Why the Type of Differencing Matters
A series can vary over time because its overall level changes, because its seasonal pattern changes, or because both happen. Those patterns call for different questions before differencing: does the series persistently drift from one observation to the next, does it persistently return to a different level each season, or are both features present? Applying the wrong difference can leave the important nonstationarity in place, or add dependence that was not in the original series.
For a series \(y_1,\ldots,y_T\) observed at equally spaced times, let \(m\) be the number of observations in one seasonal cycle. For monthly observations with annual seasonality, \(m=12\); for quarterly observations with annual seasonality, \(m=4\). Differencing compares observations separated by a specified number of time steps. It is a way to remove a changing level, not an automatic treatment for any visible pattern.
A weakly stationary series has a constant mean and variance, and its covariance between two observations depends only on the time separation between them. A series with a changing level may not meet these conditions. Differencing is warranted when the changing level reflects persistent stochastic nonstationarity; a repeating but stable seasonal mean, by itself, does not establish that seasonal differencing is needed.
Reading Seasonal Plots and Seasonal-Lag Autocorrelation
A seasonal plot groups observations by their position in the cycle: for example, all first quarters are shown together, then all second quarters, and so on. It helps distinguish a stable seasonal shape from a changing one. If the seasonal profile is roughly stable while the overall level rises, a trend is more plausible than a seasonal unit root. If observations in each season continue to move away from their previous-cycle counterparts, seasonal differencing may be relevant.
Autocorrelation provides a second diagnostic. The autocorrelation at lag \(k\) is the correlation between observations \(k\) time steps apart. In a series with seasonal period \(m\), examine lags \(m,2m,3m,\ldots\). Large, persistent positive autocorrelation at those lags can be evidence that the series retains seasonal-level persistence. It is not proof on its own: ordinary trend can also produce high autocorrelation at seasonal lags. Interpret the seasonal-lag pattern alongside the seasonal plot and the behavior at short lags.
| Evidence in the series | Question to ask | Initial differencing candidate |
|---|---|---|
| Persistent drift between consecutive observations; no changing seasonal levels | Does one-step differencing remove the trend? | Ordinary differencing |
| Persistent changes from one seasonal cycle to the next | Does comparing the same season across cycles stabilize the series? | Seasonal differencing |
| Both persistent one-step movement and seasonal-level persistence | Does one difference leave the other changing feature? | Consider both, then reassess |
| Stable seasonal profile or stable level, with no persistent movement | Would differencing remove a feature that is already stable? | Do not difference automatically |
These are diagnostic choices, not mechanical rules. In particular, a seasonal pattern with stable seasonal means can be represented without seasonal differencing. Seasonal differencing is most compelling when same-season observations retain persistent level changes, rather than merely because the seasons have different average values.
Worked Examples
Worked Example: Seasonal Differencing Without Ordinary Differencing
A hypothetical quarterly service series has seasonal period \(m=4\). The first three years of observations are \(20,25,18,23\); \(24,23,21,24\); and \(27,26,21,27\). In the longer record, the seasonal plot shows that same-quarter levels move persistently between years, while the overall mean is not drifting steadily within each year. The autocorrelation remains positive at lags 4 and 8 but is not persistently large at successive short lags. Calculate both types of first differences for the displayed observations and select a candidate.
Method. Compare each value first with the immediately preceding quarter, then with the same quarter in the preceding year. The values alone are a short excerpt; the stated plot and autocorrelation evidence summarize the longer record used to assess persistence.
The observations in order are \(20,25,18,23,24,23,21,24,27,26,21,27\). The ordinary differences, from the second observation onward, are:
and then \(-1,-2,3,3,-1,-5,6\). Thus the complete ordinary-difference sequence is \(5,-7,5,1,-1,-2,3,3,-1,-5,6\). It contains a repeating within-year pattern in this short excerpt.
The seasonal differences compare each quarter with the quarter four observations earlier:
followed by \(27-24=3,\;26-23=3,\;21-21=0,\;27-24=3\). The seasonal-difference sequence is \(4,-2,3,1,3,3,0,3\). These calculations show exactly what each operation compares; they do not, by themselves, establish stationarity from only eight seasonal differences.
Given the longer-record evidence, seasonal differencing is the better candidate: it directly compares corresponding quarters, and the seasonal-lag autocorrelation indicates persistent same-quarter dependence. Ordinary differencing alone can leave that seasonal-level persistence. The common mistake is to treat a visible quarterly pattern as sufficient reason to apply both differences. Here the stated evidence does not support adding ordinary differencing as well.
Check. There are \(12-4=8\) seasonal differences, as expected, and each is the current observation minus the value four quarters earlier.
Worked Example: Ordinary Differencing When Seasonal-Lag Correlation Is Misleading
A hypothetical monthly index has a linear increase of 2 units each month and no seasonal variation. Its first 12 values are \(40,42,44,46,48,50,52,54,56,58,60,62\). With \(m=4\) used as a candidate seasonal period, the plot shows a rising level but no recurring within-cycle profile. Choose between ordinary differencing and seasonal differencing using the displayed values.
Ordinary differences. Each successive value increases by 2:
All 11 ordinary differences equal 2. The rising level has become a constant sequence.
Seasonal differences. Each observation is 4 time steps after the one subtracted from it:
All 8 seasonal differences equal 8. This does not remove the changing level; it converts the linear increase into a nonzero constant. Taking both differences gives zero for each available value because the seasonal-difference sequence is constant.
The candidate is ordinary differencing, not seasonal differencing. A steadily rising series can have strong positive autocorrelation at lag 4, so that feature alone does not diagnose seasonal persistence. The seasonal plot and the calculated differences show that the rise is not a recurring seasonal movement. The common mistake is to read any strong seasonal-lag autocorrelation as evidence for seasonal differencing.
