Work detail

Application of Machine Learning to the Simulated Method of Moments in Financial Agent-Based Models Estimation

Author: Bc. Eric Žíla
Year: 2021 - summer
Leaders: PhDr. Jiří Kukačka Ph.D.
Work type: Bachelors
Language: English
Pages: 72
Awards and prizes: Nomination Deloitte Outstanding Thesis Award.
Abstract: We build a simple machine learning extension on top of the simulated method of
moments, an estimation technique commonly applied to analytically intractable
complex agent-based models. It offers a new approach to the selection of the
moment set. The idea beyond the extension follows the stepwise regression
framework. Two algorithms are proposed, the forward stepwise moment selection and the backward stepwise moment elimination. The methodology is
tested on four models with increasing complexity. We use three standard financial models to test the elementary functionality of the algorithms and one
popular financial agent-based model. We employ three sets of moments from
the literature as benchmarks. We find that for every model, both selection
processes consistently identify multiple moment sets that outperform all benchmark sets. In line with our expectations, the estimation improvement is minor
for the three simple models. However, for the complex agent-based model, for
which one needs to deal with moment set selection in practice, we achieve a
severe performance boost.




Patria Finance