Why Mean, ACF, and PACF Patterns Matter
A stationary time series can have a stable mean and variance while still exhibiting dependence across time. The mean describes its level; the autocorrelation function (ACF) describes its linear dependence at each lag. The partial autocorrelation function (PACF) helps distinguish direct lag relationships from dependence that is transmitted through intervening observations. Together, their characteristic patterns provide a useful first guide to whether an autoregressive (AR), moving-average (MA), or mixed ARMA model is plausible.
These patterns are theoretical properties of a process, not rigid visual rules for a finite sample. Sample ACF and PACF values vary, and different models can produce similar-looking plots. The goal is to understand what a pattern suggests and what it does not establish.
Definitions and Model Convention
Let \(X_t\) be a covariance-stationary process with finite mean \(\mu=E[X_t]\). Write \(Z_t=X_t-\mu\), and define its autocovariance at lag \(k\) as \(\gamma_k=\operatorname{Cov}(Z_t,Z_{t-k})\). Its ACF is the autocovariance scaled by the variance.
For a zero-mean process, use the lag-polynomial convention \(\phi(L)Z_t=\theta(L)\varepsilon_t\), where \(\phi(L)=1-\phi_1L-\cdots-\phi_pL^p\) and \(\theta(L)=1+\theta_1L+\cdots+\theta_qL^q\). Here \(LZ_t=Z_{t-1}\), and \(\varepsilon_t\) is a zero-mean white-noise innovation with variance \(\sigma^2\). For a nonzero-mean process, this equation applies to \(Z_t=X_t-\mu\); if instead the model is written with a constant intercept \(c\) on the right side of the AR equation, then \(\mu=c/(1-\sum_{i=1}^p\phi_i)\), provided the denominator is nonzero.
The patterns below assume a stationary causal process; for statements about the invertible MA representation and its PACF, invertibility is also assumed. In the root convention used here, causality and invertibility require the roots of the respective AR and MA polynomials to lie outside the unit circle. White noise need not be normally distributed for the covariance patterns described here.
Pure AR Processes: ACF Tails and PACF Cutoffs
For a stationary AR(\(p\)) process, the current value is a linear combination of the previous \(p\) values plus a new innovation. Its ACF generally continues beyond lag \(p\), but decays rather than cutting off. The exact decay may be monotone, alternate in sign, or oscillate with decreasing magnitude; it is governed by the AR roots. Its PACF, in contrast, is zero at lags greater than \(p\). This cutoff follows because the AR equation already expresses \(Z_t\) using only the most recent \(p\) values and an innovation orthogonal to past values.
Worked Example: Reading a Stationary AR(1) Pattern
Suppose the hypothetical process is \(X_t=10+0.6(X_{t-1}-10)+\varepsilon_t\), where \(\varepsilon_t\) is white noise with variance \(4\). The mean is \(10\), and the AR root solves \(1-0.6z=0\), giving \(z=1/0.6=1.667\), outside the unit circle. Thus the process is stationary and causal.
For an AR(1), the ACF is \(\rho_k=\phi_1^k\) for \(k\geq 0\). Substituting \(\phi_1=0.6\) gives \(\rho_1=0.6\), \(\rho_2=0.6^2=0.36\), and \(\rho_3=0.6^3=0.216\). These correlations decline geometrically rather than becoming exactly zero. The PACF is \(0.6\) at lag 1 and zero at every lag greater than 1. The innovation variance affects the process variance, but not these correlations.
Pure MA Processes: ACF Cutoffs
A pure MA(\(q\)) process is a finite weighted sum of the current and previous \(q\) innovations. At a lag greater than \(q\), the two observations in a covariance share no innovations, so their covariance is zero. Consequently, the ACF cuts off after lag \(q\). The PACF generally does not cut off there: under invertibility it typically tails off, often at a geometric or damped-oscillatory rate. Thus an MA process reverses the main cutoff pattern of an AR process.
Worked Example: Calculating the ACF of an MA(2)
Consider the hypothetical process \(Z_t=\varepsilon_t+0.5\varepsilon_{t-1}+0.25\varepsilon_{t-2}\), with innovation variance \(\sigma^2=1\). An MA process is stationary for these finite coefficients. Its MA polynomial is \(1+0.5z+0.25z^2\); its roots have modulus \(2\), so it is also invertible.
The variance is the sum of the squared coefficients times the innovation variance: \(\gamma_0=1^2+0.5^2+0.25^2=1+0.25+0.0625=1.3125\). At lag 1, the overlapping innovation terms give \(\gamma_1=(1)(0.5)+(0.5)(0.25)=0.625\), so \(\rho_1=0.625/1.3125=0.4762\). At lag 2, only one term overlaps: \(\gamma_2=(1)(0.25)=0.25\), so \(\rho_2=0.25/1.3125=0.1905\). For \(k>2\), \(\rho_k=0\).
For illustration, the lag-2 PACF is \(\alpha_{22}=(\rho_2-\rho_1^2)/(1-\rho_1^2)\). Substitution gives \((0.1905-0.4762^2)/(1-0.4762^2)\approx-0.0469\), not zero. The MA order is identified by the theoretical ACF cutoff, not a PACF cutoff.
