In a classical time-series decomposition, what do the trend, seasonal and remainder components each represent?
answer
- three parts, not two
- one is the slow-moving level
- one repeats at a fixed known period
- one is whatever is left over
- the leftovers are a residual, not proven noise
basics
~20 sDecomposition splits a series into three parts: trend-cycle, the slow movement of the level; seasonal, the pattern that repeats at a fixed known period such as 12 months; and remainder, the variation left after removing both.
solid answer
~40 sA classical decomposition writes each observation as the sum of three pieces: `y = T + S + R`. `T` is the trend-cycle, the smooth long-run movement of the level. `S` is the seasonal component, a pattern that repeats at a fixed, calendar-known period such as 12 months for monthly data or 7 days for daily data, and which sums to roughly zero over one full period so it carries no level of its own. `R` is the remainder, whatever is left once trend and season are removed. The Mauna Loa atmospheric CO2 record is the textbook picture: a steadily rising trend plus a clean 12-month cycle driven by the northern-hemisphere growing season, with a very small remainder. Decomposition is a descriptive tool for reading a series, not a forecasting model in itself.
go deeper
Be ready to name the three components and give one concrete example of each on a monthly series. Knowing that seasonality means a fixed, known period is the single fact most often checked here.
Expect to write the additive form, explain why the seasonal component is normalised over one full period, and say how the seasonally adjusted series is built and what it still contains.
Show that you read the remainder as a diagnostic: leftover structure means a mis-specified component, not noise. Be able to say why trend estimates near the end of the series are the least trustworthy ones.
Own the framing question of whether stakeholders should ever be shown a component split at all, given that different smoothers produce different splits of the same history and none of them is verifiable.
## What decomposition is for A time series usually mixes several movements that people want to talk about separately: "are we actually growing?", "is December always like this?", "was last week unusual?". Decomposition is the formal version of that separation. It expresses each observation `y` at time `t` as a combination of unobserved components that you estimate from the data. The additive form is ``` y_t = T_t + S_t + R_t ``` ## The three components **Trend-cycle (`T`).** The smooth, slowly-changing level of the series. It is deliberately not a straight line: a fitted regression line is one very restrictive trend model, whereas the decomposition trend is whatever a smoother (a moving average, a loess fit) says the local level is. It is called the *trend-cycle* rather than the trend because any long, irregular swing that is slower than the seasonal period also ends up inside it; classical decomposition has no way to split "steady growth" from "a multi-year upswing". **Seasonal (`S`).** A pattern that repeats at a **fixed, known period** tied to the calendar or clock: 12 for monthly data, 4 for quarterly, 7 for daily data with a weekly rhythm, 24 for hourly data with a daily rhythm. Two properties matter. First, the period must be known in advance, not discovered as "about eleven years". Second, the component is defined to average out to zero over one complete period in the additive case (or to average to one in the multiplicative case), so that it shifts observations around the trend without adding any level of its own. In the Mauna Loa CO2 record the seasonal component is a swing of a few parts per million: CO2 falls through the northern-hemisphere summer as plants take it up and rises again through winter, year after year, on top of the long climb in the level. **Remainder (`R`).** Everything the first two components did not explain. It is often labelled "irregular" or "error", but calling it error is misleading: it is a residual, not noise you have proved to be random. Real remainders routinely contain one-off shocks, effects you failed to model, and leftover structure. Reading the remainder is one of the most useful things you can do with a decomposition, precisely because anything systematic left in it is a component you have mis-specified. ## Seasonal adjustment The most common practical output is the **seasonally adjusted series**, `y - S` in the additive case. It answers the question "stripping out the fact that December is always big, did the underlying level move?" It still contains the remainder, so it is not smooth, and month-to-month wiggles in it are real variation rather than trend. ## Things to keep straight - **Decomposition is descriptive.** Splitting a history into components does not by itself forecast anything; you need a model that projects each component forward. - **The components are estimated, not observed.** Different smoothers give different splits of the same data. There is no ground-truth trend to check against. - **Trend estimates are weakest at the ends.** Every smoother needs data on both sides of a point, so the most recent observations, which are the ones you care about, have the least reliable trend estimate. - **The seasonal component may not be constant.** Classical decomposition assumes one repeated seasonal shape for the whole history; a business whose weekly rhythm changed after a product launch violates that assumption, and the violation shows up as structure in the remainder. ## Vocabulary that interviewers listen for Saying "trend, seasonal, remainder" with the definitions above, distinguishing *trend-cycle* from a fitted line, and noting that seasonality requires a fixed known period is usually enough to pass this question. Adding that the seasonal component is normalised over one period, and that the seasonally adjusted series keeps the remainder, marks you out as someone who has actually looked at these plots rather than read about them.
- What is a seasonally adjusted series, and when would you show one?It is the observed series with the seasonal component removed: `y - S` additively, or `y / S` multiplicatively. You show it when the question is about the underlying level rather than the calendar, for example "did sign-ups really grow this month, or is December always like this?". Note it still contains the remainder, so it is not a smooth curve and its short-run wiggles are genuine variation.
- Why is the trend component usually called the trend-cycle?Because the smoother that estimates it is tuned to remove movements at the seasonal period, and everything slower passes through. A steady climb and a long irregular swing both end up in the same component, and classical decomposition offers no way to separate them. Calling it trend-cycle is an honest admission that the piece contains both.
- What should you do if the remainder shows an obvious repeating pattern?Treat it as a specification failure, not as noise. Systematic structure left in the remainder means a component is wrong: the seasonal period may be misspecified, the seasonal shape may have changed over the history, calendar effects may be unmodelled, or the trend smoother may be too stiff or too flexible. Fix the component rather than explaining the residual away.
Like separating a recording into a bass line, a repeating drum loop and everything else: the loop is the same every bar, the bass drifts slowly, and the rest is what makes each bar different.
saying these in an interview costs you the question
- Calls the remainder error and assumes it is white noise
- Describes the trend component as a fitted straight line
- Thinks a decomposition is itself a forecasting model
- Says any repeating pattern counts as seasonality
- Forgets the seasonally adjusted series still contains the remainder