Revenue is income generated from selling goods or services before expenses. In applied machine learning, it connects forecast behavior to business value: a model with similar offline error can produce different revenue when its bias changes pricing or inventory decisions. section 12.1.2
Revenue becomes useful when teams translate model performance into operational outcomes:
- Compare models by the money their predictions enable, not only by WAPE or bias. In airline ticket forecasting, an overpredicting model earned more because it reduced discounted-ticket overbooking while still limiting unsold seats. section 12.1.2
- Use a business proxy when direct revenue measurement is inconvenient. In the retail experiment, average check served as the primary proxy because it correlated with revenue while the number of checks remained stable. section 12.4.1
- Pair revenue with guardrail and supporting measures, including checks per day, daily profit, model update frequency, and offline metrics. section 12.4.1
Airline pricing example
Across four days, Model A and Model B had the same 15% WAPE but opposite 10% biases. Using the lower of sold and forecast tickets, Model A generated $75,300 and Model B generated $83,900. The $8,600 difference for one flight illustrates why revenue can distinguish operationally useful forecasts that offline metrics treat as equal. section 12.1.2
