- Holt–Winters extends smoothing to level, trend, and a seasonal pattern.
- Choose additive or multiplicative seasonality depending on whether seasonal changes are fixed amounts or proportional to the level.
| Concept | What it means | Northstar example |
|---|---|---|
| Each seasonal position gets a state. | For weekly daily data, the method maintains weekday effects and updates them as new data arrive. | A weekly Holt–Winters model maintains separate weekday effects for Northstar's Monday-through-Sunday shopping rhythm. |
| A single period has limits. |
| A seven-day seasonal model cannot also be assumed to know Northstar's annual holiday ramp and moving Black Friday event. |
| Scale is a modeling choice. |
|
|
Add level, trend, and weekly states
The method updates a level and trend, then learns one seasonal state for each weekday. Add stages one at a time and compare the holdout forecast. Additive seasonality adds a fixed dollar pattern; multiplicative seasonality adds a percentage that scales with the level.
The question it answers: What does each Holt–Winters stage add to Northstar's forecast: a level, a trend, or a weekly rhythm?
Learned weekday effects · additive
3 · Add weekly season. The forecast adds the learned weekday state to the level and trend. The weekly pattern is learned only from past observations.
Weekly states active. The supplied recursions use β = .08 and γ = .25 to make each stage distinct.
Fine print & method
Constructed daily sales with a known upward level and a repeating weekly pattern; the last 14 days are a holdout. The recursions update level, trend, and seven seasonal states using fixed α, β, and γ by stage. No parameters are optimized; this is a teaching implementation, not a fitted ETS model. Additive and multiplicative data are generated separately, so the model form matches the data. A fixed weekly period does not automatically model annual events such as Black Friday; those need explicit event terms.