Why Read a Residual ACF?
After fitting a time-series model, the residuals are the differences between observed values and the model’s fitted values. A residual ACF plot displays the sample autocorrelations of those residuals across a range of lags. It is a visual diagnostic: it helps identify whether the fitted residuals still show serial dependence that the model has not accounted for.
A residual ACF is not a verdict based on whether any single bar crosses a line. The familiar approximate bands are usually pointwise: they give a reference for one lag at a time under a white-noise approximation. A plot may show many lags, so even if the model is adequate, some bars can cross the bands by chance. Reading the plot requires both checking individual lags and considering the pattern across all the lags inspected.
The plotted values summarize association, not causation, and a bar’s height is not itself a probability. Positive autocorrelation means residuals separated by that lag tend to have deviations in the same direction; negative autocorrelation means they tend to have deviations in opposite directions. The earlier material on residual white noise explains why serial independence, or at least the absence of serial correlation, is a desirable model-adequacy condition. The ACF plot is one way to screen for departures from that condition.
Approximate Pointwise Reference Bands
Suppose, as a reference model, that the residuals are independent, have mean zero, and have a common finite variance. For a fixed positive lag \(k\), and a sufficiently long series, the sample autocorrelation is approximately centered at zero with standard error \(1/\sqrt{n}\). Thus a commonly used approximate 95% pointwise reference band is:
These are approximate limits for the null reference, not exact bounds that must contain every sample autocorrelation. Under the reference assumptions, about 5% of individual lag statistics would be expected to fall outside the corresponding two-sided limits over repeated samples. A bar beyond a band is therefore a signal to investigate, not proof by itself that a model is wrong.
Use the number of residual observations underlying the plotted ACF, not automatically the original series length if some observations were lost to differencing, initialization, or missing data.
Check whether the plot uses approximate bands such as \( \pm 1.96/\sqrt{n} \), a different approximation, or adjusted limits. Do not assume all software uses identical formulas.
Note which bars cross the reference limits, their signs, and whether the exceedances are isolated or appear in a repeated pattern.
Because many pointwise comparisons are being viewed at once, treat a small number of borderline exceedances cautiously. A later formal diagnostic can assess a group of autocorrelations jointly.
Why the Standard Error Is Approximately \(1/\sqrt{n}\)
The approximation can be understood from the sample autocorrelation formula. To see the main idea, suppose the observations are independent white noise with mean zero and variance \(\sigma^2\), and consider a fixed lag \(k\) as \(n\) grows. Ignoring the small effect of estimating the mean for this argument, the numerator is approximately \(\sum_{t=k+1}^{n} e_t e_{t-k}\). Each product has expectation zero. Under independence, its variance is \(\sigma^4\), and the sum has variance approximately \(n\sigma^4\). The denominator is approximately \(n\sigma^2\). Dividing the numerator’s standard deviation by that denominator gives:
For a fixed lag and large \(n\), a normal approximation then motivates the \(1.96\) multiplier for a two-sided 95% pointwise reference. This argument is an approximation, not a guarantee of exact coverage for a short series. It also treats the residuals like independent observations. In practice, residuals come from a fitted model, whose parameters were estimated from the same data; estimation can affect the sampling behavior of their autocorrelations. Consequently, the simple bands are useful visual guides, but should not be presented as exact inference for every fitted model.
Worked Examples: From Individual Bars to a Whole Plot
Worked Example: Checking Several Bars Against Pointwise Bands
A model has been fitted to a hypothetical monthly claim-count series. The residual ACF plot is based on \(n=64\) residuals and shows \(r_1=0.08\), \(r_2=0.27\), \(r_3=-0.12\), \(r_4=0.23\), and \(r_5=-0.31\). Using approximate 95% pointwise bands, identify the exceedances and describe what they do and do not establish.
Method. Use \( \pm 1.96/\sqrt{n} \) for each lag, then compare each displayed autocorrelation with those limits.
Step 1: Calculate the standard error. Since \( \sqrt{64}=8 \),
Step 2: Calculate the pointwise limits.
Step 3: Compare the bars. At lag 1, \(0.08\) is inside the band. At lag 2, \(0.27>0.245\), so the bar exceeds the upper limit. At lag 3, \(-0.12\) is inside the band. At lag 4, \(0.23\) is inside the band. At lag 5, \(-0.31<-0.245\), so the bar exceeds the lower limit.
Conclusion. Lags 2 and 5 exceed the approximate pointwise limits. This merits attention, but the plot inspected at least five lags, so the exceedances alone do not prove that the model has unaccounted serial dependence. The correct conclusion is a visual warning that should be assessed in context.
Common mistake avoided: Calling the model inadequate solely because one or more bars cross pointwise limits ignores that the limits apply separately at each lag.
Worked Example: Allowing for Many Inspected Lags
A hypothetical quarterly expense model produces \(n=144\) residuals. Its ACF plot displays the first \(m=20\) positive lags. The largest positive bar is \(r_5=0.18\), and the most negative is \(r_{12}=-0.27\). Calculate the ordinary pointwise limit and a rough Bonferroni-adjusted reference limit. Explain which bars cross each limit.
Step 1: Calculate the ordinary pointwise limit. Since \( \sqrt{144}=12 \),
Thus the ordinary approximate pointwise band is \([-0.1633,0.1633]\). The lag-5 value \(0.18\) is above \(0.1633\), and the lag-12 value \(-0.27\) is below \(-0.1633\). Both cross the ordinary pointwise limits.
