CoPS/IMPACT Working Paper Number G-348

Title: Neural-Network approximation of reduced forms for CGE models explained by elementary examples

Authors: Peter B. Dixon, Maureen T. Rimmer and Florian Schiffmann

Abstract

Neural Network (NN) theory provides a powerful method for approximating the reduced form of a large-scale multi-regional CGE model. However, NN methods are relatively unknown by CGE modellers. We set out the theory of the NN approximation method and demonstrate how it works with simple examples.

The paper is motivated by a project for a client with limited in-house CGE capabilities but requiring the ability to obtain CGE solutions at short notice in a confidential environment. We describe how an NN approximation meets the client's needs. The NN approximation is more accurate and broadly applicable than earlier approaches that CGE modellers have used based on regression equations and matrices of elasticities.

JEL classification: C45, C68

Keywords: Neural network method explained; Neural network approximations to reduced forms; Multi-regional computable general equilibrium models

Working Paper Number G-348 can be downloaded in PDF format. To print this you will need the Adobe Acrobat Reader.

Go to working papers page