Mixed ARMA Processes: Both Functions Usually Tail Off
In a mixed ARMA(\(p,q\)) process, autoregressive and moving-average components operate together. In a minimal, nondegenerate model, neither the ACF nor the PACF usually cuts off at a finite lag. Their tails decay under the causal and invertible conditions, with rates and possible oscillations determined by the relevant roots. An exact cutoff or a particularly short tail can occur in special cases, including parameter cancellations, so the word “usually” matters.
The AR part imposes a recurrence on autocovariances once the lags are beyond the range directly affected by the finite MA polynomial. This recurrence produces a decaying tail rather than the finite overlap cutoff of a pure MA. Under invertibility, the process also has an infinite autoregressive representation, which explains why its PACF generally has a tail instead of the exact finite cutoff of a pure AR.
Worked Example: A Mixed ARMA(1,1) Tail
Suppose \(Z_t=0.5Z_{t-1}+\varepsilon_t+0.4\varepsilon_{t-1}\), with \(\sigma^2=1\). The AR root is \(2\), and the MA root solves \(1+0.4z=0\), giving \(-2.5\). Both roots are outside the unit circle, so the model is causal and invertible.
For this ARMA(1,1), \(\gamma_0=\sigma^2(1+\theta_1^2+2\phi_1\theta_1)/(1-\phi_1^2)\). Substituting gives \(\gamma_0=(1+0.4^2+2(0.5)(0.4))/(1-0.5^2)=(1+0.16+0.40)/0.75=2.08\). At lag 1, \(\gamma_1=\phi_1\gamma_0+\theta_1\sigma^2=(0.5)(2.08)+0.4=1.44\), so \(\rho_1=1.44/2.08=0.6923\).
For \(k\geq2\), the autocovariances satisfy \(\gamma_k=0.5\gamma_{k-1}\), and therefore \(\rho_2=0.5(0.6923)=0.3462\), \(\rho_3=0.5(0.3462)=0.1731\). This ACF does not cut off at lag 1. The lag-2 PACF is \((\rho_2-\rho_1^2)/(1-\rho_1^2)\approx(0.3462-0.6923^2)/(1-0.6923^2)=-0.2556\), also not zero. The calculations illustrate the general mixed-model pattern: both functions can display tails.
Using Sample Patterns to Propose Orders
For observed data, the sample ACF and PACF are estimates of population quantities. A pronounced sample cutoff is not proof that the corresponding population correlations are exactly zero: sampling variation can make small values appear large or large values appear small. Use the patterns as model clues, then assess candidate models with appropriate estimation and diagnostic methods.
| Candidate process | Typical theoretical ACF | Typical theoretical PACF | Initial clue |
|---|---|---|---|
| AR(\(p\)) | Tails off | Cuts off after lag \(p\) | PACF has a plausible cutoff |
| MA(\(q\)) | Cuts off after lag \(q\) | Tails off | ACF has a plausible cutoff |
| ARMA(\(p,q\)) | Usually tails off | Usually tails off | Neither function has a clear cutoff |
Worked Example: Suggesting an Initial Model From Plot Behavior
Imagine a hypothetical analyst examining a stationary series whose sample ACF has prominent values at lags 1 and 2, followed by values close to zero, while the sample PACF declines gradually across several lags. This pattern is consistent with an MA(2), because its theoretical ACF cuts off after lag 2 while its PACF generally tails off.
The analyst should describe MA(2) as a plausible candidate, not as a proven data-generating process. The assessment depends on sample size, estimation variability, and whether the apparent cutoff is stable across reasonable specifications. A mixed ARMA model or another process might also produce similar finite-sample plots. The initial pattern is a way to narrow possibilities, not a substitute for fitting and checking a model.
Common Mistakes and Practical Interpretation
- Calling every declining ACF an AR model. A declining ACF alone does not distinguish an AR process from an ARMA process. Check whether the PACF shows a plausible cutoff and consider other evidence.
- Treating a sample value near zero as an exact cutoff. Sample correlations fluctuate. The theoretical cutoff is an exact property of a model; a plotted sample pattern is only evidence about that property.
- Assuming the PACF must cut off for an MA model. The usual pure-MA signature is the ACF cutoff. Under invertibility, the PACF generally tails off.
- Ignoring the mean or stationarity conditions. ACF and PACF patterns describe dependence around a stable mean. Trend, changing variance, or other nonstationarity can make those patterns misleading.
Check Your Understanding
Answer in terms of theoretical process patterns, and distinguish them from what a finite sample can establish.
- What is the defining difference between the ACF at lag \(k\) and the PACF at lag \(k\)?
- For a stationary AR(3), which function has a theoretical cutoff, and after which lag?
- Why is the ACF of an MA(2) exactly zero at lags greater than 2?
- A sample ACF appears to cut off after lag 1, while the sample PACF tails off. What model might this suggest, and why is that not a definitive conclusion?
- What do causality and invertibility contribute to the usual interpretation of ARMA ACF and PACF tails?