Reading a Seasonal ARIMA Specification
A SARIMA model records several choices in a compact order notation. The challenge is not just recognizing six integers: it is identifying which describe ordinary behavior, which describe seasonal behavior, and what time interval the seasonal period represents. A misplaced seasonal order can change the model interpretation substantially, even when the numbers themselves look familiar.
The standard notation is \((p,d,q)(P,D,Q)_s\). The first triple gives the ordinary autoregressive, differencing, and moving-average orders. The second gives the corresponding seasonal orders, and \(s\) gives the number of observations in one seasonal cycle. These orders do not, by themselves, report estimated coefficient values or say whether the model fits the data well.
Here, AR means autoregressive and MA means moving average. The ordinary orders refer to short-run lags measured one observation at a time. The seasonal orders refer to lags that are multiples of \(s\). Thus, for monthly observations with \(s=12\), a seasonal order of one refers to lag 12, not lag one. The same order numbers therefore have different lag interpretations depending on whether they are in the first or second triple.
What Each Order Represents
In an ordinary ARMA model, an AR order \(p\) indicates the highest ordinary lag in the autoregressive part, while an MA order \(q\) indicates the highest ordinary lag in the moving-average part. In a SARIMA model, \(P\) and \(Q\) play analogous roles at seasonal lags. For example, seasonal AR order \(P=2\) uses seasonal lag positions \(s\) and \(2s\); it does not mean that every lag from 1 through \(2s\) is a seasonal AR lag.
The differencing orders describe transformations made before the stationary ARMA structure is modeled. Ordinary differencing of order \(d\) applies the first-difference operator \(d\) times. Seasonal differencing of order \(D\) applies the seasonal-difference operator \(D\) times. In lag-operator notation, these are \((1-B)^d\) and \((1-B^s)^D\), respectively, where \(B y_t=y_{t-1}\). The subscript on \(B^s\) means that the observation is shifted back by \(s\) periods.
The compact notation is commonly used for a multiplicative seasonal ARIMA model. Its differencing part can be written as \((1-B)^d(1-B^s)^D y_t\), with ordinary and seasonal AR and MA components specified by their respective orders. The notation tells us which component orders are present; it does not give their numerical coefficients. In particular, an order of two specifies the number and locations of terms, not their estimated strengths.
| Component | Meaning | Lag interpretation |
|---|---|---|
| \(p\) | Ordinary AR order | Ordinary lags through \(p\) |
| \(d\) | Ordinary differencing order | Repeated differences one observation apart |
| \(q\) | Ordinary MA order | Ordinary lags through \(q\) |
| \(P\) | Seasonal AR order | Seasonal lags \(s,2s,\ldots,Ps\) |
| \(D\) | Seasonal differencing order | Repeated differences \(s\) observations apart |
| \(Q\) | Seasonal MA order | Seasonal lags \(s,2s,\ldots,Qs\) |
| \(s\) | Seasonal period | Observations in one seasonal cycle |
The table uses “through” to describe the order and possible lag positions, not a guarantee that every coefficient will be statistically significant or nonzero after fitting. When an order is zero, that component is absent. For example, \(P=0\) means no seasonal AR terms are specified; it does not mean the data have no seasonal pattern of any kind.
Why the Seasonal Difference Uses Lag \(s\)
The seasonal difference is designed to compare an observation with the one from the same position in the preceding seasonal cycle. This directly explains why the order notation uses \(B^s\), rather than \(B\), for seasonal differencing.
By definition, \(B y_t=y_{t-1}\). Applying it \(s\) times gives \(B^s y_t=y_{t-s}\).
Applying \(1-B^s\) gives \((1-B^s)y_t=y_t-y_{t-s}\), the difference between an observation and the observation one full cycle earlier.
If \(D=2\), apply \(1-B^s\) a second time to the first seasonal differences. The value of \(D\) counts applications; it does not change the seasonal period \(s\).
This derivation justifies the seasonal differencing part of the notation. For instance, with quarterly data and \(s=4\), one seasonal difference compares each quarter with the same quarter in the previous year. The period remains four observations whether \(D\) is zero, one, or two.
