BOOTSTRAP FILTERING FOR CZECH MACRO-ECONOMIC MODEL ESTIMATION

TRNKA, P.; HAVLENA, V.

Abstract

Implementing direct Bayesian inference using Monte Carlo methods (Bootstrap filter) we identified Czech macro-economic model based on the work (Clarida et al. 1999). The main concern was to identify model parameters for the prediction of model behaviour, which is essential for taking proper economical decisions. Simultaneous estimation of model parameters led to non-linear model. Commonly used Extended Kalman filter failed in this case, therefore we used bootstrap filter, which can handle non-linear and/or non-gaussian systems. The posterior probability density function of states and parameters were obtained from the prior probabilities (represented as a large set of samples), which were updated from measured data according to Bayesian infer-ence. Given only limited data set (quarterly data from 1994) at disposal we incorporated smoothing (backward filtration) into bootstrap filter to maximize the use of information from the data. Identified model successfully revealed shocks in economy according to the history.

Coresponding author e-mail: trnkap[at]control[dot]felk[dot]cvut[dot]cz

Session: Modelling, Simulation, and Identification of Processes