Towards predicting neural net control of macro-econometric multi-compartment models

摘要

In this paper we propose a novel concept for analyzing dynamic economic and financial systems by multi-compartment modeling techniques in combination with neural networks. This is done as a two-step process: first, a multi-layer perceptron(MLP{sub}(ident)) is introduced to approximate the dynamics of observable variables of a single (domestic) compartment and to identify interdependencies; second, a multilayer perceptron (MLP{sub}(control)) evaluates series of outputs from MLP{sub}(ident)to make a complex decision about certain properties (e.g. stability) and to find a classification of economic or financial scenarios. Simulation results confirm the ability of MLP{sub}(ident) to learn the structure and parameters of the domesticcompartment and to provide good results in out-of-sample tests. Simulation results for MLP{sub}(control) show that it is able to decide whether a current parameter set leads to a stable or instable constellation in the future. Possible applications ofthis novel concept are analysis of interdependencies of economic variables between multiple countries, or control of dynamic system behaviour with regard to predicting properties.