Why MA Interpretations Go Wrong
Moving-average models are compact: each observation depends on only a finite number of current and past innovations. That compact form creates several related facts—model order, finite dependence, an autocorrelation function (ACF) cutoff, and invertibility—that are easy to conflate. The key is to ask what each property describes. The order counts innovation lags in the model; the cutoff describes population autocorrelations; invertibility concerns whether the innovations can be recovered from the observed series using a stable infinite series.
This tutorial uses the convention that the coefficients in an MA model enter with plus signs. A different sign convention changes coefficient signs in formulas, so always check the specified model before applying a result. The distinctions here are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
For a weakly stationary process, let \(\gamma_k=\operatorname{Cov}(X_t,X_{t-k})\), and let \(\rho_k=\gamma_k/\gamma_0\) when \(\gamma_0>0\). These are population quantities. In particular, a statement that the ACF “cuts off” after lag \(q\) means \(\rho_k=0\) for every \(k>q\); it does not mean that the estimated ACF in a finite sample will be exactly zero there.
Three Properties That Must Be Kept Separate
| Property | What it describes | MA(\(q\)) implication |
|---|---|---|
| Model order | Which innovation lags have nonzero coefficients | The largest lag with a nonzero coefficient is \(q\) |
| Finite dependence | Which observations can share innovations | Observations more than \(q\) periods apart share none |
| ACF cutoff | Which population autocorrelations can be nonzero | \(\rho_k=0\) for \(k>q\) |
| Invertibility | Whether innovations can be stably represented using present and past observations | Determined by the roots of the MA polynomial, not by the order alone |
Finite dependence and an ACF cutoff are connected but are not identical claims. Even if the innovations are only uncorrelated white noise rather than independent, observations separated by more than \(q\) have zero covariance because they involve disjoint innovation time points. Independence of those observations requires an additional condition, such as independent innovations. Zero covariance alone does not generally imply independence.
Deriving the MA(2) ACF Cutoff
The MA(2) model under the stated convention is \(X_t=\mu+\varepsilon_t+\theta_1\varepsilon_{t-1}+\theta_2\varepsilon_{t-2}\). Subtracting the constant mean does not affect covariances. Because distinct white-noise innovations have zero covariance, a product contributes to \(\gamma_k\) only when the two observations contain the same innovation.
The three innovation terms in one observation are uncorrelated, so their variances add: \(\gamma_0=\sigma^2(1+\theta_1^2+\theta_2^2)\).
\(X_t\) and \(X_{t-1}\) share two innovation terms. Their coefficient products sum to \(\theta_1+\theta_1\theta_2\), giving \(\gamma_1=\sigma^2\theta_1(1+\theta_2)\).
Only \(\varepsilon_{t-2}\) appears in both observations, with coefficient products \(\theta_2\) and 1. Thus \(\gamma_2=\sigma^2\theta_2\).
For \(k>2\), the two observations contain no common innovation time point. Every covariance term is therefore zero, so \(\gamma_k=0\) and \(\rho_k=0\).
This derivation explains why an ACF cutoff is a population result: it follows from which innovations the two observations share. It also shows why the ACF at every lag up to \(q\) need not be nonzero. For example, \(\gamma_1\) can vanish if \(\theta_1=0\) or if \(\theta_2=-1\). A missing spike at an intermediate lag does not, by itself, prove that the model has lower order.
Worked Examples: Checking the Model, ACF, and Roots
Worked Example: A Noninvertible MA(1) Still Has a Lag-1 Cutoff
Suppose a hypothetical monthly claim-count index follows \(X_t=\mu+\varepsilon_t+1.5\varepsilon_{t-1}\), where the innovations are white noise with variance \(\sigma^2=4\). Find the order, autocovariances and ACF cutoff, then check invertibility. The method is to calculate moments from shared innovations and test the polynomial root separately.
Order and finite dependence. The coefficient at lag 1 is \(1.5\ne0\), so this is MA(1). Each observation uses two innovations, \(\varepsilon_t\) and \(\varepsilon_{t-1}\). Observations more than one time period apart share no innovation.
