How to solve dynamic stochastic models computing expectations just once

Abstract

We introduce a computational technique—precomputation of integrals—that makes it possible to construct conditional expectation functions in dynamic stochastic models in the initial stage of a solution procedure. This technique is very general: it works for a broad class of approximating functions, including piecewise polynomials; it can be applied to both Bellman and Euler equations; and it is compatible with both continuous‐state and discrete‐state shocks. In the case of normally distributed shocks, the integrals can be constructed in a closed form. After the integrals are precomputed, we can solve stochastic models as if they were deterministic. We illustrate this technique using one‐ and multi‐agent growth models with continuous‐state shocks (and up to 60 state variables), as well as Aiyagari's (1994) model with discrete‐state shocks. Precomputation of integrals saves programming efforts, reduces computational burden, and increases the accuracy of solutions. It is of special value in computationally intense applications. MATLAB codes are provided.

Exploiting Symmetry in High-Dimensional Dynamic Programming

“…In fact, we only need to change one parameter value in the deep learning code that solves the model: everything shocks are Gaussian) would miss the typical set of the distribution, which in high dimensions does not include the mode of the distribution (i.e., zero). Our argument of drawing once from the distribution of idiosyncratic shocks, thanks to concentration of measure, is different from Judd et al (2017), who propose precomputing the integrals required in Bellman or Euler equations.…”

Using a Hyperbolic Cross to Solve Non-Linear Macroeconomic Models

“…Although the hyperbolic cross approximation is also very accurate, its average absolute Euler error is −8.9, which is about the same as the approximation using just 65 points and slightly less accurate than the five-layer Smolyak approximation. 15…”

Programming FPGAs for Economics: An Introduction to Electrical Engineering Economics

“…After the pioneering contribution of Krusell and Smith (1998), heterogeneous agent models have been used extensively to study business cycle fluctuations, monetary and fiscal policy, life-cycle decisions, industry dynamics, and international trade, and to answer many other questions. Also, there has been tremendous interest in the development of solution methods well-suited for these models, like Rust (1997), Algan et al (2008), Reiter (2009), Den Haan and Rendahl (2010), Maliar et al (2010), Reiter (2010), Young (2010), Algan et al (2014), Sager (2014) Pröhl (2015), Nuño and Thomas (2016), Achdou et al (2021), Bhandari et al (2017), Brumm and Scheidegger (2017), Judd et al (2017), Bayer and Luetticke (2018), Childers (2018), Mertens and Judd (2018), Winberry (2018), Fernández-Villaverde et al (2019), Auclert et al (2020), Bilal (2021), and Kahou et al (2021), among score of other papers.…”