Agent-based Model Parameters Estimation with Wasserstein Distance
This is a joint work with Sebastiano Manzan.
The goal of this paper is to propose an estimation method that is accurate when the interest is to estimate the parameters of high-dimensional models, such as DSGE and ABM. In particular, we propose a Simulated Minimum Distance (SMD) estimator based on the Wasserstein distance between the model-simulated distribution and the empirical distribution. There are several advantages of this methodology. First, the SMD does not require costly and difficult approximations of the likelihood function. Second, the Wasserstein distance is more stable, and it is defined also for distributions with no overlapping support. We first conduct comparison experiments between the Wasserstein distance estimator and several popular agent-based model estimation methods, on the Brock and Hommes (1998) asset pricing model and a set of classical time series models. The results show the verified statistical properties of the estimator and its significantly higher accuracy relative to established estimation methods such as Bayesian estimation, Simulated Method of Moments, and other novel information theory-based estimation methods. Finally, we estimate the Brock and Hommes model to financial data and discuss the results relative to the existing literature.
I presented this paper at 30th Symposium of the Society for Nonlinear Dynamics & Econometrics, Orlando, March 16th, 2023, and 28th Computing in Economics and Finance Conference, Dallas, June 19th, 2022