STOP 23 / 31An approachable toolkit, not an oracle.
Layout
STOP 23 · More models

An approachable toolkit, not an oracle.

EXPLORE~6 min
Back to route map
  • Prophet combines a trend, seasonal components, holiday effects, and optional regressors.
  • It can be convenient for business time series, but its assumptions still need testing.
Prophet & changepoints: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
Trend can change slope.
  • Changepoints let a piecewise trend bend.
  • Their flexibility is regularized; too much flexibility chases noise, too little misses real changes.

A Northstar market launch may change the growth slope; overly flexible changepoints can also chase a short promotion spike.

Calendar knowledge helps.
  • Fourier seasonality and explicit holiday windows can represent recurring behavior.
  • Add Black Friday to the holiday calendar rather than expecting a smooth yearly wave to find it perfectly.

Give Prophet Northstar's actual Black Friday event dates and windows instead of assuming smooth annual seasonality will identify them exactly.

Treat intervals critically.
  • Uncertainty depends on the model and how future trend changes are assumed to behave.
  • Evaluate empirical coverage; do not treat displayed bounds as guarantees.

Northstar evaluates whether Prophet's intervals cover held-out outcomes, especially when future trend behavior differs from fitted history.

FIG 23.1

Zoom out. Does the story change?

WHAT THIS SHOWS

Select how much Northstar history the forecasting rule can use. The fitted forecast changes with that selection; revealing the future does not refit it.

The question it answers: Does the recent rise represent lasting growth or a repeating shopping season?

History available

Fitted short-window trend only · 3 training months

Northstar forecast fitted only to the selected monthly history30.0240.1450.1660.2Nov 2025 · forecast originJan 2023Jan 2024Jan 2025Jan 2026
Earlier context · not used by this fitSelected training historyForecast fitted to selected history
January forecast · sales index266.0
Training months3
Future used to fitNone
Fine print & method

Constructed monthly sales index, not dollars. Three months fit a log-linear trend; twelve months use seasonal naïve because a single cycle cannot separately identify annual seasonality and trend. With at least two cycles, ordinary least squares fits log sales to time and calendar-month indicators using only the selected training rows. The red reference deliberately retains the three-month fit. More data helps this specified model, not every possible forecasting algorithm.

CHECK YOURSELF · CHOOSE AN ANSWER

Prophet fits historical sales very closely. What should happen next?

Your learning note

Private · just for you