STOP 29 / 31More flexibility. Same rules.
Layout
STOP 29 · In practice

More flexibility. Same rules.

EXPLORE~6 min
Back to route map
  • Tree models, neural networks, and ensembles can learn nonlinear relationships and share information across products.
  • They still need sensible features, honest evaluation, and reliable future inputs.
Machine learning & combinations: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
Supervised learning needs time-aware features.
  • Lagged sales, rolling past-only statistics, known calendars, and planned promotions can become predictors.
  • Build multi-step strategies deliberately: recursive and direct forecasts behave differently.
  • Northstar can use past-only lags, known holidays, and planned promotions.
  • A direct 28-day forecast and repeated one-day recursion need separate evaluation.
Global models can share information.

Learning across many related products can help sparse series, provided the evaluation reflects new products and real deployment conditions.

A model trained across Northstar's related products may help a sparse category, provided the test includes the kinds of new or changing products it will actually face.

Combinations are strong contenders.
  • Averaging diverse forecasts can reduce model-specific errors.
  • A simple average is a useful baseline; learn weights only from appropriate training or validation data.

Average sensible Northstar forecasts as a benchmark; learn combination weights from training/validation, never the future period used to report the final score.

FIG 29.1

Would you know it at forecast time?

WHAT THIS SHOWS

Inspect which candidate Northstar features were genuinely available before the forecast origin.

The question it answers: Does this predictor cross the information boundary?

Tap a feature to check whether it crosses the information boundary.

Availability is a timestamped fact, not just a column name. Store what was actually known at each forecast origin.

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.

CHECK YOURSELF · CHOOSE AN ANSWER

Why might averaging several sensible forecasts help?

Your learning note

Private · just for you