J.Q. Smith and Antonio Santos
for example, by using Markov chain Monte Carlo (MCMC) techniques, the main
aim of this paper is to present certain modifications of particle filter methods which
have recently been proposed to predict the process. The predictions made by this
model, in contrast with ARCH family models, are expressed through the posterior
density of the states f ( αt∣Dt ) and the predictive density of returns f ( yt+1 ∣Dt ) rather
than through point predictions. Henceforth we assume a closed forecasting system
and we let Dt = {y0, y1, . . . , yt} represent the available information at time t. Our
modifications are straightforward but nevertheless appear to improve the predictive
performance dramatically when these models are applied to stock return series. We
first review current particle filter methods.
2 Particle Filter methods
The Bayes’ rule allows us to assert that the posterior density f (αt∣Dt) of states is
related to the density f (αt∣Dt- 1) prior to yt and the density f (yt∣αt) of yt given αt
by
f (αtD) α f (yt∖at) f (αtlDt- 1) (3)
and the predictive density of yt+1 given Dt is
f (yt+1|Dt) =
f (yt+1 |at+1) f (αt+1 lDt) dαt+1
(4)
Instead of numerically estimating these integrals, the particle filter approximates
these densities using a simulated sample.
Particle filters approximate the posterior density of interest, f (αt∣Dt), through
a set of m “particles” {αt,1, . . . , αt,m} and their respective weights {πt,1, . . . , πt,m}
where πt,j ≥ 0 and jm=1 πt,j = 1. To implement these filters, we must first be able
to sample from the nonstandard density f (αt∣Dt). It is possible to develop simula-
tion procedures to approximate the distribution of interest and to calculate certain
G.E.M.F - F.E.U.C.