Estimating Parameters of Structural Models Using Neural Networks

Abstract

Machine learning tools such as neural networks see increasing applications in marketing and economics for predictive tasks, such as classifying images and forecasting choices. Instead of these predictive tasks, we explore using neural nets to estimate the parameter values for an economic model. The neural net is trained with model-generated datasets. Through training, the neural net learns a direct mapping from (the moments of) a dataset to the parameter values under which the dataset is generated. We show this Neural Net Estimator (NNE) converges to Bayesian parameter posterior when the number of training datasets is sufficiently large. We examine the performance of NNE in two Monte Carlo studies. NNE incurs substantially smaller simulation costs compared to simulated MLE and GMM, while achieving no worse estimation accuracy. NNE is also easy to implement with the wide availability of neural net training packages.