Seasonal Differencing Versus Seasonal Indicators

Why the Kind of Seasonality Matters

A quarterly series may rise in the same quarter every year, but that visible repetition can have different causes. The average level might be predictably higher in one quarter than another, while deviations from those averages have little seasonal persistence. Alternatively, the series may retain a seasonal level from one year to the next, so that a shock to a given quarter carries forward into the same quarter of later years. A model should distinguish these situations rather than treating every repeating pattern the same way.

Seasonal indicators represent a deterministic repeating mean: each season has its own fixed average, after allowing for any other specified terms such as a trend. Seasonal differencing represents a different feature: dependence across observations one full seasonal cycle apart, including a seasonal unit root when the seasonal level does not revert to a fixed mean. The two tools are not necessarily alternatives; a model can include seasonal indicators and also model dependence in the errors.

Definition: For seasonal period \(s\), a deterministic seasonal mean is a repeating set of mean levels, one for each of the \(s\) seasons. Seasonal indicators estimate those levels. Stochastic seasonal dependence means that observations, or their deviations from the mean, are related across seasonal lags such as \(s, 2s,\ldots\). Seasonal differencing uses \(\nabla_s y_t=y_t-y_{t-s}\) to remove a persistent seasonal level, especially one arising from a seasonal unit root.

Seasonal Indicators and Seasonal Differencing

Suppose \(y_t\) is observed at equally spaced times and there are \(s\) seasons per year. For quarterly data, \(s=4\); for monthly data, \(s=12\). With an intercept, one standard indicator model uses one season as the reference:

y_t=\beta_0+\sum_{j=1}^{s-1}\beta_j I_{j,t}+\varepsilon_t.

Here \(I_{j,t}=1\) when observation \(t\) is in season \(j\), and is 0 otherwise. The reference season is season \(s\). The coefficient \(\beta_0\) is the mean in that reference season if the errors have mean zero, and \(\beta_j\) is the difference between the mean in season \(j\) and the reference-season mean. Using indicators for all \(s\) seasons and an intercept would create exact collinearity, because the indicators sum to 1 at every time. A trend can also be included if the mean is expected to change over time.

An indicator model concerns the conditional mean. It does not require the errors \(\varepsilon_t\) to be independent: they may instead follow a stationary ARMA model, including seasonal dependence. By contrast, seasonal differencing transforms the series itself. It subtracts the observation from one full seasonal cycle earlier:

\nabla_s y_t=(1-B^s)y_t=y_t-y_{t-s},

where \(B\) is the backshift operator, defined by \(B y_t=y_{t-1}\). Differencing is suited to a seasonal level that persists stochastically, as in \(y_t=y_{t-s}+\varepsilon_t\), when the seasonal difference is stationary. It is not automatically needed merely because the raw series has a large and stable difference between seasons.

Worked Example: Estimating a Fixed Quarterly Pattern

A hypothetical claims unit records quarterly processing totals (in hundreds of claims) for three years. The observations are: Year 1: 101, 129, 91, 109; Year 2: 99, 131, 89, 111; Year 3: 100, 130, 90, 110. Determine the seasonal indicator estimates using Quarter 4 as the reference, and calculate the seasonal differences.

Method. Calculate each quarter’s sample mean, then express the other quarter means as contrasts with Quarter 4. This estimates fixed seasonal means without imposing a seasonal unit root.

The quarterly sums are \(101+99+100=300\), \(129+131+130=390\), \(91+89+90=270\), and \(109+111+110=330\). Dividing each by 3 gives means \(100\), \(130\), \(90\), and \(110\), respectively. Thus \(\hat\beta_0=110\), the Quarter 4 mean. The indicator estimates are:

\hat\beta_1=100-110=-10,\qquad \hat\beta_2=130-110=20,\qquad \hat\beta_3=90-110=-20.

For example, the fitted mean for Quarter 2 is \(110+20=130\). The other fitted means are \(110-10=100\), \(110-20=90\), and \(110\), matching the four calculated quarterly means.

