Repeated Seasonal Differencing and Information Loss

Why Repeating a Seasonal Difference Deserves Care

A seasonal difference compares an observation with the one from the same season in the previous cycle. For quarterly data, it compares a quarter with the same quarter one year earlier; for monthly data, it compares a month with the same month one year earlier. This can remove a seasonal unit root when the series has one. Applying the same difference again is a separate modeling choice, not a harmless way to make the first difference “stronger.”

The key risk is that one seasonal difference may already have removed the nonstationary seasonal component. A second difference then acts on the remaining variation, can induce strong dependence, and discards additional information. This tutorial focuses on those consequences and on how to reason through them before specifying a model.

Definition: Let \(B\) denote the lag operator, so that \(By_t=y_{t-1}\), and let \(s\) be the seasonal period, a positive integer. The seasonal difference of order \(s\) is \(\Delta_s y_t=(1-B^s)y_t=y_t-y_{t-s}\). Repeating it gives \(\Delta_s^2y_t=(1-B^s)^2y_t=y_t-2y_{t-s}+y_{t-2s}\). More generally, \(k\) repeated seasonal differences give \(\Delta_s^k y_t=(1-B^s)^k y_t\).

The notation \(\Delta_s^2\) means that the same seasonal differencing operator is applied twice. It does not mean a difference at period \(2s\). In particular, the second seasonal difference compares two first seasonal differences \(s\) time steps apart.

What a Second Difference Changes

A first seasonal difference measures the change between matching seasons in adjacent cycles. A second seasonal difference measures the change in that seasonal change between adjacent cycles. That may be appropriate if the seasonal component has a second unit root, but it is unnecessary if one seasonal difference has already produced a stationary series.

The sample-size cost is immediate. If the original series has \(n\) observations, a first seasonal difference supplies \(n-s\) values, starting at time \(s+1\). A second supplies \(n-2s\) values, starting at time \(2s+1\), provided \(n>2s\). These counts describe the available differenced values; the differencing operation also removes level information, so the original series cannot generally be recovered from the differenced values alone.

Worked Example: Tracking Quarterly Differences and the Usable Sample

A hypothetical quarterly claims index is observed for 12 quarters. Its values, in index points, are \(100,120,90,110,105,123,94,113,111,125,98,117\). Compute its first and second seasonal differences for \(s=4\), and state how many transformed observations remain at each stage. This is a calculation on the stated data; no probability distribution, interest rate, or mortality assumption is being imposed.

Method. First subtract the value four quarters earlier. Then apply the same subtraction to the resulting seasonal differences, again four quarters apart.

The first differences begin at quarter 5:

\begin{aligned} \Delta_4y_5&=105-100=5, & \Delta_4y_6&=123-120=3,\\ \Delta_4y_7&=94-90=4, & \Delta_4y_8&=113-110=3,\\ \Delta_4y_9&=111-105=6, & \Delta_4y_{10}&=125-123=2,\\ \Delta_4y_{11}&=98-94=4, & \Delta_4y_{12}&=117-113=4. \end{aligned}

There are \(12-4=8\) first differences. The second differences begin at quarter 9, because they compare the first differences at quarters 9–12 with those at quarters 5–8:

\begin{aligned} \Delta_4^2y_9&=6-5=1, & \Delta_4^2y_{10}&=2-3=-1,\\ \Delta_4^2y_{11}&=4-4=0, & \Delta_4^2y_{12}&=4-3=1. \end{aligned}

There are \(12-2(4)=4\) second differences. The results check directly against the three-term formula; for example, \(\Delta_4^2y_9=111-2(105)+100=1\). The second difference describes changes in the year-to-year seasonal changes, not simply the original year-to-year changes. It also leaves only four transformed values in this short series.

Common mistake avoided: subtracting \(y_{t-8}\) from \(y_t\) and calling it a second seasonal difference. That is a lag-8 difference, \((1-B^8)y_t\), whereas the repeated difference is \(y_t-2y_{t-4}+y_{t-8}\).

Why Repeated Differencing Induces Dependence

The effect is especially clear when a seasonal difference is applied to white noise. Suppose \(\varepsilon_t\) is white noise with mean zero, variance \(\sigma^2\), and zero covariance at every nonzero lag. A white-noise series has no serial dependence, but differencing it creates overlapping combinations of the same noise terms. Repeated seasonal differencing therefore does not leave white noise unchanged.