Worked Example: Quarterly Claims With Seasonal AR and Differencing
A hypothetical analyst models quarterly counts of equipment claims using SARIMA order \((1,1,0)(1,1,0)_4\). Interpret all six orders and identify the lag positions involved. Do not infer coefficient values.
Step 1: Read the ordinary triple. The first triple is \((p,d,q)=(1,1,0)\). Thus, the ordinary AR order is \(p=1\), the ordinary differencing order is \(d=1\), and the ordinary MA order is \(q=0\). There is one ordinary AR lag, lag 1, and no ordinary MA term.
Step 2: Read the seasonal triple. The second triple is \((P,D,Q)=(1,1,0)\). Thus, there is one seasonal AR order, one seasonal difference, and no seasonal MA term.
Step 3: Interpret the period. The subscript is \(s=4\), appropriate to four quarterly observations per year. The seasonal AR lag position is \(1s=4\), so it is lag 4. The seasonal difference compares \(y_t\) with \(y_{t-4}\).
Step 4: Summarize the specification. The model has an ordinary first difference and a quarterly seasonal difference; an ordinary AR term at lag 1 and a seasonal AR term at lag 4; and neither ordinary nor seasonal MA terms. This interpretation avoids the mistake of reading the second AR order as an AR term at lag 1.
Check. Since \(q=Q=0\), both MA components are absent. Since \(d=D=1\), there is one difference of each type, with the seasonal difference spanning four quarters.
Interpreting the Period in Context
The subscript \(s\) is not a count of seasonal terms. It identifies the number of observations in one seasonal cycle. Its calendar meaning depends on how often the series is observed. Monthly data commonly use \(s=12\) for an annual cycle; quarterly data commonly use \(s=4\); daily data can use \(s=7\) for a weekly cycle. These are interpretations of observation spacing, not universal rules: the analyst must know the data frequency and the seasonal cycle being modeled.
Changing the observation frequency changes the numerical period used to represent the same calendar cycle. An annual pattern is described by \(s=12\) for monthly observations but \(s=4\) for quarterly observations. A period of \(s=12\) on quarterly data would span 12 quarters, or three years, rather than one year.
Worked Example: Monthly Policy Applications
A hypothetical insurer records new policy applications monthly and uses the specification \((2,0,1)(0,1,2)_{12}\). Translate the specification into ordinary and seasonal components, and state the seasonal MA lag positions.
Step 1: Ordinary orders. The first triple is \((2,0,1)\). The ordinary AR order is \(p=2\), so the ordinary AR lag positions are 1 and 2. The ordinary differencing order is \(d=0\), so no ordinary differencing is specified. The ordinary MA order is \(q=1\), so there is an ordinary MA term at lag 1.
Step 2: Seasonal orders. The seasonal triple is \((0,1,2)\). There is no seasonal AR term, one seasonal difference, and a seasonal MA order of 2.
Step 3: Convert seasonal orders to lags. Monthly data have \(s=12\), so seasonal MA order \(Q=2\) corresponds to seasonal lag positions \(12\) and \(24\). The one seasonal difference compares an observation with the one 12 months earlier.
Step 4: State the interpretation in context. The model uses two ordinary AR lags and one ordinary MA lag, no ordinary difference, one annual seasonal difference, and seasonal MA terms at 12 and 24 months. The seasonal AR component is absent.
Check. \(D=1\) does not imply \(d=1\): the specification has seasonal differencing but no ordinary differencing. The seasonal period is 12 monthly observations, not 12 years.
Zero Orders, Repeated Orders, and the Limits of Notation
An order of zero is an explicit absence of that model component. For example, \((0,0,0)(0,0,0)_{12}\) has no AR, differencing, or MA orders specified, ordinary or seasonal. The period is still identified as 12, but no seasonal component is included in that order specification. This notation alone does not establish that the data are seasonally constant or that the specification is suitable.
Likewise, \(D=2\) means the seasonal difference is applied twice; it does not mean there are two different seasonal periods. With \(s=4\), the same period of four observations is used for both applications. Repeated differencing can have important statistical consequences, but the order notation itself only records the number of applications.
