Stochastic simulation of KVARTS91

Abstract

This report documents stochastic simulations of the quarterly macroeconometric model KVARTS91, implemented in the TROLL software system. By means of stochastic simulation we want to see whether the standard deterministic model solution approximates the expectation of a stochastic model solution, and to quantify (some of) the econometric uncertainty in that stochastic solution. A built-in Stochastic Simulator is used to simulate the model with stochastic residuals and stochastic parameter estimates, assuming normal distributions for the stochastic input. The model is simulated ex ante through 1993 and 1994. The results show less than 1 percent deterministic bias in the endogenous variables. On the other hand do the widths of the simulated (95 percent) prediction intervals vary a lot. But, for most variables the interval widths stay below 10 percent of the level of the variable. The uncertainty in the model solutions imply that an analysis based on stochastic rather than deterministic model simulations may lead to more subtle conclusions. To make stochastic simulation a feasible and realistic alternative to standard deterministic simulation, commands that prepare and govern the stochastic simulation by TROLL's Stochastic Simulator have been collected into macros. Some effort has also been put into the writing of small programs that make the documentation of a stochastic simulation experiment a swift semi-automatic procedure.