pfR

Abstract

pfR is a new R package for particle filtering. In an attempt to promote a more interactive use through an R session, it provides an R interface to the C++ template library pf. All details of the particle filtering algorithms are abstracted away, so no technical expertise in particle filtering is required to specify and use any state-space model. Moreover, it is extensible so that small changes in the code can change the filtering algorithm or any of its tuning parameters. Two things make pfR unique. First, models are specified directly into C++ code, so there is no need to learn a domain-specific language. Second, pfR makes heavy use of class templates. This means almost every customizable parameter of a particle filtering algorithm can be specified at compile time, which can lead to dramatic speed gains. For example, the number of particles in a particle filter, the dimensions of vectors, the resampling strategy, and even the precision of floating points are all compile-time constants. A variety of examples using financial time series are given. Different stochastic volatility models are specified, and performance comparisons of filtering and (approximate) likelihood evaluations are provided. Last we demonstrate a few computationally expensive algorithms that make use of particle filters as a building blocks.