04.01.2013 Views

Springer - Read

Springer - Read

Springer - Read

SHOW MORE
SHOW LESS

Create successful ePaper yourself

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

8.8 Generalized State-Space Models 305<br />

Remark 1. The preceding analysis for the Poisson-distributed observation equation<br />

holds, almost verbatim, for the general family of exponential densities (8.8.31).<br />

(One only needs to take care in specifying the correct range for x and the allowable<br />

parameter space for α and λ in (8.8.37).) The relations (8.8.43)–(8.8.44), as well<br />

as the exponential smoothing formula (8.8.48), continue to hold even in the more<br />

general setting, provided that the parameters αt|t−1 and λt|t−1 satisfy the relations<br />

(8.8.41)–(8.8.42).<br />

Remark 2. Equations (8.8.41)–(8.8.42) are equivalent to the assumption that the<br />

prior density of Xt given y (t−1) is proportional to the δ-power of the posterior distribution<br />

of Xt−1 given Y (t−1) , or more succinctly that<br />

f(xt; αt|t−1,λt|t−1) f(xt; δαt−1|t−1,δλt−1|t−1)<br />

∝ f δ (xt; αt−1|t−1,λt−1|t−1).<br />

This power relationship is sometimes referred to as the power steady model (Grunwald,<br />

Raftery, and Guttorp, 1993, and Smith, 1979).<br />

Remark 3. The transformed state variables Wt e Xt have a gamma state density<br />

given by<br />

p wt+1|y (t) g(wt+1; αt+1|t,λt+1|t)<br />

(see Problem 8.26). The mean and variance of this conditional density are<br />

E Wt+1|y (t) αt+1|t and Var Wt+1|y (t) αt+1|t/λ 2<br />

t+1|t .<br />

Remark 4. If we regard the random walk plus noise model of Example 8.2.1 as<br />

the prototypical state-space model, then from the calculations in Example 8.8.1 with<br />

G F 1, we have<br />

and<br />

E Xt+1|Y (t) E Xt|Y (t)<br />

Var Xt+1|Y (t) Var Xt|Y (t) + Q>Var Xt|Y (t) .<br />

The first of these equations implies that the best estimate of the next state is the same<br />

as the best estimate of the current state, while the second implies that the variance<br />

increases. Under the conditions (8.8.41), and (8.8.42), the same is also true for the<br />

state variables in the above model (see Problem 8.26). This was, in part, the rationale<br />

behind these conditions given in Harvey and Fernandes (1989).<br />

Remark 5. While the calculations work out neatly for the power steady model,<br />

Grunwald, Hyndman, and Hamza (1994) have shown that such processes have degenerate<br />

sample paths for large t. In the Poisson example above, they argue that the<br />

observations Yt convergeto0ast →∞(see Figure 8.12). Although such models

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

Saved successfully!

Ooh no, something went wrong!