248 S. Stijven et al.
Fig. 6 The constituent models of an ensemble will diverge when asked to extrapolate. This is
illustrated for a 3-variable model ensemble consisting of 14 unique models
on both contemporaneous and dynamic relationships. The last option assumes using
lagged inputs up to an expected maximum lag. The selected nonlinear transforms
from the GP-generation phase are used in ARIMAX model generation but they
might not be selected in the final model if they are statistically insignificant. If
distributed lags of a statistically significant nonlinear transform exist, it is repre-
sented by the corresponding transfer function in the ARIMAX model. As a result,
GP complements ARIMAX models with adding contemporaneous and dynamic
nonlinear explanatory variables and the final models are with all the benefits of
this well-known forecasting approach, such as building multi-step forecasts of all
inputs, statistically defined confidence limits, and available software.
An important area of business forecasting with big economic impact is raw mate-
rials prices forecasting where 3-to-6 months forecasts are critical in high-volume
price negotiation. A recent application of large-scale raw materials forecasting in
the chemical industry is discussed in (Kordon 2012 ). An example of applying
the hybrid system in two typical cases in raw materials forecasting: (1) when the
relationships between the forecasted variable and the related economic drivers are