- Before a complicated model earns its place, it has to beat a simple forecast on unseen data.
- Northstar can start with the latest day, the same weekday last week, or a carefully aligned period last year.
| Baseline | Forecast = | Assumes | Northstar example |
|---|---|---|---|
| Mean | Average of the training history | A roughly stable level; no trend or seasonal adjustment | Northstar forecasts a quiet, stable category using its historical average sales. |
| Naïve | Last observed value | The latest level is useful; the last shock is carried forward | Tomorrow's Northstar sales equal today's, even if today contained an unusual promotion. |
| Seasonal naïve | Same position in the latest season | The seasonal pattern repeats, without added growth |
|
| Drift | Last value + horizon × average change per period | The endpoint-to-endpoint average slope continues | Extend Northstar's average historical daily dollar change; this does not automatically reproduce weekly shopping patterns. |
| Seasonal naïve × growth | Aligned prior-season value × a growth factor | The seasonal pattern repeats and the growth assumption holds |
|
Give a simple forecast a fair test
Each simple rule uses only Northstar's pre-origin history. Compare its forecast with the same 56 unseen days and compute genuine holdout errors.
The question it answers: Does a more complex candidate earn its place against an honest baseline?
All baseline inputs come from before the forecast origin. The 18% growth factor is a supplied scenario assumption, not learned from holdout sales. This window illustrates ordinary-day performance; test event windows separately.
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.