02.12.2012 Views

Applications of state space models in finance

Applications of state space models in finance

Applications of state space models in finance

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

44 4 Markov regime switch<strong>in</strong>g<br />

who employ a hidden Markov model to derive stylized facts <strong>of</strong> daily S&P 500 return<br />

series.<br />

The standard regime switch<strong>in</strong>g model can be extended to provide an even higher<br />

degree <strong>of</strong> flexibility. Important specifications <strong>in</strong>clude Markov switch<strong>in</strong>g <strong>models</strong> with<br />

time-vary<strong>in</strong>g switch<strong>in</strong>g probabilities (Diebold et al. 1994; Filardo 1994) and hidden<br />

semi-Markov <strong>models</strong>, which allow for nonparametric <strong>state</strong> occupancy distributions, as<br />

proposed by Ferguson (1980). As these extensions are beyond the scope <strong>of</strong> this thesis,<br />

they will not f<strong>in</strong>d any consideration hereafter; for further read<strong>in</strong>g, see, for example, Bulla<br />

(2006, ¢ 5) and the references given there<strong>in</strong>.<br />

HMMs are employed <strong>in</strong> a wide spectrum <strong>of</strong> applications due to their overall versatility<br />

and mathematical tractability. Their attractiveness is grounded on the fact that<br />

the likelihood can be evaluated <strong>in</strong> a straightforward fashion, either by numerical maximization<br />

or by employ<strong>in</strong>g the EM algorithm. Moment properties can be easily derived,<br />

miss<strong>in</strong>g observations can be easily dealt with, and HMMs are <strong>of</strong>ten <strong>in</strong>terpretable <strong>in</strong> a<br />

natural way. Besides, they are moderately parsimonious with a simple two-<strong>state</strong> model<br />

provid<strong>in</strong>g a good fit <strong>in</strong> many cases. Comprehensive treatments <strong>of</strong> HMMs are provided<br />

by Elliott et al. (1995), MacDonald and Zucch<strong>in</strong>i (1997), Bhar and Hamori (2004) and<br />

Hamilton (1993), with the latter two focus<strong>in</strong>g on applications <strong>in</strong> economics and f<strong>in</strong>ance.<br />

The organization <strong>of</strong> this chapter is as follows. Section 4.1 briefly reviews <strong>in</strong>dependent<br />

mixture <strong>models</strong> and basic properties <strong>of</strong> Markov cha<strong>in</strong>s. Section 4.2 outl<strong>in</strong>es the basic<br />

hidden Markov model. Section 4.3 briefly discusses parameter estimation based on the<br />

ML method. Section 4.4 looks at forecast<strong>in</strong>g and decod<strong>in</strong>g procedures, before Section 4.5<br />

provides an overview <strong>of</strong> model selection and validation. To assure notational conformability,<br />

the outl<strong>in</strong>e <strong>in</strong> this chapter is, unless <strong>state</strong>d otherwise, based on Bulla (2006, ¢ 2–3)<br />

who also provided the code for conduct<strong>in</strong>g the computations related to HMMs <strong>in</strong> this<br />

thesis us<strong>in</strong>g the statistical s<strong>of</strong>tware package R 2.1.1 (R Development Core Team 2005).<br />

4.1 Basic concepts<br />

Before <strong>in</strong>troduc<strong>in</strong>g the theory <strong>of</strong> HMMs, it is <strong>in</strong>structive to look at two fundamental<br />

concepts as a basis to understand the basic structure <strong>of</strong> a hidden Markov model. Subsection<br />

4.1.1 reviews the concept <strong>of</strong> mixture distributions. Subsection 4.1.2 provides a<br />

brief <strong>in</strong>troduction to the basic properties <strong>of</strong> Markov cha<strong>in</strong>s, which are needed to construct<br />

HMMs.<br />

4.1.1 Independent mixture distributions<br />

Consider the weekly log-return series <strong>of</strong> the Technology sector. As discussed <strong>in</strong> ¢ 2.2.2,<br />

the series <strong>of</strong> all sectors show signs <strong>of</strong> positive autocorrelation <strong>in</strong> the squared returns.<br />

This characteristic is commonly referred to as volatility cluster<strong>in</strong>g, a common feature<br />

<strong>of</strong> f<strong>in</strong>ancial returns data that usually <strong>in</strong>duces excess kurtosis. Panel (a) <strong>of</strong> Figure 4.1<br />

shows the weekly percentage log-returns on the DJ Technology sector. It can be seen<br />

that the sector’s volatility is not constant over time: the level <strong>of</strong> volatility is clearly<br />

lower at the beg<strong>in</strong>n<strong>in</strong>g and higher <strong>in</strong> the second half <strong>of</strong> the sample. As a consequence, a<br />

normal distribution is not capable <strong>of</strong> describ<strong>in</strong>g the return series adequately. Panel (b)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!