Constructing Efficient Simulated Moments Using Temporal Convolutional Networks

Abstract

We propose a method to estimate model parameters using temporal convolutional networks (TCNs). By training the TCN on simulated data, we learn the mapping from sample data to the model parameters that were used to generate this data. This mapping can then be used to define exactly identifying moment conditions for the method of simulated moments (MSM) in a purely data-driven manner, alleviating a researcher from the need to specify and select moment conditions. Using several test models, we show by example that this proposal can outperform the maximum likelihood estimator, according to several metrics, for small and moderate sample sizes, and that this result is not simply due to bias correction. To illustrate our proposed method, we apply it to estimate a jump-diffusion model for a financial series.