Teaching Economics to the Machines

Abstract

Structural models in economics often suffer from a poor fit with the data and demonstrate suboptimal forecasting performances. Machine learning models, in contrast, offer rich flexibility but are prone to overfitting and struggle to generalize beyond the confines of training data. We propose a transfer learning framework that incorporates economic restrictions from a structural model into a machine learning model. Specifically, we first construct a neural network representation of the structural model by training on the synthetic data generated by the structural model and then fine-tune the network using empirical data. When applied to option pricing, the transfer learning model significantly outperforms the structural model, a conventional deep neural network, and several alternative approaches for bringing in economic restrictions. The out-performance is more significant i) when the sample size of empirical data is small, ii) when market conditions change relative to the training data, or iii) when the degree of model misspecification is likely to be low. 

Keywords: transfer learning, structural model, deep neural networks, option pricing

JEL Classification: C01, G12