04.01.2013 Views

Springer - Read

Springer - Read

Springer - Read

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

308 Chapter 8 State-Space Models<br />

Table 8.3 Transition probabilities for the<br />

number of goals scored by England<br />

against Scotland.<br />

yt+1<br />

p(yt+1|yt ) 0 1 2 ≥ 3<br />

0 .214 .500 .214 .072<br />

yt 1 .409 .272 .136 .182<br />

2 .250 .375 .125 .250<br />

≥ 3 0 .857 .143 0<br />

for t 1, 2,..., where f is given by (8.8.36) and α1|0 0,λ1|0 0. The power<br />

steady conditions (8.8.41)–(8.8.42) are assumed to hold for αt|t−1 and λt|t−1. The only<br />

unknown parameter in the model is δ. The log-likelihood function for δ based on the<br />

conditional distribution of y1,...,y52 given y1 is given by (see (8.8.27))<br />

ℓ δ,y (n) n−1<br />

ln p yt+1|y (t) , (8.8.49)<br />

t1<br />

where p yt+1|y (t)) is the negative binomial density (see Problem 8.25(c))<br />

p yt+1|y (t) nb yt+1; αt+1|t,(1 + λt+1|t) −1 ,<br />

with αt+1|t and λt+1|t as defined in (8.8.44) and (8.8.43). (For the goal data, y1 0,<br />

which implies α2|1 0 and hence that p y2|y (1) is a degenerate density with unit<br />

mass at y2 0. Harvey and Fernandes avoid this complication by conditioning the<br />

likelihood on y (τ) , where τ is the time of the first nonzero data value.)<br />

Maximizing this likelihood with respect to δ, we obtain ˆδ .844. (Starting the<br />

equations (8.8.43)–(8.8.44) with α1|0 0 and λ1|0 δ/(1 − δ), we obtain ˆδ .732.)<br />

With .844 as our estimate of δ, the prediction density of the next observation Y53 given<br />

y (52) is nb(y53; α53|52,(1+λ53|52) −1 . The first five values of this distribution are given in<br />

Table 8.4. Under this model, the probability that England will be held scoreless in the<br />

next match is .471. The one-step predictors, ˆY1 0, ˆY2,...,Y52 are graphed in Figure<br />

8.10. (This graph can be obtained by using the ITSM option Smooth>Exponential<br />

with α 0.154.)<br />

Figures 8.11 and 8.12 contain two realizations from the fitted model for the goal<br />

data. The general appearance of the first realization is somewhat compatible with the<br />

goal data, while the second realization illustrates the convergence of the sample path<br />

to 0 in accordance with the result of Grunwald et al. (1994).

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

Saved successfully!

Ooh no, something went wrong!