Step 2: Construct a rough family-wise reference. To make the probability of at least one false rejection across \(m=20\) separate comparisons no greater than approximately \(0.05\), a Bonferroni calculation allocates \(0.05/20=0.0025\) to each two-sided comparison. The normal critical value is approximately \(z_{0.99875}=3.023\). The resulting rough limit is:
The lag-5 value \(0.18\) does not exceed \(0.2519\). The lag-12 value does, because \(|-0.27|=0.27>0.2519\). This adjustment is conservative and is a rough visual comparison, not a replacement for a model-specific diagnostic. ACF values at different lags are not generally independent.
Step 3: See why multiplicity matters. If 20 null comparisons were independent and each had a 5% false-alarm probability, the chance of at least one false alarm would be:
Conclusion. Two bars cross the ordinary pointwise bands, but only the lag-12 bar crosses this rough Bonferroni limit. The plot calls for further assessment, not a mechanical decision based on a count of crossings.
Common mistake avoided: Treating 20 pointwise 95% bands as though they jointly provide 95% protection against any false alarm is incorrect.
Worked Example: Interpreting Repeated Spikes
A hypothetical insurer models a weekly operational series. The residual ACF uses \(n=100\) residuals, and the first eight values are \(0.04,-0.07,0.03,0.24,0.01,-0.05,0.06,0.22\). The plot displays pointwise 95% bands. Identify the bars outside the bands and explain why their arrangement deserves attention.
Step 1: Calculate the pointwise limits. Since \( \sqrt{100}=10 \),
Step 2: Compare all eight values. The first three values, \(0.04\), \(-0.07\), and \(0.03\), lie inside the band. The lag-4 value \(0.24\) exceeds \(0.196\). The lag-5, lag-6, and lag-7 values \(0.01\), \(-0.05\), and \(0.06\) lie inside. The lag-8 value \(0.22\) also exceeds \(0.196\).
Step 3: Interpret the arrangement. The two exceedances are positive and separated by four lags. One possible reading is that dependence at a recurring interval has not been fully captured. This is a more structured visual signal than a single isolated bar, but it is still not proof: several lags were inspected, and the plotted values are approximate diagnostics.
Conclusion. Lags 4 and 8 cross the pointwise limits, and their repeated spacing is worth investigating in the fitted model and the data’s time structure. Do not infer a definitive cause from the ACF alone.
Common mistake avoided: Noting only whether bars cross the line misses potentially informative organization among the exceedances.
Practical Limits and Common Traps
Residual ACF plots are most informative when the analyst reports the basis for the comparison: the number of residuals, the band convention, and the lags inspected. Software may display a pointwise approximation, use a different standard-error calculation, or provide an adjustment. If residuals were lost at the start of a series, the number actually used in the ACF can differ from the number of original observations. Use the plotted sample size or the software’s stated convention rather than silently substituting a convenient count.
A frequent arithmetic error is to use \(1/\sqrt{n}\) itself as the 95% limit. For \(n=100\), that gives \(0.10\), but it is the approximate standard error, not the two-sided 95% reference limit. The pointwise limit is \(1.96(0.10)=0.196\). Using \(0.10\) as the band would incorrectly flag many bars.
A second trap is treating every crossing as equally persuasive. A single small exceedance among many displayed lags may be a chance result; repeated, substantial, or organized spikes warrant closer attention. Conversely, no crossing does not prove that the residuals are independent: the approximation has finite-sample limitations, and a plot only displays selected lags. The appropriate conclusion is calibrated to what the plot actually shows.
Check Your Understanding
For each problem, use the stated approximation and distinguish an individual-lag comparison from a conclusion about the full plot.
- A fitted model has 81 residuals. What are the approximate 95% pointwise limits, rounded to four decimals? (Answer: \([-0.2178,\ 0.2178]\))
- A residual ACF is based on 64 observations. At lag 3 the plotted value is \(-0.26\). Does it cross the approximate 95% pointwise band, and what is the band rounded to three decimals? (Answer: Yes; the band is \([-0.245,\ 0.245]\).)
- A plot displays 12 lags, each compared with a 5% two-sided pointwise test. Under the simplifying assumption that the comparisons are independent, what is the probability of at least one false alarm, rounded to four decimals? (Answer: \(1-0.95^{12}=0.4596\).)
- A model has 100 residuals and displays 10 lags. A bar at lag 6 is \(0.20\), while the largest absolute value at the other displayed lags is \(0.17\). Does the lag-6 bar cross the ordinary approximate 95% pointwise limit, and does it exceed a rough Bonferroni limit using \(z_{0.9975}\approx2.807\)? (Answer: The pointwise limit is \(\pm0.196\), so \(0.20\) crosses it; the Bonferroni limit is \(\pm0.2807\), so it does not cross that limit.)
- A plot based on 121 residuals shows \(r_2=0.05\), \(r_6=-0.19\), and \(r_{10}=-0.23\). Using approximate 95% pointwise bands, which of these three bars cross, and what feature of the values should be noted without claiming a definitive cause? (Answer: The band is approximately \(\pm0.1782\); lags 6 and 10 cross, and both exceedances are negative and four lags apart.)