Autocovariances. The variance is \(\gamma_0=4(1+1.5^2)=4(3.25)=13\). At lag 1, the shared innovation has coefficient product \(1.5\), so \(\gamma_1=4(1.5)=6\). At all lags \(k>1\), \(\gamma_k=0\).
ACF. \(\rho_1=6/13=0.461538\ldots\), and \(\rho_k=0\) for \(k>1\). The population ACF therefore cuts off after lag 1.
Invertibility. The MA polynomial is \(1+1.5z\). Its root is \(z=-1/1.5=-2/3\), whose absolute value is less than 1. The process is not invertible under the standard root condition. This does not change its MA(1) order or its lag-1 ACF cutoff.
Check and mistake avoided. The computed variance and lag-1 covariance give \(\rho_1=6/13\), which lies between \(-1\) and 1. A common mistake is to infer that a noninvertible model cannot have a finite ACF cutoff. Here the root test and the covariance calculation show that invertibility and the cutoff are separate properties.
Worked Example: A Zero Lag-1 ACF Does Not Reduce an MA(2) to MA(0)
A hypothetical quarterly operational index is modeled as \(X_t=\mu+\varepsilon_t+0\varepsilon_{t-1}+0.5\varepsilon_{t-2}\), with white-noise innovations of variance \(\sigma^2=9\). Determine its actual order and calculate \(\gamma_0,\gamma_1,\gamma_2,\rho_1,\rho_2\), and the cutoff. The method is to use the nonzero coefficient at the largest lag and then apply the MA(2) moment formulas.
Order. Although the lag-1 coefficient is zero, the lag-2 coefficient is \(0.5\ne0\). The order is 2.
Autocovariances. The variance is \(\gamma_0=9(1+0^2+0.5^2)=9(1.25)=11.25\). At lag 1, \(\gamma_1=9(0)(1+0.5)=0\). At lag 2, \(\gamma_2=9(0.5)=4.5\). At lags greater than 2, there are no shared innovations, so the autocovariances are zero.
Autocorrelations. \(\rho_1=0/11.25=0\). Also, \(\rho_2=4.5/11.25=0.4\). Thus the ACF is zero at lag 1, nonzero at lag 2, and zero at every lag greater than 2.
Check and mistake avoided. Since \(\rho_2=0.4\ne0\), the ACF has not cut off after lag 1. Calling this MA(0) or MA(1) just because \(\rho_1=0\) would mistake an individual autocorrelation for the model order. The result is internally consistent: \(\gamma_2/\gamma_0=4.5/11.25=0.4\).
Worked Example: A Check for Invertibility in an MA(2)
A hypothetical monthly exposure index follows \(X_t=\mu+\varepsilon_t+0.5\varepsilon_{t-1}+0.2\varepsilon_{t-2}\), with white-noise innovations of variance \(\sigma^2=4\). Calculate its first three autocovariances and first two autocorrelations, state its cutoff, and test invertibility. The method is to calculate covariances from shared innovations and solve the MA polynomial roots.
Autocovariances. \(\gamma_0=4(1+0.5^2+0.2^2)=4(1.29)=5.16\). At lag 1, \(\gamma_1=4(0.5)(1+0.2)=4(0.6)=2.4\). At lag 2, \(\gamma_2=4(0.2)=0.8\). For \(k>2\), \(\gamma_k=0\).
Autocorrelations. \(\rho_1=2.4/5.16=20/43\approx0.4651\). Also, \(\rho_2=0.8/5.16=20/129\approx0.1550\). The ACF cuts off after lag 2.
Invertibility. The MA polynomial is \(1+0.5z+0.2z^2\). Its roots solve \(0.2z^2+0.5z+1=0\), so \(z=(-0.5\pm\sqrt{0.25-0.8})/0.4=(-0.5\pm i\sqrt{0.55})/0.4\). The squared modulus of either root is \((0.5^2+0.55)/(0.4^2)=0.8/0.16=5\), so each root has modulus \(\sqrt{5}>1\). The model is invertible.
Check and mistake avoided. Both autocorrelations are nonzero through lag 2, and all later population autocorrelations are zero, as expected for this MA(2). A common mistake is to test invertibility by looking for a nonzero ACF after lag 2. The cutoff follows from finite innovation overlap; invertibility instead follows from the roots, which here lie outside the unit circle.
