A predictive distribution represents the possible future demand values for a specific item, store, and period, rather than reporting only one point forecast. section 5.3.1
For low-volume count data, the distribution is often asymmetric, so optimizing for the median can produce biased forecasts when a mean-oriented decision is needed. section 5.3.1
Use the distribution to:
- Represent discrete future outcomes for supply-chain forecasting and inventory control. section 5.3.1
- Assess whether point forecasts and metrics such as MAE, wMAPE, or MAPE fit the decision. section 5.3.1
- Avoid MAPE when any future realization can be zero, because it becomes undefined. section 5.3.1
