Why Seasonal Order Identification Requires Care
A seasonal pattern in a time series can persist from one season to the next. For monthly data with annual seasonality, for example, observations twelve months apart may be related; for quarterly data with yearly seasonality, the corresponding separation is four quarters. The ACF and PACF help describe that dependence, but a visible spike at a seasonal lag is evidence to interpret, not an order to read off mechanically.
The goal is to use the pattern at seasonal multiples to propose seasonal autoregressive and moving-average orders, while keeping the ordinary, nonseasonal dependence in view. This identification is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
For order identification, examine the ACF and PACF of a series that is reasonably stationary, or of the appropriately transformed series being modeled. At seasonal multiples, a simple seasonal AR pattern tends to have a PACF cutoff and a gradually declining ACF; a simple seasonal MA pattern tends to have an ACF cutoff and a gradually declining PACF. These signatures are useful guides under simple model assumptions, not guarantees for every observed series.
Reading Seasonal-Lag Patterns
A cutoff means that correlations beyond a particular order are zero in the ideal theoretical model. A tail-off means that correlations tend to decline rather than becoming exactly zero after a fixed lag. In a finite sample, neither pattern will usually appear perfectly. Look for a coherent pattern across several seasonal multiples, not one isolated bar.
| Simple seasonal model | Seasonal-lag ACF | Seasonal-lag PACF | Initial candidate |
|---|---|---|---|
| Seasonal AR of order \(P\) | Typically tails off | Cuts off after seasonal order \(P\) | \(P\) from the PACF cutoff |
| Seasonal MA of order \(Q\) | Cuts off after seasonal order \(Q\) | Typically tails off | \(Q\) from the ACF cutoff |
| Seasonal ARMA | Typically tails off | Typically tails off | Consider mixed candidates |
The table summarizes idealized signatures when the seasonal component is simple and other dependence does not obscure it. It is not a rule that every significant seasonal-lag ACF value implies an MA term, or that every significant seasonal-lag PACF value implies an AR term. Ordinary AR and MA effects can also contribute at seasonal multiples, and a multiplicative seasonal model can produce dependence at nearby cross-lags, such as \(s-1\) and \(s+1\).
Always record the seasonal period before reading a plot. If \(s=12\), the seasonal sequence to inspect is at lags 12, 24, 36, and so on; lag 2 is not a seasonal lag merely because it is greater than 1. The seasonal order counts seasonal steps, not individual observations: a cutoff after lag 24 in a seasonal-lag pattern corresponds to two seasonal steps.
Why the Simple Signatures Arise
The cutoff and tail-off descriptions have a mathematical basis. Consider a seasonal AR(1) model with no ordinary AR or MA component:
where \(\varepsilon_t\) is white noise with mean zero and finite variance. Looking only at observations separated by \(s\) gives an ordinary AR(1)-type relationship within each seasonal subsequence. Its correlation at \(k\) seasonal steps is \(\Phi^k\), so the seasonal-lag ACF declines geometrically. Once the intervening seasonal values are accounted for, only the first seasonal step contributes a direct AR relationship; the seasonal PACF is zero after that step in the theoretical model.
For comparison, consider a seasonal MA(1) model with the plus-sign convention:
Its variance is \((1+\Theta^2)\sigma_\varepsilon^2\), because the two white-noise terms are uncorrelated. The covariance at lag \(s\) is \(\Theta\sigma_\varepsilon^2\), since that is the single shared innovation; at lags \(2s,3s,\ldots\), there is no shared innovation and the covariance is zero. Dividing covariance by variance gives the theoretical seasonal ACF:
This justifies the ACF cutoff for the simple seasonal MA(1). The seasonal PACF need not cut off there. With ordinary terms, mixed seasonal terms, or only a short sample, observed patterns may depart substantially from these ideal cases.
Worked Examples
Worked Example: A Seasonal PACF Cutoff Suggests Seasonal AR(1)
A hypothetical monthly claim-count series has been transformed to a series considered reasonably stationary for modeling. Its length is \(n=100\), and its annual seasonal period is \(s=12\). The seasonal-lag sample ACF values at lags 12, 24, 36, and 48 are \(0.61, 0.37, 0.22, 0.08\); the corresponding PACF values are \(0.58, 0.04, -0.06, 0.03\). The ordinary-lag ACF is also notably positive at lag 1. Use the seasonal pattern to suggest a candidate, and state the limitation.
