liveBook · Manning

Asymmetric cost function

As explained in Machine Learning System Design

A cost function that assigns different penalties to different types or directions of error.

An asymmetric cost function makes some errors more costly than others. In retail demand forecasting, underpredicting demand can cause out-of-stock events, missed profit, customer dissatisfaction, and lost customers, while overprediction mainly risks expired goods. section 9.4.1

Use the asymmetry to interpret residuals and deployment risk:

  • Prefer slight overprediction to underprediction, but treat excessive overprediction as harmful rather than automatically desirable. section 9.4.1
  • Inspect residual distributions, average bias, and spread together; a small positive bias with low variance is the preferred outcome. section 9.4.1
  • Treat negative skew as a deployment red flag because it indicates a tendency toward underprediction and missed demand. section 9.4.1

Retail stock management

For a store-item forecast, predicting 120 units when demand is 100 may create waste, but predicting 80 when demand is 100 can create an out-of-stock event. The preferred model is therefore mildly optimistic, not maximally high. section 9.4.1

How Machine Learning System Design covers it

How deep it goes
Applies cost asymmetry to residual bias, spread, and skew but stops short of formal mathematical derivation.
Examples
Uses overstock, expired goods, out-of-stock events, missed profit, and customer dissatisfaction as contrasting error consequences.
What it assumes
Assumes demand forecasting and residuals, while teaching why business impact makes their preferred direction unequal.
What you can do after
Diagnose demand-forecast residuals and flag negatively skewed models before deployment.
How it connects
Connects demand forecasting with residual analysis, overprediction, underprediction, and stock management.
Cover of Machine Learning System Design
liveBookmanning.com
Read Machine Learning System Design on liveBook Get the book at manning.com

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.