J.Q. Smith and Antonio Santos
g (αt∣αt- 1 ,j) = f (αt∣αt- 1 ,j) is used to define a tentative draw from f (αt∖Dt). The
“plain” SIR algorithm can be expressed using the following steps:
1. Generate a new set of particles using the transition equation, αt,j = gt(αt∣αt-1,j),
j = 1, . . . , m. A new set of weights is calculated for each of the m particles
using the formula
(5)
wt,j = f(yt∣αt,j)
wt,j
∏t,j = ÿ™-----, j = 1 ,...,m (6)
i=1 wt,i
2. Resample from {αt,1 , . . . , αt,m} using the weights {πt,1, . . . , πt,m}, and thus
obtain a new set of particles with equal weights. These are then used in the
next iteration.
Although this method was regarded as a considerable breakthrough, it is now
widely recognized that this algorithm suffers from several weaknesses:
1. sample impoverishment (so the quality of the approximation thus deteriorates
as time passes);
2. a lack of robustness regarding outliers, and
3. typically poor approximation of the tails of the posterior distribution.
Improvements to the basic SIR algorithm focusing on robust filters to outliers
were recently proposed by Pitt and Shephard (1999, 2001). This algorithm - called
the auxiliary particle filter (APF) - is widely recognized as an important improve-
ment on the basic algorithm when implemented in time series such as those in
finance, where the weaknesses referred to above become critical.
From a sequential perspective the main objective is to update the particles at t -
1, and the respective weights, {αt-1,1, . . . , αt-1,m} and {πt-1,1, . . . , πt-1,m}. Using the
G.E.M.F - F.E.U.C.