STOP 04 / 31Pull the line apart.
Layout
STOP 04 · Foundations

Pull the line apart.

EXPLORE~7 min
Back to route map
  • Decomposition asks: what contributed to this observation?
  • It splits the past into a trend-cycle, seasonal patterns, and a remainder.
  • It describes history; it is not, by itself, a forecast.
Decomposition: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
Start with a simple smoother.
  • A centered 7-day average suppresses a repeating weekly pattern.
  • It needs three days on either side, so the first and last three trend values are unavailable.

A centered seven-day average can reveal Northstar's slower-moving level, but Wednesday's value uses days through Saturday, so it is not a real-time feature for Wednesday.

Ask for the patterns you need.
  • A weekly decomposition will not magically remove annual holiday effects.
  • STL handles one seasonal period; MSTL can handle several.
  • A moving Black Friday event still benefits from an explicit regressor.

Remove weekly and yearly patterns separately; add Black Friday indicators rather than expecting a weekly decomposition to explain a moving annual promotion.

The remainder is not automatically noise.
  • A holiday spike left in the remainder may be an omitted predictable event.
  • Look for structure before dismissing it as random.
  • A recurring Cyber Monday spike in the remainder suggests a missed event effect.
  • It is unexplained by this model, not necessarily unpredictable.
sales = trend-cycle + seasonal components + remainder
FIG 04.1

What does a weekly smoother miss?

WHAT THIS SHOWS

Compare a real centered weekly average with the simulator's known components, then include or omit the known event adjustment.

The question it answers: What structure can remain when a weekly smoother misses an annual event?

Interactive time series chart$11k$44k$76k$109kJan ’23Jul ’23Jan ’24Jul ’24Jan ’25Jul ’25Dec ’25
Observed salesCentered 7-day mean
TRUE TREND · SIMULATOR REFERENCE
Interactive time series chart$22k$29k$35k$41k
Known underlying level
AFTER REMOVING KNOWN COMPONENTS
Interactive time series chart−$9k$15k$39k$63k
Unexplained event + noise

The 7-day average is an actual centered smoother, not a full STL decomposition. Lower plots use oracle components from the simulator, not estimated components. The contrast shows why omitting a predictable event leaves structure in the remainder.

Fine print & assumptions

Constructed teaching data and illustrative calculations, not a production forecasting service. Read the diagram's method notes and distinguish supplied simulator components from fitted estimates.

CHECK YOURSELF · CHOOSE AN ANSWER

A weekly-only decomposition leaves a huge Black Friday spike. Why?

Your learning note

Private · just for you