The orders do not identify an intercept, drift, estimated parameter values, or innovation distribution. Those details must be supplied separately when they are part of a model specification. Nor can the order tuple alone establish stationarity, invertibility, or forecast accuracy. It describes the model structure; assessment of whether that structure is appropriate requires additional information.
Worked Example: Weekly Sensor Readings and an Order Trap
A hypothetical monitoring system collects one sensor reading each day. An analyst writes \((0,1,1)(2,0,0)_7\). A reviewer claims this means two ordinary AR terms at lags 1 and 2, no seasonal effects, and a seasonal difference of order one. Identify the errors and give the correct interpretation.
Step 1: Read the first triple. The ordinary orders are \((p,d,q)=(0,1,1)\). Thus, there is no ordinary AR term, one ordinary difference, and one ordinary MA term at lag 1.
Step 2: Read the second triple. The seasonal orders are \((P,D,Q)=(2,0,0)\). Thus, the model has seasonal AR order 2, no seasonal difference, and no seasonal MA term.
Step 3: Convert the seasonal AR order to lags. The period is \(s=7\), so the seasonal AR lag positions are \(7\) and \(14\). They are not ordinary lags 1 and 2.
Step 4: Correct each claim. There are no ordinary AR terms because \(p=0\). Seasonal effects are not entirely absent because \(P=2\). There is no seasonal difference because \(D=0\); instead, \(d=1\) specifies one ordinary difference. The daily series uses a seven-observation seasonal cycle.
Check. The reviewer has confused the ordinary and seasonal triples and has swapped \(d\) with \(D\). The correct specification includes an ordinary MA lag 1, seasonal AR lags 7 and 14, and an ordinary first difference.
Common Exam Traps
- Swapping the triples. In \((p,d,q)(P,D,Q)_s\), the first triple is ordinary and the second is seasonal.
- Treating \(D\) as \(d\). \(d\) differences adjacent observations; \(D\) differences observations \(s\) periods apart.
- Using the wrong time unit for \(s\). The period is counted in observations, not years, months, or weeks unless each observation happens to be one such unit.
- Reading seasonal order \(P=2\) as ordinary lags 1 and 2. Its seasonal positions are \(s\) and \(2s\).
- Assuming all seasonality is removed when a seasonal order is zero. A zero says only that the corresponding component is absent; it does not prove anything about the data's seasonal behavior.
- Assuming the tuple contains coefficient estimates. It specifies component orders, not coefficient signs, magnitudes, or statistical significance.
Check Your Understanding
For each specification, identify the ordinary and seasonal components and interpret the period in context.
- A hypothetical quarterly series has order \((0,1,1)(1,0,0)_4\). State the ordinary difference and MA lag, and the seasonal AR lag. (Answer: One ordinary difference, ordinary MA lag 1, and seasonal AR lag 4; no seasonal difference.)
- A hypothetical monthly series has order \((1,0,0)(0,1,1)_{12}\). How many differences are specified, and what seasonal MA lag is present? (Answer: No ordinary difference and one seasonal difference; seasonal MA lag 12.)
- A hypothetical daily electricity-demand series uses \((2,0,0)(0,0,2)_7\) to represent a weekly cycle. List its ordinary AR and seasonal MA lag positions. (Answer: Ordinary AR lags 1 and 2; seasonal MA lags 7 and 14.)
- A hypothetical quarterly sales series is specified as \((0,0,0)(2,1,0)_4\). Explain why it is incorrect to call this an ordinary AR(2) model with an ordinary first difference. (Answer: The AR order 2 and difference order 1 are seasonal: seasonal AR lags 4 and 8, and one seasonal difference comparing observations four quarters apart; there are no ordinary AR terms or ordinary differences.)
- A hypothetical monthly service-call series has order \((1,2,0)(0,0,1)_{12}\). State the differencing operations and the seasonal MA lag, without confusing the two difference orders. (Answer: Two ordinary differences and no seasonal differences; one seasonal MA term at lag 12.)