Recognizing MA Orders From ACF And PACF

Why MA Order Recognition Requires Both Theory and Caution

For a moving-average process, the theoretical autocorrelation function (ACF) has a distinctive feature: it becomes exactly zero after a finite lag. That makes the ACF a useful guide to the order. In observed data, however, sample autocorrelations fluctuate, so a correlogram rarely shows a perfectly clean cutoff. The partial autocorrelation function (PACF) offers another clue, but for an MA process it usually tails off rather than providing a sharp cutoff.

The practical task is therefore to separate the population pattern from its noisy sample estimate. A candidate order should fit the main ACF pattern, while isolated sample spikes should be treated as uncertain evidence rather than automatic proof of a higher order. This identification skill is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Definition: A finite MA process of order $q$ is $$ X_t=\mu+\sum_{j=0}^{q}\theta_j\varepsilon_{t-j}, \qquad \theta_0=1, $$ where the innovations $\varepsilon_t$ have mean zero, variance $\sigma_\varepsilon^2>0$, and are uncorrelated across time. The order $q$ is the largest lag with a nonzero coefficient: $\theta_q\ne0$. Trailing zero coefficients are excluded when stating the order. The strictly positive innovation variance ensures that the process has positive variance and that its ACF is defined.

Why the Theoretical ACF Cuts Off at the MA Order

Write $\gamma_k=\operatorname{Cov}(X_t,X_{t-k})$ and $\rho_k=\gamma_k/\gamma_0$. The two process values share innovations only when their finite sets of innovation times overlap. At lag $k$, this gives

\gamma_k= \begin{cases} \sigma_\varepsilon^2\displaystyle\sum_{j=0}^{q-k}\theta_j\theta_{j+k}, & 0\le k\le q,\\[6pt] 0, & k>q. \end{cases}

To see why, expand the covariance of $X_t$ and $X_{t-k}$. A term involving $\varepsilon_{t-j}$ in the first value can have nonzero covariance with a term in the second only if the innovation time is the same. In $X_{t-k}$, that matching term has index $j+k$, so the matching coefficient products are $\theta_j\theta_{j+k}$. If $k>q$, no indices can match and the covariance is zero.

At lag zero, $\gamma_0=\sigma_\varepsilon^2\sum_{j=0}^{q}\theta_j^2>0$, because $\sigma_\varepsilon^2>0$ and $\theta_0=1$. At the final lag, the formula reduces to $\gamma_q=\sigma_\varepsilon^2\theta_q$, which is nonzero because both factors are nonzero. Dividing by $\gamma_0$ is therefore valid. This proves that the theoretical ACF is zero for all $k>q$, but not at lag $q$: its cutoff is exactly $q$. Some intermediate autocovariances can be zero without changing that final cutoff.

Theorem: For a nondegenerate MA process of order $q$, the theoretical ACF is zero at every lag greater than $q$, and is nonzero at lag $q$. Its theoretical cutoff therefore identifies $q$. The PACF of an MA process typically tails off rather than cutting off at $q$; it is not generally correct to identify the MA order by looking for a PACF cutoff.

The result does not require an invertibility assumption: it follows from the finite innovation representation and the positive innovation variance. Invertibility matters for other questions, including whether the representation is recoverable in the usual way from the observed process. For order recognition, the essential theoretical clue here is the finite ACF cutoff.

Worked Example: Finding the Exact Cutoff for an MA(2)

Suppose $X_t=\mu+\varepsilon_t+0.5\varepsilon_{t-1}+0.25\varepsilon_{t-2}$, where the innovations are uncorrelated, have mean zero, and have variance $\sigma_\varepsilon^2=4$. Find the autocovariances and ACF through lag 3, then identify the order.

Method. Use the MA autocovariance formula, retaining only innovation terms shared by the two observations.

At lag zero, the coefficient squares sum to $1+0.5^2+0.25^2=1+0.25+0.0625=1.3125$. Thus $\gamma_0=4(1.3125)=5.25$.

At lag one, the matching coefficient products are $(1)(0.5)$ and $(0.5)(0.25)$. Therefore $\gamma_1=4(0.5+0.125)=2.5$.

At lag two, the only matching product is $(1)(0.25)$, so $\gamma_2=4(0.25)=1$. At lag three, there are no shared innovations, so $\gamma_3=0$.