Method. First calculate the rough reference bound:
At seasonal lag 12, the ACF value \(0.61\) and PACF value \(0.58\) exceed \(0.196\) in absolute value. At lag 24, the ACF value \(0.37\) exceeds the bound, whereas the PACF value \(0.04\) does not. At lag 36, the ACF value \(0.22\) is just above the bound, whereas the PACF value \(-0.06\) is not. At lag 48, neither value exceeds the bound.
The seasonal PACF has one clear initial spike followed by values close to zero; the seasonal ACF declines across the seasonal multiples. That pattern supports trying a seasonal AR(1), \(P=1\), as a candidate. The positive ordinary lag-1 dependence means ordinary AR behavior also needs consideration; it does not change the definition of \(P\). The isolated ACF value \(0.22\), only slightly beyond the rough bound, is not by itself evidence for a higher seasonal order.
Check. The proposed seasonal order counts steps of twelve months: \(P=1\) refers to lag 12, not ordinary lag 1. The common mistake is to call this seasonal AR(12); the correction is that the model has seasonal order one with period twelve.
Worked Example: A Seasonal ACF Cutoff Suggests Seasonal MA(1)
A hypothetical quarterly portfolio series has seasonal period \(s=4\) and \(n=121\) observations. Its seasonal-lag sample ACF at lags 4, 8, 12, and 16 is \(0.49, 0.06, -0.02, 0.04\); its seasonal-lag PACF is \(0.45, 0.24, 0.13, 0.09\). The ordinary-lag ACF at lags 1 and 2 is \(0.31\) and \(0.12\). Identify a seasonal candidate and interpret the pattern.
Method. The rough reference bound is:
The ACF at lag 4, \(0.49\), exceeds \(0.1782\). The next three seasonal ACF values, \(0.06,-0.02,0.04\), are all within the bound. In contrast, the PACF values at lags 4 and 8, \(0.45\) and \(0.24\), exceed the bound, while the values at lags 12 and 16 do not. The seasonal ACF therefore shows a strong first seasonal spike followed by small values, while the PACF declines rather than displaying the same immediate cutoff.
This supports a seasonal MA(1), \(Q=1\), as an initial candidate. As a check on the idealized signature, if the model used \(X_t=\varepsilon_t+\Theta\varepsilon_{t-4}\) with \(\Theta=0.8\), its theoretical lag-4 ACF would be:
which is close to the reported sample value \(0.49\). This is a consistency check, not proof that the data-generating coefficient equals \(0.8\). The ordinary-lag values indicate nonseasonal dependence may also be present, so the seasonal MA candidate should be considered alongside ordinary terms.
Common mistake. Treating every PACF bar outside the reference bound as a separate seasonal AR order would overstate the evidence. The seasonal ACF pattern is the clearer cutoff here; the approximate bound is a guide, and the complete model still needs checking.
Worked Example: Both Seasonal Functions Persist
A hypothetical monthly energy-use series has \(n=144\) observations and seasonal period \(s=12\). At lags 12, 24, 36, and 48, its ACF values are \(0.48, 0.31, 0.20, 0.08\), and its PACF values are \(0.42, 0.29, 0.18, 0.04\). The ordinary ACF at lags 1 and 2 is \(0.52\) and \(0.28\); the ACF values at cross-lags 11 and 13 are \(0.26\) and \(0.23\). Use these summaries to choose sensible candidate seasonal orders without claiming a unique answer.
Method. Calculate the rough reference bound:
For the ACF, the first three seasonal values exceed \(0.1633\): \(0.48,0.31,0.20\); the fourth, \(0.08\), does not. For the PACF, the first three also exceed \(0.1633\): \(0.42,0.29,0.18\); the fourth, \(0.04\), does not. Neither function has a clean one-step cutoff. Both decay across the seasonal multiples, which is compatible with mixed seasonal AR and MA dependence, but does not establish it.
