STOP 24 / 31Separate the rhythms. Forecast what remains.
Layout
STOP 24 · More models

Separate the rhythms. Forecast what remains.

EXPLORE~8 min
Back to route map
  • Northstar has a weekly shopping rhythm and a smoother yearly pattern.
  • MSTL—Multiple Seasonal-Trend decomposition using Loess—estimates several seasonal components.
  • Pair it with a forecasting model for the seasonally adjusted series, then put the seasonal forecasts back.
  • The result is a forecasting pipeline, not a single ETS model with several seasonal periods.
MSTL + ETS: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
1. Decompose using training data only.
  • Specify the seasonal periods you have enough history to identify.
  • MSTL iteratively estimates seasonal components, a trend, and a remainder.
  • Seasonal smoothing windows control how much the historical pattern may evolve; robust options can limit the influence of unusual observations.

At each Northstar backtest cutoff, estimate weekly and annual components using only the available training history—not the later evaluation dates.

2. Forecast the seasonally adjusted series with nonseasonal ETS.
  • Subtract all estimated seasonal components from sales.
  • The adjusted series is trend plus remainder—not the remainder alone.
  • Fit an ETS model without seasonal states to it; the level, trend, damping, and error choices still need estimation and validation.

After removing seasons, Northstar still forecasts the underlying sales level/trend plus remainder, rather than trying to forecast the remainder alone.

3. Project the seasonal components and recombine.
  • A common implementation forecasts each seasonal component with seasonal naïve: repeat its most recent estimated cycle.
  • Add those forecasts to the ETS forecast.
  • Historical seasonal shapes may evolve, but repeating their final cycles does not predict new future changes in shape.
  • If you used a log or Box–Cox scale, back-transform carefully; means may need bias adjustment.

Repeat the latest estimated weekly and annual component cycles, forecast the adjusted series, and add the pieces back on a consistent scale.

4. Keep calendar effects and uncertainty honest.
  • An annual period of 365 does not automatically handle leap years or Black Friday’s moving date.
  • Fit event adjustments using past data and add their future effects separately, or compare a calendar-regression approach.
  • Refit the entire decomposition inside each backtest window.
  • Intervals based only on the adjusted-series forecast can omit seasonal and decomposition uncertainty; check empirical coverage.
  • A repeated annual slot will not always land on Black Friday.
  • Northstar handles the moving event explicitly and tests the final recombined intervals.
adjustedₜ = salesₜ − weeklyₜ − yearlyₜ forecast sales = ETS(adjusted) forecast + projected weekly + projected yearly (+ separately modeled event)
FIG 24.1

Remove, forecast, project, recombine

WHAT THIS SHOWS

Follow the separate, forecast, project, and recombine stages using known simulator components.

The question it answers: What does the adjusted series contain, and how are the seasonal forecasts put back?

Schematic only: known simulator components and a supplied trend projection. This is not an estimated MSTL + ETS fit or an accuracy comparison.

1. Remove seasons

Subtract the weekly and yearly components. What remains is the seasonally adjusted series: trend plus remainder. It still contains noise. Toggle each component to see what is left when a seasonal pattern is not removed.

Schematic MSTL and ETS pipeline$29k$34k$38k$43kTraining endsTraining historyForecast origin+42 days
Observed training salesAfter selected seasonal removals

The example excludes Black Friday to isolate the seasonal pipeline. Neither a repeated annual slot nor MSTL itself solves a moving event calendar. Revisit event handling →

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

After removing the weekly and yearly components, what should the ETS model forecast?

Your learning note

Private · just for you