Multistep ahead prediction of future data using Monte Carlo sampling
In an AR(p) model of data evolution in a discrete time with a Gaussian noise we are trying to predict the distribution of data up to the horizon t + h,knowing the data at time t. Because of the use of Bayesian paradigm, we assume one source of uncertainty to be model parameters, therefore modelled as random variables. Since the analytical solution of such a prediction is dificult due to the high dimensionality of the problem (in our case the dimension may be as high as in nite), we are forced to search for approximative solutions. We propose one such solution using Monte Carlo sampling from the prior distribution and later reconstruction of the final distribution. This seminar is going to inform about the progress of the project to present date.