The ordinary correlations at lags 1 and 2 are substantial, and the cross-lag values at 11 and 13 also exceed the rough bound. These features make it risky to attribute the seasonal-multiple pattern to seasonal orders alone. A reasonable next step is to compare a mixed seasonal candidate such as \((P,Q)=(1,1)\) with simpler seasonal alternatives, including \((1,0)\) and \((0,1)\), while allowing suitable ordinary terms. The seasonal evidence does not prove that either seasonal order must be one, or that the ordinary terms have any particular order.
Check. ACF and PACF both tailing off is not a clean signature for a pure seasonal AR or pure seasonal MA model. The conclusion is “compare plausible candidates,” not “select the largest significant lag as the order.”
Accounting for Ordinary Dependence and Ambiguity
Ordinary dependence matters because seasonal multiples are also ordinary lags. In monthly data, lag 12 is both the first seasonal lag and an ordinary lag twelve observations away. A regular AR or MA component can affect its ACF and PACF. Therefore, a seasonal spike should be interpreted in the context of the pattern across ordinary lags and seasonal multiples, not in isolation.
In multiplicative seasonal ARIMA models, ordinary and seasonal factors combine. Their expansion can create cross-lag terms as well as terms at the seasonal multiples. For a monthly series, inspect whether notable values near lags 11, 12, and 13 fit a broader pattern. Such nearby values do not automatically imply seasonal period 11 or 13; they may be associated with interaction between ordinary and seasonal dependence.
When both seasonal ACF and PACF tail off, possible explanations include a mixed seasonal ARMA component, ordinary dependence that complicates the simple pattern, or sampling variability. When a pattern looks like a cutoff, remember that a small sample can hide later correlations and a large number of inspected lags can produce isolated apparent spikes. Use plots to generate candidates, then fit and compare plausible specifications and examine whether their residuals retain a systematic seasonal pattern.
Mark \(s,2s,3s,\ldots\) explicitly; do not confuse a lag number with the seasonal order.
Look for a seasonal PACF cutoff with ACF tail-off as evidence for seasonal AR, or a seasonal ACF cutoff with PACF tail-off as evidence for seasonal MA.
Review ordinary lags and, in multiplicative models, nearby cross-lags before assigning the pattern solely to seasonal terms.
Use the pattern to nominate a small set of candidate orders, then assess fitted models and residual behavior.
Check Your Understanding
For each problem, identify what the seasonal-lag evidence suggests and note any important qualification.
- A quarterly series has \(n=100\) observations and seasonal period \(s=4\). Its seasonal ACF values at lags 4, 8, and 12 are \(0.44,0.03,0.01\), while its seasonal PACF values are \(0.39,0.21,0.12\). What seasonal order is a reasonable first candidate? (Answer: Seasonal MA(1), \(Q=1\); the seasonal ACF cuts off after its first seasonal lag.)
- A monthly series has \(n=64\), \(s=12\), and seasonal ACF values \(0.55,0.34,0.20\) at lags 12, 24, and 36. Its seasonal PACF values are \(0.51,0.02,-0.04\). Compute the rough reference bound and give a candidate. (Answer: \(1.96/\sqrt{64}=0.245\); seasonal AR(1), \(P=1\), is a candidate because the PACF has one clear initial spike and the ACF declines.)
- A monthly series has seasonal period 12. A visible ACF spike occurs at lag 24, while the seasonal PACF is near zero at lag 12 and lag 24. How many seasonal steps does lag 24 represent, and what should be checked before assigning an order? (Answer: Two seasonal steps; inspect the full seasonal ACF pattern, ordinary dependence, and other relevant lags rather than assigning \(P=2\) from that one ACF spike.)
- For a quarterly series with \(n=144\), seasonal ACF values at lags 4, 8, and 12 are \(0.40,0.25,0.04\), and seasonal PACF values are \(0.36,0.22,0.03\). Both functions have two values beyond the rough bound. Calculate the bound and state the appropriate conclusion. (Answer: \(1.96/\sqrt{144}\approx0.1633\); both functions show persistence rather than a clean single cutoff, so compare mixed or otherwise plausible seasonal candidates instead of declaring a unique order.)
- A monthly model has a strong ACF pattern at lags 11, 12, and 13, with seasonal period \(s=12\). Does this establish seasonal order 11 or 13? (Answer: No. Those are ordinary and nearby cross-lags around the first seasonal lag; consider ordinary-seasonal interaction and the complete ACF/PACF pattern.)