Check. The ordinary-difference sequence has \(12-1=11\) values; the seasonal-difference sequence has \(12-4=8\).
Worked Example: Considering Both Differences
A hypothetical quarterly index has \(m=4\). Its first three years are \(100,110,105,115\); \(112,124,117,131\); and \(128,142,133,151\). In the full record, the seasonal plot shows changing same-quarter levels as well as movement in the overall level. Autocorrelation persists at short lags and at seasonal lags. Compute the seasonal differences and then their ordinary differences.
Step 1: Seasonal differences. Comparing the second year with the first gives \(112-100=12,\;124-110=14,\;117-105=12,\;131-115=16\). Comparing the third year with the second gives \(128-112=16,\;142-124=18,\;133-117=16,\;151-131=20\). The seasonal-difference sequence is therefore \(12,14,12,16,16,18,16,20\).
Step 2: Ordinary differences of that sequence. Subtract each seasonal difference from the one immediately before it:
followed by \(18-16=2,\;16-18=-2,\;20-16=4\). The combined-difference sequence is \(2,-2,4,0,2,-2,4\). It has \(12-4-1=7\) values, also as given by applying the two lag operations.
The seasonal differences still rise across years, so seasonal differencing alone has not removed all changing-level behavior. Both differences are reasonable candidates given the stated evidence. The resulting sequence still has a repeating pattern; this example does not claim that the combined differences are white noise or that differencing alone completes model selection. The common mistake is to assume that using both differences guarantees a satisfactory stationary series. After differencing, inspect the result again.
Check. Each seasonal difference uses a four-quarter comparison, and each combined difference subtracts adjacent values in the seasonal-difference sequence.
Why the Combined-Difference Formula Holds
The ordinary and seasonal differences are linear subtraction operations. Apply ordinary differencing to the seasonal-difference series: the value at time \(t\) is \(y_t-y_{t-m}\), and the value at time \(t-1\) is \(y_{t-1}-y_{t-m-1}\). Subtracting the second from the first gives:
The same result follows if the ordinary difference is computed first and then seasonally differenced. This is because both operations are fixed linear differences, so their order commutes. The formula makes clear that combined differencing removes both a one-step and a same-season comparison; it also makes clear why the combined operation uses four observations.
Overdifferencing: Symptoms and a Useful Check
Overdifferencing occurs when a difference is applied to a series that does not need it, or when too many differences are taken. One warning sign is a strong negative autocorrelation at the lag corresponding to the unnecessary difference: lag 1 after unnecessary ordinary differencing, or lag \(m\) after unnecessary seasonal differencing. Another is an increase in variation without a corresponding improvement in stability. These are warnings, not conclusive tests; assess them together with the plot and the substantive pattern.
The origin of the negative correlation can be seen for ordinary differencing of white noise. Suppose \(y_t=e_t\), where \(e_t\) is independent over time, has mean zero and variance \(\sigma^2\). Then \(x_t=\nabla y_t=e_t-e_{t-1}\). Its variance is \(2\sigma^2\), because the two independent terms have variances \(\sigma^2\) each. Adjacent values share one innovation with opposite signs, so:
Therefore \(\operatorname{Corr}(x_t,x_{t-1})=-\sigma^2/(2\sigma^2)=-0.5\). Differencing has introduced negative dependence into a series that was already stationary. For seasonal differencing of white noise, the same reasoning gives covariance \(-\sigma^2\) between values \(m\) time steps apart and variance \(2\sigma^2\), so the induced correlation at lag \(m\) is \(-0.5\).
Common trap. A candidate sees a high positive autocorrelation at lag \(m\) and immediately applies seasonal differencing, even though the seasonal plot shows a stable profile and the series has an ordinary trend. That decision confuses persistence caused by an overall rising level with persistence in the seasonal levels. Check whether the same-season pattern itself changes, compare short- and seasonal-lag behavior, and reassess the differenced series for an unusually negative correlation at the lag just differenced.
Check Your Understanding
For each problem, identify the comparison made by the proposed difference and select a reasonable initial choice from the evidence.
- A quarterly series has \(m=4\). Four consecutive observations are \(30,34,31,36\), and the next year begins with \(33,37,34,39\). Calculate the four seasonal differences available from those two years. (Answer: \(3,3,3,3\))
- A monthly series has \(m=12\). Its seasonal plot has a stable yearly profile that shifts upward steadily; autocorrelation is high at lags 1 and 12. The first difference removes the rise, while same-month comparisons continue to show a trend. Which difference is the better initial choice, and why? (Answer: Ordinary differencing; the evidence describes an overall rise, not persistent changing seasonal levels.)
- A quarterly series has \(m=4\), and its seasonal differences are \(5,7,6,8,5,7,6,8\). Calculate the ordinary differences of this seasonal-difference sequence. (Answer: \(2,-1,2,-3,2,-1,2\))
- A series is already stationary white noise with variance \(\sigma^2\). It is unnecessarily seasonally differenced at period \(m\). What is the variance of the differenced series and its autocorrelation at lag \(m\)? (Answer: Variance \(2\sigma^2\); autocorrelation at lag \(m\) is \(-0.5\).)
- A quarterly series has both a changing overall level and persistent movement in same-quarter levels. Seasonal differencing reduces the seasonal-lag persistence, but the differenced series still has a changing trend. What candidate should be considered next, and what must be checked afterward? (Answer: Consider ordinary differencing as well; then reassess the resulting series for remaining patterns and signs of overdifferencing.)