Result: If \(\varepsilon_t\) is white noise with variance \(\sigma^2\), then \(w_t=\Delta_s^2\varepsilon_t=\varepsilon_t-2\varepsilon_{t-s}+\varepsilon_{t-2s}\) has variance \(6\sigma^2\), covariance \(-4\sigma^2\) at lag \(s\), covariance \(\sigma^2\) at lag \(2s\), and zero covariance at other positive lags. Thus its autocorrelations at \(s\) and \(2s\) are \(-2/3\) and \(1/6\), respectively.

Here is a direct justification. The three noise terms in \(w_t\) occur at distinct times, so their pairwise covariances are zero. The variance is the sum of their squared coefficients times \(\sigma^2\). At lag \(s\), \(w_t\) and \(w_{t-s}\) share two noise terms with opposite-signed coefficient products; at lag \(2s\), they share one term. At any other positive lag, no noise term is shared, so the covariance is zero.

\begin{aligned} \operatorname{Var}(w_t)&=(1^2+(-2)^2+1^2)\sigma^2=6\sigma^2,\\ \operatorname{Cov}(w_t,w_{t-s})&=(1)(-2)\sigma^2+(-2)(1)\sigma^2=-4\sigma^2,\\ \operatorname{Cov}(w_t,w_{t-2s})&=(1)(1)\sigma^2=\sigma^2. \end{aligned}

Dividing each covariance by \(6\sigma^2\) gives the stated autocorrelations. Equivalently, the repeated difference filters the noise by the seasonal moving-average polynomial \((1-B^s)^2=1-2B^s+B^{2s}\). The induced pattern is not evidence that the original white noise had this dependence; it was created by the differencing operation.

Worked Example: Quantifying Dependence Created by Repetition

Suppose a hypothetical quarterly series is already stationary white noise, with \(\varepsilon_t\) independent and normally distributed as \(N(0,4)\). Thus the mean is zero and the variance is \(4\); take \(s=4\). Find the variance and autocovariances of \(w_t=\Delta_4^2\varepsilon_t\), and its autocorrelations at lags 4 and 8. No interest rate or mortality assumption is relevant to this time-series calculation.

Method. Expand the repeated difference, then use the independence and variance of the three terms:

w_t=\varepsilon_t-2\varepsilon_{t-4}+\varepsilon_{t-8}.

The variance contributions are \(1^2(4)=4\), \((-2)^2(4)=16\), and \(1^2(4)=4\), giving

\operatorname{Var}(w_t)=4+16+4=24.

At lag 4, the shared terms contribute \((1)(-2)(4)=-8\) and \((-2)(1)(4)=-8\), so \(\operatorname{Cov}(w_t,w_{t-4})=-16\). At lag 8, the single shared term contributes \((1)(1)(4)=4\). There are no shared terms at other positive lags, so those autocovariances are zero.

The lag-4 autocorrelation is \(-16/24=-2/3\), and the lag-8 autocorrelation is \(4/24=1/6\). These values check against the general result after substituting \(\sigma^2=4\). A common error is to calculate only the variance and treat \(w_t\) as white noise; the shared terms establish that it is dependent.

Seasonal Unit-Root Removal Versus Over-Differencing

A seasonal unit root is a unit root associated with the seasonal lag structure; seasonal differencing is used to remove that nonstationary component. If the seasonal unit root has multiplicity one, one application of \(1-B^s\) removes it. Applying the operator again multiplies the transformed series by another \(1-B^s\), even though the targeted unit root has already been removed. The resulting dependence can complicate a model and make estimates less informative.

For contrast, if a series genuinely contains a seasonal random walk, one seasonal difference can reveal its innovations. A second difference then differences those innovations. This simple construction shows why the order of differencing must be tied to the behavior of the series, not selected just because further differencing makes a graph look less variable.

Worked Example: A Seasonal Random Walk and an Unneeded Second Difference

A hypothetical quarterly series follows \(y_t=y_{t-4}+\varepsilon_t\), with initial values \(y_1=10,y_2=20,y_3=30,y_4=40\). For \(t\geq5\), the innovations are independent \(N(0,1)\) random variables; the listed realized innovations are \(\varepsilon_5=2,\varepsilon_6=-1,\varepsilon_7=0,\varepsilon_8=3,\varepsilon_9=1,\varepsilon_{10}=2,\varepsilon_{11}=-2,\varepsilon_{12}=0\). Find \(y_5,\ldots,y_{12}\), the first seasonal differences, and the second seasonal differences. There is no discounting or mortality assumption in this hypothetical time-series example.