Divide each autocovariance by $\gamma_0=5.25$: $\rho_1=2.5/5.25=10/21\approx0.4762$, $\rho_2=1/5.25=4/21\approx0.1905$, and $\rho_3=0$. The ACF is zero at every later lag as well. The final nonzero coefficient is $\theta_2=0.25$, so the model is MA(2) and its theoretical cutoff is exactly lag 2.

Check and common mistake avoided. The lag-2 autocovariance is $\sigma_\varepsilon^2\theta_2=4(0.25)=1$, matching the full calculation. Do not call the process MA(1) just because lag 3 is zero: the cutoff is the last nonzero lag, not the first zero lag.

What the PACF Contributes

The PACF at lag $k$ is the correlation between $X_t$ and $X_{t-k}$ after accounting linearly for the intervening values $X_{t-1},\ldots,X_{t-k+1}$. Equivalently, for a stationary process with a nonsingular covariance matrix, it is the coefficient on $X_{t-k}$ in the linear regression of $X_t$ on $X_{t-1},\ldots,X_{t-k}$. This is a different quantity from the ACF: it measures the remaining linear association after intermediate lags are controlled for.

For an MA process, the PACF generally does not become exactly zero after a fixed lag. Instead, its values typically diminish with lag, though the shape and rate depend on the coefficients. “Typically decays” is a pattern to use when reading a correlogram, not a guarantee that every successive value is smaller or that the PACF must follow a smooth curve.

Worked Example: Why an MA(1) Can Have a Nonzero PACF at Lag 2

Let $X_t=\mu+\varepsilon_t-0.6\varepsilon_{t-1}$, with uncorrelated, mean-zero innovations of variance 9. Find the theoretical ACF at lags 1 and 2, then calculate the PACF at lag 2. State what the two patterns imply.

Method. First calculate the autocovariances and ACF. For lag 2, use the two-lag partial autocorrelation formula $\phi_{22}=(\rho_2-\rho_1^2)/(1-\rho_1^2)$.

The coefficient squares at lag zero sum to $1+(-0.6)^2=1.36$, so $\gamma_0=9(1.36)=12.24$. At lag one, the only matching product is $(1)(-0.6)$, giving $\gamma_1=9(-0.6)=-5.4$. There is no shared innovation at lag two, so $\gamma_2=0$.

Consequently, $\rho_1=-5.4/12.24=-15/34\approx-0.4412$, and $\rho_2=0$. The theoretical ACF is zero for every lag greater than 1, so the order is 1.

For the PACF at lag 2, square $\rho_1$: $\rho_1^2=225/1156$. Then $\phi_{22}=(0-225/1156)/(1-225/1156)=-225/931\approx-0.2417$. Thus the PACF is not zero at lag 2 even though the ACF has already cut off. A nonzero PACF at this lag does not contradict an MA(1) model.

Check and common mistake avoided. The computed lag-2 PACF is nonzero, while the lag-2 ACF is zero. Using a PACF cutoff rule to assign this process order 2 would confuse the two functions. For an MA candidate, look primarily for the ACF cutoff and use the PACF’s broader shape as supporting evidence.

Reading a Sample Correlogram

A sample ACF is calculated from a finite observed series, so its values need not equal the theoretical autocorrelations. Even if the true autocorrelation at a lag is zero, its sample estimate will usually not be exactly zero. A common rough screening guide for a series of length $n$ is the pair of bounds

\pm\frac{1.96}{\sqrt{n}}.

These are approximate 95% reference bounds for individual sample autocorrelations under a white-noise null setting. They are a guide, not a definitive order-selection test. Checking many lags makes isolated exceedances more likely; dependence, estimation of a mean or other model parameters, and other departures from the simple reference setting can also affect the interpretation.

Worked Example: Proposing an Order from a Noisy ACF

A hypothetical monthly series has $n=144$ observations. Its sample ACF at lags 1 through 6 is $0.36,\ 0.22,\ 0.08,\ -0.10,\ 0.17,\ 0.03$, and its sample PACF is $0.35,\ 0.12,\ 0.04,\ -0.07,\ 0.11,\ 0.02$. Use the approximate bounds to propose an MA order, and explain how strongly the correlogram supports it.

Method. Calculate the reference bounds, identify the prominent early ACF values, and then check whether later values provide a convincing alternative cutoff.

Since $\sqrt{144}=12$, the approximate bound is $1.96/12=0.1633$. The lag-1 ACF $0.36$ and lag-2 ACF $0.22$ exceed the bound. The lag-3 value $0.08$, lag-4 value $-0.10$, and lag-6 value $0.03$ do not. The lag-5 value $0.17$ exceeds the bound only slightly.

