Abstract: The invention provides a computer-implemented method of training an approximation architecture for a forecasting system including performing a pre-training stage comprising steps of: obtaining training samples each including an input value for a first plurality of input variables and corresponding parameter values of the functions; obtaining dependencies between different variables of the first plurality of input variables; determining, based on the obtained dependencies, dimensionality-reducing rules for determining a second plurality of input variables, wherein there are fewer degrees of freedom in the second plurality than in the first plurality; and, determining, by applying the dimensionality-reducing rules to the training samples, modified training samples including input values for at least some of the second plurality of variables and corresponding parameter values approximating the functions.