Method. Use the model recursively, adding each innovation to the observation four quarters earlier:

\begin{aligned} y_5&=10+2=12, & y_6&=20-1=19,\\ y_7&=30+0=30, & y_8&=40+3=43,\\ y_9&=12+1=13, & y_{10}&=19+2=21,\\ y_{11}&=30-2=28, & y_{12}&=43+0=43. \end{aligned}

The first seasonal differences are, in order from \(t=5\) to \(12\), \(12-10=2\), \(19-20=-1\), \(30-30=0\), \(43-40=3\), \(13-12=1\), \(21-19=2\), \(28-30=-2\), and \(43-43=0\). They reproduce the innovations exactly, as the model \(y_t-y_{t-4}=\varepsilon_t\) requires.

The second differences are the later first differences minus the corresponding earlier ones: at \(t=9\), \(1-2=-1\); at \(t=10\), \(2-(-1)=3\); at \(t=11\), \(-2-0=-2\); and at \(t=12\), \(0-3=-3\). Thus the second differences are \(-1,3,-2,-3\), which are differences of innovations, not the original innovations.

The calculations check against \(\Delta_4^2y_t=\varepsilon_t-\varepsilon_{t-4}\). For example, at \(t=10\), \(2-(-1)=3\). One seasonal difference is the transformation that removes the seasonal random walk in this model. Repeating it changes the white-noise innovations into a dependent series.

Information Lost and an Exam Trap

A seasonal difference removes level information within each seasonal position. If \(z_t=y_t-y_{t-s}\) is known, adding a constant separately to all observations in any one seasonal position leaves every \(z_t\) unchanged. For example, with quarterly data, adding a constant to every first-quarter observation, without changing other quarters, does not change the first seasonal differences. Recovering levels therefore requires starting values or other information.

A second seasonal difference loses more: its formula has three terms, and its finite-sample output begins at \(2s+1\). It also annihilates sequences that are linear across cycles for each seasonal position, not just constant across cycles. These facts explain why over-differencing is more than a loss of a few initial observations: the transformation can erase structure and impose dependence on the retained series.

Common exam trap: seeing a seasonal pattern and automatically applying two seasonal differences. The correct reasoning is to establish whether the seasonal difference is needed and whether one application is sufficient. If the series after one seasonal difference is already stationary, applying the same operator again can over-difference it. The induced negative autocorrelation at seasonal lag is one clue, but it must be assessed with the model and data rather than treated as conclusive by itself.

Identify the seasonal period.
State \(s\) and confirm that the seasonal difference compares matching positions in successive cycles.

Ask what remains after one difference.
Do not infer that another difference is needed merely because the transformed series still varies. Variation can be stationary noise.

Assess the cost of repetition.
For \(n\) observations, a second seasonal difference leaves \(n-2s\) values and may create moving-average dependence at seasonal lags.

Interpret the transformed quantity.
A first difference measures seasonal changes; a second measures changes in those changes. Confirm that this is what the model requires.

Key takeaway: Repeated seasonal differencing is justified only when the remaining series still requires that transformation. When one seasonal difference has already removed a seasonal unit root, another can discard useful level and sample information while inducing a seasonal moving-average dependence pattern.

Check Your Understanding

For each problem, identify the operation carefully before calculating.

A monthly series has \(n=48\) observations and seasonal period \(s=12\). How many first and second seasonal differences are available, and at what first time index does the second-difference series begin? (Answer: 36 first differences; 24 second differences, beginning at \(t=25\).)
For \(s=3\), a series has \(y_7=18\), \(y_4=15\), and \(y_1=10\). Calculate \(\Delta_3y_7\) and \(\Delta_3^2y_7\). (Answer: \(\Delta_3y_7=3\) and \(\Delta_3^2y_7=18-2(15)+10=-2\).)
White noise has variance \(9\). After applying two seasonal differences at period \(s=6\), find the variance and the autocovariance at lag 12. (Answer: variance \(54\); lag-12 autocovariance \(9\).)
A series follows \(y_t=y_{t-4}+\varepsilon_t\), where the innovations are white noise with variance \(2\). State what the first seasonal difference equals and find the variance of the second seasonal difference. (Answer: \(\Delta_4y_t=\varepsilon_t\); \(\operatorname{Var}(\Delta_4^2y_t)=2(2)=4\).)
A 30-observation quarterly series is seasonally differenced twice. A candidate claims the result is a lag-8 difference and gives 22 values. Identify both errors and give the correct expression and number of values. (Answer: the operator is \((1-B^4)^2=1-2B^4+B^8\), not \(1-B^8\); there are \(30-2(4)=22\) values, so the count is correct.)

What You'll Learn