Now calculate \(y_t-y_{t-4}\), comparing each quarter with the same quarter in the prior year. From Year 1 to Year 2, the differences are \(99-101=-2\), \(131-129=2\), \(89-91=-2\), and \(111-109=2\). From Year 2 to Year 3, they are \(100-99=1\), \(130-131=-1\), \(90-89=1\), and \(110-111=-1\). The fixed quarter means cancel, leaving only changes in deviations from those means. Common mistake avoided: a seasonal difference is not the difference between adjacent quarters; it compares the same quarter across years. These modest differences do not by themselves establish a seasonal unit root.

Why Differencing Independent Errors Creates Dependence

A key reason not to difference automatically is that seasonal differencing changes the error structure. Suppose a stable seasonal mean has been represented by indicators and the remaining errors are independent with mean zero and variance \(\sigma^2\). For two observations from the same season in successive years, the seasonal difference cancels the fixed mean, but the new error is a difference of two errors.

\nabla_s y_t=\varepsilon_t-\varepsilon_{t-s}.

The variance follows from independence: the variance of a difference is the sum of the two variances when the covariance is zero. To see why differenced errors at seasonal lag \(s\) are negatively related, write the two differences explicitly. Their only shared error is \(\varepsilon_{t-s}\), with opposite signs:

\begin{aligned} \operatorname{Var}(\varepsilon_t-\varepsilon_{t-s}) &=\operatorname{Var}(\varepsilon_t)+\operatorname{Var}(\varepsilon_{t-s})=2\sigma^2,\\ \operatorname{Cov}(\varepsilon_t-\varepsilon_{t-s}, \varepsilon_{t-s}-\varepsilon_{t-2s}) &=-\operatorname{Var}(\varepsilon_{t-s})=-\sigma^2. \end{aligned}

Consequently, the correlation between these differenced errors at lag \(s\) is \(-\sigma^2/(2\sigma^2)=-1/2\), provided \(\sigma^2>0\). This induced seasonal moving-average dependence is a mathematical consequence of differencing independent errors; it is not evidence that the original errors had seasonal persistence.

Worked Example: The Dependence Induced by Seasonal Differencing

Suppose quarterly observations have fixed quarter means \(20,35,18,27\), and independent errors \(\varepsilon_t\sim N(0,9)\). In two consecutive years, the error values by quarter are Year 1: \(2,-2,1,-1\), and Year 2: \(-1,1,0,2\). Calculate the four seasonal differences from Year 1 to Year 2 and the theoretical variance and lag-4 correlation of a seasonally differenced error.

Method. Add each error to its quarter mean to obtain the observations. Then subtract the prior year’s observation in the same quarter. The fixed mean cancels quarter by quarter.

The observations are \(22,33,19,26\) in Year 1 and \(19,36,18,29\) in Year 2. The four differences are therefore \(19-22=-3\), \(36-33=3\), \(18-19=-1\), and \(29-26=3\). These also equal the error changes: \(-1-2=-3\), \(1-(-2)=3\), \(0-1=-1\), and \(2-(-1)=3\). This checks that the seasonal means have canceled.

Since \(\sigma^2=9\), the variance of a differenced error is \(2(9)=18\). The covariance between differences four time steps apart is \(-9\), so their correlation is \(-9/18=-0.5\). Common mistake avoided: it would be incorrect to interpret this induced negative correlation as proof of an original seasonal moving-average process. The calculation assumed independent original errors; the correlation arose from the shared error with opposite signs.

When the Seasonal Level Is Stochastic

A seasonal unit-root model illustrates when differencing is purposeful. In \(y_t=y_{t-s}+\varepsilon_t\), each season’s level is carried forward from the corresponding season one year earlier, with a new shock added. Applying the seasonal difference gives \(\nabla_s y_t=\varepsilon_t\). If the innovations are stationary, the differenced series is stationary even though the original seasonal levels do not settle around fixed quarter means.

This differs from the indicator model. Indicators estimate a stable mean for each season; they do not remove a seasonal unit root. Seasonal differencing removes the evolving seasonal level; it does not estimate a separate long-run mean for each season. If both stable quarter effects and seasonal persistence in deviations are plausible, a model can represent the quarter means with indicators and use a suitable time-series error model. The choice concerns the mean structure and the dependence structure separately.

