197 Model Validation and Diagnostics
residual ACF plots in the chapter) display dashed lines representing the usual approxi-
mate two standard error bounds for departure of the relevant ordinate from zero. The left
panel refers to the cuts data set and the right panel to the iceberg order data. Both indicate
that this simple model does not adequately capture the dependencies in the data. From
the left panel, the dynamic misspecication previously observed in Figure 9.3 is not so evi-
dent. However, the residual autocorrelation at lag 12 appears quite large and may suggest
a neglected seasonal component. The right panel, associated with the iceberg order data,
tells a similar story to that already gleaned from Figure 9.3.
The sample means of the Pearson residuals for both data sets are close to zero. However,
their variances are 1.607 and 1.289 for the cuts and iceberg order data, respectively. Both
numbers are considerably larger than unity, suggesting potential misspecication of the
Poisson innovation distribution specied in the PINAR(1) tted model.
9.3.2 Component Residuals
At the outset, note that the two parts of the right-hand side of (9.1) and the special case (9.2)
are typically unobserved. The rst part can be thought of as a specication for the (random)
number of survivors from stochastic operations performed at, or prior to, time t,or its
complement, the number of departures. The second part reects the number of new arrivals
to the system at time t. It is transparent to derive the concepts in the following for the
model variant provided in (9.2). For this case, though each of the α
k
X
t−k
is unobservable,
following Freeland and McCabe (2004), we dene a set (t = 1, ..., T) of departure residuals
for each operator in (9.2) (k ∈{1, ..., p})by
r
k,t
= E
t
[α
k
X
t−k
] − E
t−1
[α
k
X
t−k
] , (9.5)
where E
t
is the expectation conditional on all information up to and including time t.
Generally, E
t
[α
k
X
t−k
] = E
t−1
[α
k
X
t−k
] as the conditioning sets are different. Similarly,
dene the set of arrivals residuals (t = 1, ..., T)as
r
p +1,t
= E
t
[ε
t
] − E
t −1
[ε
t
] . (9.6)
By considering the sum of the set of p + 1 component residuals thereby dened as
p+1 p p
r
k,t
= E
t
[α
k
X
t−k
] + E
t
[ε
t
] − E
t−1
[α
k
X
t−k
] + E
t−1
[ε
t
]
k=1 k=1 k=1
= E
t
[X
t
] − E
t−1
[X
t
] = X
t
− E
t−1
[X
t
] = r
t
,
it is seen that the component residuals add up to the usual raw residuals for model (9.2).
One advantage of sets of residuals being associated with each unobserved part of the
model is to offer the potential that they be used to identify the source of problems asso-
ciated with a tted model in the following way. Initially, any of this set of p + 1 residuals
may be used to check specication, either informally through the use of time series plots
(or other graphical devices) and/or more formally through the construction of statistical
specication tests. If some component residual indicates that the corresponding compo-
nent of the model is not well specied, it may be possible to suggest modications for
improvement. For example, a cyclical pattern in a residual plot may indicate the presence