A Pre-Trained Estimator for Consumer Search Model

Abstract

We explore pre-training estimators for structural econometric models. The estimator is "pre-trained" in the sense that most computations and effort occur once during its construction. Subsequent applications of it to estimate different datasets require little computation costs or effort. The estimation relies on a neural net to recognize structural model's parameter from data patterns. This paper focuses on a sequential search model that is known to be hard to estimate. We evaluate our pre-trained estimator on 11 real datasets. The estimation takes seconds to run and shows high accuracy. We provide it at pnnehome.github.io. More generally, pre-trained estimators make structural models more accessible and easier to apply. They can also facilitate privacy-preserving estimation because they need only aggregate data patterns.