Worked Example: Detecting a Seasonal Random Walk

A hypothetical quarterly index begins with \(y_1=50\), \(y_2=60\), \(y_3=55\), and \(y_4=65\). For quarters 5 through 8, suppose the data follow \(y_t=y_{t-4}+\varepsilon_t\), with innovations \(\varepsilon_5=2\), \(\varepsilon_6=-1\), \(\varepsilon_7=3\), and \(\varepsilon_8=0\). Calculate the next four observations and seasonal differences. Explain why the initial quarterly pattern alone does not justify seasonal indicators as the solution.

Method. Apply the seasonal random-walk equation separately to each quarter, using the observation four steps earlier.

The calculations are \(y_5=y_1+\varepsilon_5=50+2=52\), \(y_6=y_2+\varepsilon_6=60-1=59\), \(y_7=y_3+\varepsilon_7=55+3=58\), and \(y_8=y_4+\varepsilon_8=65+0=65\). The seasonal differences are \(y_5-y_1=52-50=2\), \(y_6-y_2=59-60=-1\), \(y_7-y_3=58-55=3\), and \(y_8-y_4=65-65=0\). Each equals its specified innovation, as required by the model.

The starting values differ by quarter, but those four observations alone do not show that fixed quarter means govern the full series. Under the stated model, each quarter’s level is carried forward and changes with new shocks. Seasonal indicators alone would not remove that persistence. Common mistake avoided: treating one observed cycle of high and low values as proof of a deterministic seasonal mean, without considering how the same-season levels evolve over time.

A Practical Distinction

The central modeling question is not simply whether a seasonal pattern is visible. Ask what is repeating and what happens to deviations from that pattern. When the same seasonal means recur while deviations are stationary, indicators can account for the mean pattern; any remaining serial correlation should be modeled separately. When same-season levels persist and the seasonal difference is plausibly stationary, seasonal differencing addresses that evolving level. A stable average pattern and stochastic persistence may also coexist.

A large seasonal difference in the raw data is not by itself a reason to difference. If the difference between quarters is stable, seasonal differencing removes that mean pattern but can also create dependence in otherwise independent errors. Conversely, seasonal indicators cannot substitute for differencing when a seasonal unit root is present. The assumptions behind the proposed mean and error structure matter more than the visual size of a seasonal swing.

Key takeaway: Use seasonal indicators to represent fixed, repeating seasonal means. Use seasonal differencing to address persistent stochastic seasonal levels when the differenced series is stationary. If the data contain both features, represent the seasonal mean and the error dependence separately rather than expecting one device to do both jobs.

Check Your Understanding

For each problem, distinguish the seasonal mean question from the seasonal dependence question.

A monthly service count has a stable mean of 40 in January and 70 in July. After subtracting each month’s mean, the errors are independent with variance 4. What is the variance of a January-to-January seasonal difference of the errors? (Answer: 8)
A quarterly series follows \(y_t=y_{t-4}+\varepsilon_t\), where the innovations are stationary with mean zero. What transformation removes the seasonal level, and what is the transformed series? (Answer: \(\nabla_4 y_t=y_t-y_{t-4}=\varepsilon_t\))
An indicator model uses Quarter 4 as the reference, with estimated intercept 200 and Quarter 2 coefficient \(-15\). What is the fitted Quarter 2 mean, and what does the coefficient represent? (Answer: 185; Quarter 2’s mean is 15 below Quarter 4’s mean)
Errors are independent with variance 16, and a seasonal difference is formed with period \(s=12\). What is the theoretical correlation between differenced errors at lag 12? (Answer: \(-0.5000\))
A series has stable seasonal means, but its indicator-adjusted errors remain correlated at lag 4. Should the seasonal indicators necessarily be discarded, and what feature needs additional modeling? (Answer: No; retain indicators if the seasonal means are stable and model the remaining lag-4 error dependence)

What You'll Learn