Estimating multivariate conditional distributions via neural networks and global optimization

A new concept for modelling and forecasting is introduced. The maximum likelihood principle is used to identify the multivariate distribution of forecast variables, conditional on given attributes of forecast context. The distribution parameters are conditional on input features such as properties of the product. The conditional distribution parameters are estimated by a global optimization method, using neural networks for functional approximation. The goal is to construct a general attribute-based forecast model, which can be applied to novel cases with new attribute combinations. The information about a complete distribution of forecasts can be used to quantify the reliability of the forecast. The reliability information is particularly useful for decision support, e.g. if the forecast error causes strongly asymmetric costs. This is illustrated on a case study concerning the spare parts demand forecast.