- A predictor is useful only if it is available when the forecast is made.
- Realized ad clicks, future stockouts, or next week’s actual conversion rate can make a backtest look magical—and fail in production.
| Driver type | What it means | Northstar example |
|---|---|---|
| Known in advance | Calendar events and plans recorded at the forecast origin may be legitimate inputs. |
|
| Must be forecast | Unknown future inputs add their own uncertainty to the final sales prediction. | Northstar can use a visitor-traffic forecast or traffic scenarios, not future realized visitors hidden inside a backtest. |
| Future-outcome leakage | Realized future outcomes and preprocessing that uses later observations cross the information boundary. | Using actual future ad clicks or a centered average that needs future sales can flatter Northstar's test score without producing a usable forecast. |
Would you know it at forecast time?
Northstar's daily orders are related to visitor traffic. A regression can use forecast visitors, but it cannot know the realized visitors of future days. Compare honest supplied visitor forecasts with the deliberately leaky alternative.
The question it answers: How does uncertainty in an outside driver enter Northstar's sales forecast?
Training: visitors and daily orders
Next 14 days · driver uncertainty included
Fine print & method
Regression coefficients and residual spread use only the first 120 observations. Future driver forecasts are supplied synthetic forecasts with controlled errors, not a fitted traffic forecasting service. Realized future visitors are used only in the explicitly marked leakage option and for evaluation. Bands combine residual variance and supplied driver-error variance under a normal linear approximation; they are not empirically calibrated Northstar business intervals or causal estimates.