EcoSta 2021: Start Registration
View Submission - EcoSta2021
A0647
Title: Optimal moment-subset selection for the simulated method of moments using machine learning Authors:  Jiri Kukacka - UTIA AV CR, v.v.i. (Czech Republic) [presenting]
Abstract: Simulation-based estimation inference is expanded via machine learning techniques. The setup of the simulated method of moments (SMM) is extended with an automated selection of the optimal set of moments. To briefly demonstrate the importance of the issue, a relatively rich set of nine moments has been previously employed, while for a similar type of a financial agent-based model, four but also fifteen moments have been used. An insufficient set of moments will likely ignore some important dynamic properties of the model while overfilling the moment set will likely lead to estimation inefficiencies and problems with identifying parameters. Algorithmic subset selection methods generally developed for model construction are thus utilized. The methods evaluate subsets of features, moments in our case, in terms of their suitability for a given purpose and retain only the optimal ones. We conduct an extensive comparative study of the accuracy and computational complexity of the proposed machine learning extension of the SMM w.r.t. the original method and the simulated maximum likelihood method. As a laboratory, we take advantage of the New Keynesian macroeconomic model under rational expectations and various behavioral heuristics.