When does a monthly series need seasonal differencing at lag 12 rather than a first difference?
answer
- compare like month with like month
- one lag-12 step, not twelve lag-1 steps
- it also flattens a straight-line climb
- seasonal first, then re-test the level
- transform the spread before touching the level
basics
~20 sSeasonal differencing is needed when a repeating annual pattern, not a drifting level, is what moves the mean. Subtracting the value twelve months earlier removes a stable yearly pattern that a lag-1 difference leaves intact.
solid answer
~50 sA lag-1 difference removes a wandering level; it does nothing about a pattern that repeats every twelve months, because January still differs systematically from July after you subtract the previous month. The seasonal difference `X_t - X_{t-12}` compares like month with like month and removes a stable annual pattern outright. Monthly electricity demand is the standard case: winter and summer peaks recur every year. A bonus is that the lag-12 difference of a linear trend is a constant, so one seasonal difference often removes a steady trend too. The order of work is: stabilise the variance first with a log or Box-Cox transform if the swings grow with the level, then take one seasonal difference, then test again and add a single first difference only if the level still wanders. It costs the first twelve observations.
go deeper
Be ready to write the seasonal difference as the current value minus the value one full cycle earlier, and to say why subtracting the previous month leaves an annual pattern untouched.
Expect to explain the ordering — variance transform, then seasonal difference, then re-test, then at most one first difference — and to know that a lag-12 difference also removes a straight-line trend.
Show that you weigh the cost: twelve observations gone, and a judgment about whether the annual pattern is stable enough to remove wholesale on a short series rather than differencing reflexively.
Own the preprocessing convention across the team's forecasting work, so the same series is not transformed three different ways by three analysts and the resulting forecasts stay comparable.
## The two difference operators There are two distinct operations, and interviewers ask this question to check that a candidate does not blur them. - **First difference:** `X_t - X_{t-1}`. Compares each period with the one immediately before. Removes a wandering level. - **Seasonal difference at lag m:** `X_t - X_{t-m}`. For monthly data with an annual cycle, `m = 12`. Compares each period with the same period one cycle earlier — January with the previous January. They are not interchangeable, and neither is the other applied repeatedly. Taking twelve successive first differences is a completely different transformation from taking one lag-12 difference, and it is a common confusion worth naming explicitly. ## Why a first difference does not remove a yearly pattern Suppose demand follows a fixed annual shape: high in January, low in April, high again in July. The first difference measures the month-to-month change, and that change itself follows a fixed annual shape — the January-to-February drop happens every year. The differenced series therefore still has a mean that depends on the calendar month, which violates the constant-mean condition. Subtracting the value twelve months earlier is what lines up like with like: whatever fixed contribution the month of January makes appears in both terms and cancels. ## The trend bonus A lag-12 difference also removes a deterministic linear trend. If the underlying level is `a + b*t`, then `(a + b*t) - (a + b*(t - 12)) = 12*b` a constant. So a single seasonal difference handles both a stable annual pattern and a steady linear climb, leaving a level that no longer drifts. This is why the recommended order is seasonal difference first, then re-test — quite often no further differencing is needed at all. ## When you need both You need a first difference *in addition* when, after the seasonal difference, the level still wanders in a way that does not look like it will come back — that is, when the series carries a stochastic trend on top of its annual pattern rather than a steady deterministic climb. The workflow is: 1. Stabilise the variance if the swings grow with the level. 2. Take one seasonal difference at lag 12. 3. Re-check stationarity on the result. 4. Only if the level still wanders, take one first difference of that already-seasonally-differenced series. Do not run this in the opposite order as a habit, and do not take two seasonal differences. A second seasonal difference is very rarely justified; more than one of each is a strong signal that something other than integration — a variance that moves, or a misspecified deterministic component — is driving the diagnostics. ## Variance stabilisation comes first Differencing addresses the level. It does nothing about a spread that grows as the series grows. On a subscription-revenue series that is compounding, or on electricity demand where the size of the seasonal swing widens as total consumption rises, the differenced series will still show small early wiggles and large late ones. The fix is a transform applied to the raw values before any differencing: a log, or more generally a Box-Cox transform with a fitted parameter — the log is the Box-Cox case where that parameter is zero. After the transform, proportional swings become roughly constant-sized swings, and the differencing that follows produces a series with a stable spread. Applying the log after differencing does not work, because differences can be negative and, more fundamentally, because the transform has to act on the multiplicative structure while it is still present in the levels. ## The costs A seasonal difference at lag 12 discards the first twelve observations, which is a real price on a series of only a few years. Three years of monthly data become two years of usable differenced values. That argues for not differencing more than the evidence requires and for weighing whether the annual pattern is stable enough that removing it wholesale is the right choice — a seasonal difference assumes the pattern is *not* fixed forever and lets it evolve, which is its strength on long series and its cost on short ones. Finally, remember that the period is a property of the data's frequency and its cycle, not a constant. Monthly data with an annual cycle uses 12; quarterly data with an annual cycle uses 4; daily data with a weekly cycle uses 7. Picking the wrong period leaves the pattern in place and wastes observations.
- When does a monthly series need both a seasonal and a first difference?When a stable annual pattern sits on top of a level that wanders stochastically rather than climbing steadily. Take the seasonal difference first, since it also removes a deterministic linear trend, then re-test. If the level of the seasonally differenced series still drifts without reverting, add exactly one first difference and re-test again. Needing more than one of each type usually points to unstabilised variance rather than to deeper integration.
- Why apply a log or Box-Cox transform before differencing rather than after?Because differencing fixes a moving level, not a moving spread. If the size of the seasonal swings grows in proportion to the level, the differenced series still has changing variance. A log — the Box-Cox case with parameter zero — turns proportional swings into roughly constant-sized ones while the multiplicative structure is still present in the levels. After differencing, that structure is gone and the values can be negative, so the transform no longer applies.
- How much data does one seasonal difference of monthly data cost?The first twelve observations, since each differenced value needs its counterpart from a year earlier. On three years of monthly data that turns 36 points into 24, which is a substantial loss. If a first difference is then also taken, one more observation goes. This is a concrete argument for differencing only as much as the evidence supports, especially on short series.
saying these in an interview costs you the question
- Applies twelve successive first differences instead of one lag-12 difference
- Takes a second seasonal difference to be safe
- Believes differencing fixes a spread that grows with the level
- Applies the log transform after differencing rather than before
- Uses a period of 12 for quarterly or weekly data
- Ignores the twelve observations lost to the seasonal difference