Working Paper: NBER ID: w16708
Authors: Kenneth Judd; Lilia Maliar; Serguei Maliar
Abstract: In conventional stochastic simulation algorithms, Monte Carlo integration and curve fitting are merged together and implemented by means of regression. We perform a decomposition of the solution error and show that regression does a good job in curve fitting but a poor job in integration, which leads to low accuracy of solutions. We propose a generalized notion of stochastic simulation approach in which integration and curve fitting are separated. We specifically allow for the use of deterministic (quadrature and monomial) integration methods which are more accurate than the conventional Monte Carlo method. We achieve accuracy of solutions that is orders of magnitude higher than that of the conventional stochastic simulation algorithms.
Keywords: stochastic simulation; Monte Carlo; numerical methods
JEL Codes: C63
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| integration method (F15) | solution accuracy (C62) |
| polynomial degree (C69) | error rates (C52) |
| one-node Monte Carlo integration (C29) | integration errors (F15) |
| one-node Gauss-Hermite quadrature integration (C29) | solution errors (C62) |
| multi-node quadrature (C30) | solution errors (C62) |
| simulation length (C41) | integration errors (F15) |
| simulation length (C41) | curve fitting errors (C51) |