Trend, Seasonality and Noise

Splitting a series into direction, repeating pattern and leftover.

Also known as: time series decomposition, trend-seasonality-noise, decomposing a series

Every time series can be read as three layers added together: the slow trend (direction), the repeating seasonality (calendar pattern), and the leftover noise (everything else). Splitting a series this way — decomposition — turns “sales went up” into three answerable questions: did the direction change, was it the season, or was it a one-off?

observed  =  trend  +  season  +  noise
  120     =    100   +    25    +   (-5)

In practice the split is approximate: smooth the series for the trend, average each calendar position for the season, and whatever remains is noise. The value is not precision but attribution — each layer gets judged on its own terms instead of all three arguing inside one number.

The classic mistakes:

Reading the raw series as one story. A December record with a flat trend is a seasonal success and a strategic flatline at once. Decompose before judging.
Treating noise as signal. A single odd week is usually leftover, not a break. Reacting to noise with strategy changes is expensive churn.
Assuming the layers are independent. Promotions (noise by design) can pull forward seasonal demand; a trend break can look like a seasonal shift. The split organises thinking; it does not prove the layers never interact.
Over-smoothing the trend. An over-aggressive smoother bends the trend toward the season and hides real breaks. Keep the trend flexible enough to turn.

When to go further: models like exponential smoothing and ARIMA formalise this split to forecast. For reading and explaining a series, the informal three-layer split is usually enough.