High-Dimensional Dynamic Stochastic Economic Modeling using Adaptive Sparse Grids

The Faculty of Informatics is pleased to announce a seminar given by Simon Scheidegger

DATE: Tuesday, May 27th 2014
PLACE: USI Lugano Campus, room SI-004, Informatics building (Via G. Buffi 13)
TIME: 16.30

ABSTRACT:
Considering the economy as a system of many interrelated markets where a large number of individuals interact to exchange goods and services is a central theme of modern economics, going back to Arrow, Debreu and Radner. In the last 2 decades researchers have used versions of this model to examine the macro-economy. However, due to the heterogeneity across different consumers, workers, households, firms, sectors, or countries one easily ends up with versions of the model that are thought to be unsolvable.
Model-based economics has for the most part reacted to this challenge in two ways. Either by focusing on qualitative results obtained from extremely simplified models with little heterogeneity, or by only locally solving the equation systems that describe the dynamics around a so-called steady state. In contrast, solving for the global solution of a model with substantial heterogeneity is very costly: the computation time and storage requirements increase dramatically with the amount of heterogeneity, i.e. with the dimensionality of the problem. It is therefore often far beyond the scope of current methods to include as much heterogeneity as a natural modeling choice would suggest: for instance, a yearly calibration in an overlapping generations (OLG) model, or an international real business cycle (IRBC) model that includes all economically important countries. Building on Brumm & Scheidegger 2013 (submitted, under revision), we aim to use modern numerical methods and cutting-edge supercomputing facilities to compute global solutions of high-dimensional dynamic stochastic economic models in a way that fits their generic structure. No matter whether these are solved by iterating on a value function (parametric dynamic programming) or on the functions that represent economic choices (time-iteration), the computational challenge is similar:
i)  In each iteration step, an economic function needs to be approximated. For this purpose, the function value has to be determined at many points in the high-dimensional state space, and
ii) each point involves solving a  high-dimensional maximization problem (for dynamic programming) or a system of nonlinear equations (for time-iteration).
To make this problem solvable within reasonable time scales, i.e. hours or days, we minimize both the number of points to be evaluated and the time needed for each evaluation. For the first purpose (i) we use adaptive sparse grids, while the second task  (ii) is achieved by a hybrid parallelization scheme that minimizes interprocess communication and offloads the function evaluations to accelerators. This scheme enables us to make efficient use of modern hybrid high-performance computing facilities such as CSCS's Piz Daint Cray XC30. Our framework thus offers the promise to economic modelers of being able to solve models that include much more heterogeneity than was previously possible.

BIO:
Dr. Simon Scheidegger is a Postdoc at the Department of Banking and Finance in the Group  of F. Kübler, where he works since 2012. In 2010, he obtained his PhD in theoretical physics at the University of Basel (Supervisors: Prof. Dr. M. Liebendörfer, Prof. Dr. F-K. Thielemann), being  awarded with the Faculty prize of the Department of Science. From 2010 to 2012, he worked as a Credit Risk Modeler at Credit Suisse.His current research activities cover high performance computing in finance & economics, the numerical solution of real business cycle models,  overlapping generation models and optimal taxation problems.

HOST: Prof. Olaf Schenk