| 摘要 |
A method and system for forecasting demand for order configurations are provided. The method and system, in one aspect, expresses attach rates within a family of n options as a set of n positive numbers that sum to 1. By applying suitable transformations to the attach rates, they are modeled as a random vector in (n−1)-dimensional Euclidean space. The distribution of the transformed attach rates are modeled as a distribution family specified by a location vector and a dispersion matrix. The dispersion matrix is simplified, for example, using historical data or expert judgment or both to identify option families that have dependent demand. Simplifying may also include expressing dependence between options by a simple model that involves few parameters. Location vector is estimated by computing point forecasts of transformed attach rates. The parameters of the dispersion matrix are estimated by calibration on historical data, using the dispersion of the errors in historical point forecasts. |
