Surrogate modelling of a recursive-dynamic single country computable general equilibrium model

Abstract

We design, detail and test an approach which allows optimizing policy instruments to maximize welfare over results emerging from a Computable General Equilibrium (CGE) model. To technically enable policy optimization over a CGE model, we train a deep learning-based surrogate model that approximates the behavior of the CGE model. The final policy optimizing is then subject to the surrogate model. To show case our approach, we optimize the timing of emission abatement and a carbon tax recycling strategy for Tanzania from 2018 to 2050 with a detailed recursive-dynamic single-country CGE model. Our proposed approach and the provided code to train surrogate models for CGE models is quite generic, allowing its application to differently structured CGE models and associated policy instruments. Even though generation of the observation sample is computing time intensive, the surrogate model enables policy optimization not possible by using the CGE model directly. The trained neural network replicates the simulation behavior of the CGE model quite accurately with an average R² of 99.99% over the outputs. Besides optimization, such a surrogate model representing the key input-output relations of a CGE model could also be easily integrated into other modelling frameworks.