Invertibility Is Not the Same as Order
For the plus-sign convention, the MA polynomial is \(\theta(z)=1+\theta_1z+\cdots+\theta_qz^q\). An MA process is invertible when every root of \(\theta(z)=0\) lies outside the unit circle. For an MA(1), \(1+\theta_1z=0\) has root \(-1/\theta_1\); when \(\theta_1\ne0\), invertibility therefore requires \(|\theta_1|<1\).
For example, the MA(1) process with coefficient \(1.5\) in the first worked example is noninvertible, but still has \(\rho_k=0\) for \(k>1\). In some cases a noninvertible MA representation can also be rewritten using an invertible representation with adjusted coefficients and innovation variance. That fact does not justify ignoring the given representation’s root condition; identify the stated coefficients first, then apply the test.
Common Exam Traps
- Using an intermediate zero to claim a lower order. In an MA(2), \(\rho_1=0\) can occur even when \(\theta_2\ne0\). Check the largest nonzero model coefficient and the later ACF values.
- Calling sample ACF values exact. The population autocorrelation is zero beyond the MA order, but estimated sample autocorrelations generally fluctuate around zero. A small sample spike does not automatically change the model order.
- Equating zero covariance with independence. White noise guarantees zero covariance across time, not necessarily independence. Independence follows if the innovations are independent and the observations use disjoint innovation sets.
- Using invertibility to determine the ACF cutoff. Root locations determine invertibility. Shared innovations determine the finite covariance range.
- Forgetting the coefficient sign convention. If a model is written with minus signs, translate it to the stated convention before applying formulas; otherwise a calculated autocorrelation may have the wrong sign.
A reliable interpretation proceeds in a fixed order: identify the model convention, find the largest nonzero coefficient lag, calculate the relevant covariances if needed, state the population cutoff, and test invertibility separately. This prevents one observed ACF feature from being asked to answer several different questions.
Check Your Understanding
For each original problem, distinguish model order, population ACF behavior, and invertibility where requested. Answers are given at the end of each item.
- A hypothetical weekly index follows \(X_t=\mu+\varepsilon_t+0.4\varepsilon_{t-1}\), with white-noise innovations of variance 5. Find \(\gamma_0,\gamma_1,\rho_1\), and state the ACF cutoff and invertibility. (Answer: MA(1); \(\gamma_0=5.8,\ \gamma_1=2,\ \rho_1=0.3448\); cutoff after lag 1; invertible since the root is \(-2.5\).)
- A hypothetical quarterly measure follows \(X_t=\mu+\varepsilon_t+0.3\varepsilon_{t-1}+0\varepsilon_{t-2}\). What is its order, and what is the population ACF cutoff? (Answer: MA(1), because the lag-2 coefficient is zero and the lag-1 coefficient is nonzero; the ACF cuts off after lag 1.)
- For \(X_t=\mu+\varepsilon_t+0\varepsilon_{t-1}-0.4\varepsilon_{t-2}\), with innovation variance 2, calculate \(\gamma_0,\gamma_1,\gamma_2\), and determine the order and cutoff. (Answer: MA(2); \(\gamma_0=2.32,\ \gamma_1=0,\ \gamma_2=-0.8\); cutoff after lag 2.)
- A hypothetical MA(1) has coefficient \(\theta_1=-1.25\). Find the MA polynomial root, state whether it is invertible, and say whether that changes the population ACF cutoff. (Answer: root \(z=0.8\), so it is not invertible; it remains MA(1) and its population ACF cuts off after lag 1.)
- For \(X_t=\mu+\varepsilon_t+0.2\varepsilon_{t-1}+0.1\varepsilon_{t-2}\), with innovation variance 3, calculate \(\rho_1\) and \(\rho_2\), give the cutoff, and determine invertibility. (Answer: \(\gamma_0=3.15,\ \gamma_1=0.66,\ \gamma_2=0.30\); \(\rho_1=0.2095,\ \rho_2=0.0952\); cutoff after lag 2; invertible, since the roots of \(1+0.2z+0.1z^2\) have modulus \(\sqrt{10}>1\).)