsilon_{t+4}-0.5\varepsilon_{t+3}+0.25\varepsilon_{t+2}+0.5\varepsilon_{t+1}\). There is no shared innovation, so the covariance is 0.

Check. The order is \(q=3\). Lag 4 is greater than \(q\), and disjoint windows guarantee zero covariance. Lag 2 is within the window range but happens to have zero covariance because the two shared contributions cancel. Lag 3 has nonzero covariance. These results are consistent: the cutoff guarantees zero beyond \(q\), but shorter-lag values depend on the coefficients.

Common mistake avoided. Seeing zero covariance at lag 2 does not justify concluding that the windows are disjoint or that every lag of at least 2 has zero covariance.

How to Apply the Finite-Window Argument

When a problem asks whether observations in an MA(\(q\)) process are related, begin with the model order and the actual innovation indices rather than relying on a vague memory of a cutoff. The following procedure keeps the time indexing explicit.

Write the model through lag \(q\).
Record the current innovation and every lagged innovation in the stated process. Keep coefficients attached to their corresponding lags.

Expand both observations.
For a pair \(X_t\) and \(X_{t+h}\), write the innovation indices in each expression. This avoids confusing a lag in the observation with a lag in the innovation.

Check for overlap.
If \(|h|>q\), the windows are disjoint and the observations are uncorrelated under white-noise assumptions. At \(|h|\le q\), overlap is possible, so a zero-covariance conclusion needs further reasoning.

State only what the assumptions support.
Uncorrelated innovations justify a zero-covariance result. To claim independence of observations from disjoint windows, assume independent innovations.

A useful interpretation is that an MA(\(q\)) observation has a limited innovation history: its current value is affected directly by at most \(q\) earlier shocks. Two observations more than \(q\) steps apart cannot share any of those shocks. This is why the finite cutoff arises from the model specification rather than from a numerical approximation.

Common Exam Traps

Using the wrong cutoff inequality. The guaranteed result is for \(|h|>q\), not \(|h|\ge q\). At lag \(q\), the windows can share one innovation.
Assuming every shorter lag is correlated. Overlap permits nonzero covariance but does not require it. Contributions can cancel, as in the MA(3) example.
Confusing innovations with observations. An MA model combines \(\varepsilon\) terms, not lagged \(X\) terms. Substituting past observations changes the model.
Equating zero covariance with independence. White noise specifies uncorrelated innovations, not necessarily independent ones. Make the stronger claim only with an independence assumption.
Dropping the mean without explanation. The constant mean does not affect covariance, but when evaluating an observation it must still be included.

Key takeaway: An MA(\(q\)) observation uses innovations from time \(t\) back through \(t-q\). If two observations are more than \(q\) time steps apart, these finite windows do not overlap, so the observations are uncorrelated under white-noise assumptions. Shorter lags require coefficient-specific reasoning.

Check Your Understanding

For each problem, identify the innovation windows before deciding what follows from the assumptions.

A hypothetical monthly severity index follows \(X_t=8+\varepsilon_t+0.3\varepsilon_{t-1}-0.2\varepsilon_{t-2}\). List the innovations in \(X_{t+1}\). Does the cutoff result alone establish that \(X_t\) and \(X_{t+1}\) are uncorrelated? (Answer: \(X_{t+1}\) uses \(\varepsilon_{t+1},\varepsilon_t,\varepsilon_{t-1}\); no, since lag 1 is not greater than \(q=2\).)
An MA(1) process has white-noise innovations. Are \(X_t\) and \(X_{t+2}\) guaranteed to be uncorrelated? State the index-window reason. (Answer: Yes; their windows are \(\{t,t-1\}\) and \(\{t+2,t+1\}\), which are disjoint, and \(2>1\).)
Let \(X_t=\varepsilon_t+0.4\varepsilon_{t-1}-0.1\varepsilon_{t-2}\), with white-noise variance 5. Identify the shared innovation and its covariance contribution for \(X_t\) and \(X_{t+2}\). Does the cutoff guarantee zero covariance? (Answer: The shared innovation is \(\varepsilon_t\), with contribution \(1(-0.1)(5)=-0.5\); no, lag 2 equals \(q\), not a lag greater than \(q\).)
An MA(2) process has independent innovations. Are \(X_t\) and \(X_{t+3}\) merely uncorrelated, or independent as well? Explain. (Answer: They are independent as well as uncorrelated, because their innovation windows are disjoint and the innovations are independent.)
An MA(3) process has white-noise innovations, and a calculation gives \(\operatorname{Cov}(X_t,X_{t+2})=0\). Can one conclude that the two observations use disjoint innovation sets? (Answer: No; lag 2 allows overlapping windows, and zero covariance may result from cancellation among shared-innovation contributions.)

What You'll Learn