- Northstar doesn’t just want an estimate for next month.
- Its inventory team needs to know the plausible upside and downside, especially over the Black Friday period.
| Source | What it means | Northstar example |
|---|---|---|
| Future randomness | Outcomes fluctuate even when the model and its parameters are correct. | Northstar's individual buyer decisions can make actual daily sales differ from the conditional mean. |
| Estimated parameters | Trend and seasonal effects are learned with uncertainty, not known perfectly. | A short Northstar history may not pin down next year's holiday lift precisely. |
| Model or structural uncertainty | Omitted mechanisms and future business changes may not be captured by a model's default intervals. | Northstar may face a new competitor, changing assortment, or different customer behavior after the model was fitted. |
The further ahead, the wider the range
Each dot is one independent simulated forecast case. Green shading is the stated prediction interval; red dots missed it. Change the nominal level or the assumed uncertainty, then count actual coverage.
The question it answers: Does Northstar's stated interval cover unseen outcomes as often as it claims?
Fine print & method
The generator supplies independent normal errors with known, horizon-dependent variance. An uncertainty multiplier of 100% uses that known variance; other values deliberately misstate it. Actual finite-sample coverage need not equal the nominal percentage. The displayed reference uses a rough two-standard-error binomial comparison, not proof of calibration. This is not one correlated future sales path or a fitted production interval. Wider is not automatically better; assess both coverage and sharpness across honest backtests.