liveBook · Manning

Intermittent demand

As explained in Machine Learning System Design

Demand that occurs irregularly, with periods of zero demand interspersed among nonzero observations.

Intermittent demand occurs irregularly, with zero-demand periods between nonzero observations. Forecast evaluation must account for this pattern rather than rely on generic error metrics. section 5.3.1

Metric choices require specific cautions:

  • Avoid MAPE when any future realization is zero; it becomes undefined. section 5.3.1
  • Treat sMAPE cautiously because a zero actual contributes 2 regardless of the point forecast. section 5.3.1
  • Question MSE when unusually high forecast errors could dominate evaluation. section 5.3.1
  • Check relative errors for undefined values when a benchmark has zero error. section 5.3.1
  • Consider MSR, MAR, or sMSE, while making any short-term weighting explicit. section 5.3.1

How Machine Learning System Design covers it

How deep it goes
Contrasts percentage, squared, relative, rate-based, and scaled errors, but does not develop their mathematics.
Examples
Uses zero actuals, zero benchmark errors, and unusually high forecast errors as diagnostic cases.
What you can do after
Evaluate intermittent-demand forecasts while recognizing undefined values, disproportionate penalties, and implicit time weighting.
How it connects
Connects intermittent demand with count data, MAPE, sMAPE, MSE, relative errors, MSR, MAR, and sMSE.
Cover of Machine Learning System Design
liveBookmanning.com
Read Machine Learning System Design on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.