AR And MA Identification With Correlogram Traps

Reading a Correlogram as Evidence

A correlogram can make a time-series model look easier to identify than it really is. In the population, pure autoregressive and pure moving-average processes have characteristic ACF and PACF patterns. In a finite sample, however, estimated correlations fluctuate: a process with zero population autocorrelation can show a few visibly nonzero sample values, and a small model coefficient may leave almost no recognizable pattern.

The practical task is therefore not simply to count spikes. It is to compare the estimated ACF and PACF with the patterns expected under candidate models, judge the size of the apparent signals relative to sampling uncertainty, and state how strongly the evidence supports an order. This is a useful identification step, not a proof that a particular model generated the data.

For a stationary pure AR(\(p\)) process, the population PACF cuts off after lag \(p\), while its ACF generally tails off. For an invertible pure MA(\(q\)) process, the population ACF cuts off after lag \(q\), while its PACF generally tails off. For white noise, both population functions are zero at every positive lag. These are ideal population patterns; sample estimates do not usually reproduce them exactly.

Approximate Bands and Their Limits

Suppose \(X_1,\ldots,X_n\) are observations from a zero-mean white-noise process with constant finite variance and no serial dependence. For moderately large \(n\), the sample autocorrelation at a fixed positive lag is approximately normal with mean zero and variance \(1/n\). This gives the familiar approximate 95% reference bands:

The formula is a large-sample approximation under the white-noise null, not a universal confidence interval for every model or every lag. A bar outside the bands is often called significant in a rough diagnostic sense, but the bands do not prove that the corresponding population correlation is nonzero. Conversely, a bar inside the bands does not prove that the population correlation is exactly zero.

The approximation is useful because under white noise the estimated lag-\(k\) correlation is built from averages of products of observations \(k\) steps apart. Those products have mean zero, and averaging many of them produces sampling variation of order \(1/\sqrt{n}\); the resulting variance is approximately \(1/n\). Standardizing by its approximate standard deviation \(1/\sqrt{n}\) and using the central normal range of about 1.96 standard deviations gives the displayed bands.

When assessing a candidate AR or MA model, do not automatically treat the white-noise bands as exact tests of each model coefficient. They are a useful scale for judging sample ACF fluctuations, especially when asking whether apparent correlations might be noise. Estimated PACF values also fluctuate, but a correlogram’s display bands are approximate. In either case, the model pattern and the size and consistency of the evidence matter.

When a Nominal Order Leaves a Weak Clue

A process can have a specified order while a coefficient associated with its highest lag is so small that the correlogram barely reveals it. The observed pattern can then resemble a lower-order process. This is not a contradiction: order describes the model’s coefficients, while a sample correlogram provides uncertain estimates of correlation.

For an MA(2) process \(X_t=\varepsilon_t+\theta_1\varepsilon_{t-1}+\theta_2\varepsilon_{t-2}\), assume the innovations are uncorrelated with mean zero and variance \(\sigma^2\). Its autocorrelations at lags 1 and 2 are \( \rho_1=\theta_1(1+\theta_2)/(1+\theta_1^2+\theta_2^2) \) and \( \rho_2=\theta_2/(1+\theta_1^2+\theta_2^2) \). In particular, the lag-2 population correlation is small when \(\theta_2\) is small. A weak lag-2 feature can be difficult to distinguish from sampling noise.

White Noise and Isolated Spikes

A white-noise process has zero population ACF and PACF at every positive lag. That does not mean every sample bar will be zero or fall inside approximate bands. If many lags are inspected, occasional bars beyond the bands are unsurprising even when the white-noise model is suitable.

For example, if a test at each lag had exactly a 5% false-positive probability under white noise, then across 10 inspected lags the expected number of false positives would be \(10(0.05)=0.5\). This expectation does not say there will always be half a spike, and the lag estimates are not generally independent. It illustrates why a single isolated exceedance deserves less weight than a sustained, structured pattern.

A Disciplined Identification Process

Use the following sequence to keep a correlogram reading tied to both the model patterns and their limitations.

An important trap is to count all bars outside the bands and equate that count with the model order. A pure AR order is suggested by where the PACF cuts off, not by the number of ACF spikes outside a reference band. Likewise, a pure MA order is suggested by the ACF cutoff. A bar crossing a band is a sample observation; a population cutoff is a property of the model. The two should not be treated as interchangeable.

This distinction matters when comparing plausible specifications. If a coefficient is near zero, a simpler model may summarize the visible pattern adequately, while the generating model may still have a higher nominal order. Conversely, an isolated sample spike can make white noise appear to have structure. Be explicit about the evidence and its uncertainty rather than forcing a precise order from an ambiguous correlogram. This interpretation is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.