The PACF has a large first value and then smaller values through the displayed lags; it does not show a clear sharp cutoff. A reasonable candidate is MA(2): the ACF’s two prominent early values suggest a cutoff after lag 2, and the later values are mostly small. But the lag-5 sample spike makes this evidence imperfect. It is not sound to conclude that the population process must have order 5 from one marginal exceedance.

Check and common mistake avoided. The threshold is approximately $0.1633$, so $0.17$ is indeed just outside it. Retain that observation as uncertainty, but do not let one isolated, barely significant sample autocorrelation outweigh the overall pattern. The correlogram proposes an order; it does not prove one.

Start with the ACF.
Look for the last prominent early lag followed by values that are mostly small. For an MA(q), the theoretical cutoff is after lag $q$.

Use the PACF as a pattern check.
A PACF that generally tails off is consistent with an MA description, but it does not identify $q$ through a fixed cutoff.

Account for sampling variation.
Compare sample values with approximate reference bounds, while remembering that a marginal or isolated exceedance may occur by chance.

State a candidate, not a certainty.
Describe the order suggested by the full pattern and note any spikes or ambiguity that weaken the conclusion.

Common Interpretation Traps

Stopping at the first zero sample ACF. A theoretical MA(q) ACF is zero beyond $q$, but a sample ACF fluctuates. One small or even zero-looking sample value does not establish a cutoff.
Calling the first zero lag the order. The order is the last lag with a nonzero theoretical autocorrelation. In a nondegenerate MA(q), the lag-$q$ autocovariance is nonzero.
Demanding every intermediate ACF value be nonzero. Intermediate autocovariances may cancel to zero. That does not reduce the order when $\theta_q\ne0$.
Using the PACF cutoff rule for an MA process. An MA PACF generally tails off; it need not cut off at the MA order.
Treating approximate bounds as a pass-fail rule. Bounds help screen sample evidence. They do not remove sampling variation, multiple-lag issues, or the need to judge the overall correlogram.

Key takeaway: For a nondegenerate MA(q) with no trailing zero coefficients, the theoretical ACF is nonzero at lag $q$ and zero at every lag beyond $q$. The PACF typically tails off. In sample data, use the ACF cutoff as an order clue, the PACF as supporting pattern information, and approximate bounds with appropriate caution.

Check Your Understanding

For each problem, distinguish the theoretical result from the uncertain evidence in a sample correlogram.

A process has $X_t=\mu+\varepsilon_t+0.2\varepsilon_{t-1}-0.1\varepsilon_{t-2}+0.4\varepsilon_{t-3}$, with uncorrelated innovations of variance 2. What is its order, and what are $\gamma_0$ and $\gamma_3$? (Answer: MA(3); $\gamma_0=2.42$ and $\gamma_3=0.80$.)
A series has $n=100$ observations and sample ACF values $0.31,-0.22,0.07,0.12,-0.19$ at lags 1 through 5. Using the approximate 95% bounds, what low-order MA candidate is suggested, and what caution is appropriate? (Answer: The bound is approximately $\pm0.196$; MA(2) is suggested, with the lag-5 value $-0.19$ close to the bound and the conclusion remaining tentative.)
For an MA(2) with $\theta_1=0.5,\theta_2=-0.5$, and innovation variance 1, calculate $\gamma_0,\gamma_1,\gamma_2$. Does the relatively small lag-1 autocovariance imply order 1? (Answer: $\gamma_0=1.50,\ \gamma_1=0.25,\ \gamma_2=-0.50$; no, the order is 2 because $\gamma_2\ne0$.)
A series has $n=225$ observations and sample ACF values $0.42,0.19,0.03,-0.08,0.14$ at lags 1 through 5. What order is a reasonable initial candidate, and how should the lag-5 value affect the conclusion? (Answer: MA(1) is a reasonable initial candidate; the approximate bound is $\pm0.1307$, so the lag-5 value is a slight exceedance and makes the evidence less conclusive rather than establishing MA(5).)
A theoretical correlogram has a nonzero ACF at lag 3, zero ACF at every lag after 3, and a PACF that remains nonzero at some later lags. What order does the ACF indicate, and is the PACF pattern inconsistent with it? (Answer: MA(3); no, a nonzero later PACF is compatible with an MA process because its PACF typically tails off rather than cutting off.)

What You'll Learn