doi:10.1111/j.1467-9892.2008.00604.

BOOTSTRAP PREDICTION INTERVALS IN STATE–SPACE MODELS

By Alejandro Rodriguez and Esther Ruiz

Universidad Carlos III de Madrid
First Version received March 2008

Abstract. Prediction intervals in state–space models can be obtained by assuming

Gaussian innovations and using the prediction equations of the Kalman filter, with the
true parameters substituted by consistent estimates. This approach has two limitations.
First, it does not incorporate the uncertainty caused by parameter estimation. Second,
the Gaussianity of future innovations assumption may be inaccurate. To overcome these
drawbacks, Wall and Stoffer [Journal of Time Series Analysis (2002) Vol. 23, pp.
733–751] propose a bootstrap procedure for evaluating conditional forecast errors that
requires the backward representation of the model. Obtaining this representation
increases the complexity of the procedure and limits its implementation to models for
which it exists. In this article, we propose a bootstrap procedure for constructing
prediction intervals directly for the observations, which does not need the backward
representation of the model. Consequently, its application is much simpler, without
losing the good behaviour of bootstrap prediction intervals. We study its finite-sample
properties and compare them with those of the standard and the Wall and Stoffer
procedures for the local level model. Finally, we illustrate the results by
implementing the new procedure to obtain prediction intervals for future values of a
real time series.

Keywords. Backward representation; Kalman filter; local level model; unobserved

components.

1. INTRODUCTION

Time series models with unobserved components may be useful in many

contexts (see, e.g. Durbin and Koopman, 2001). After casting the model in
state–space (SS) form, the Kalman filter allows one to obtain predictions of
future values of the series together with their corresponding mean square errors
(MSE). Then, prediction intervals are obtained by assuming Gaussian errors. In
practice, the parameters of the model are unknown and are substituted by
consistent estimates. Therefore, prediction intervals may be inaccurate caused
by the estimation uncertainty and/or when the normal distribution is not
adequate. In the context of autoregressive integrated moving-average (ARIMA)
models, several authors have proposed the use of bootstrap procedures to
construct prediction intervals that overcome these limitations (see Thombs and
Schucany, 1990, for the original bootstrap procedure for AR(p) models and
Pascual et al., 2004, for a simpler procedure not based on the backward
representation).

0143-9782/09/02 167–178 JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2

Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd., 9600 Garsington Road, Oxford OX4 2DQ, UK
and 350 Main Street, Malden, MA 02148, USA.
168 A. RODRIGUEZ AND E. RUIZ

Unlike ARIMA models, models with unobserved components may have several
disturbances. Therefore, the bootstrap procedures proposed by Thombs and
Schucany (1990) and Pascual et al. (2004) cannot be directly applied to them. To
overcome this problem, Wall and Stoffer (2002), from now on WS, propose that
bootstrap prediction intervals for future observations can be obtained by using
the innovation form (IF) of the SS models, which is defined in terms of a unique
disturbance. They show that their procedure works well in the context of
Gaussian SS models. Moreover, Pfeffermann and Tiller (2005) show that the
bootstrap estimator of the underlying unobserved components based on the IF is
asymptotically consistent. Following Pascual et al. (2004), in this article, we
propose a bootstrap procedure to obtain prediction intervals of future
observations in SS models that simplifies the WS procedure both from the
computational point of view and because it does not require the backward
representation. Similar to Wall and Stoffer, our proposed bootstrap procedure is
based on the IF. We show that the new procedure has the advantage of being
much simpler without losing the good behaviour of bootstrap prediction intervals.
The rest of the article is organized as follows. In section 2, we describe the
model, the filters and propose a new bootstrap procedure. We analyse its finite-
sample properties, comparing them with those of the standard and the Wall and
Stoffer prediction intervals. Section 3 presents an application of the new bootstrap
procedure to a real time series. Section 4 concludes the paper with our conclusions
and some suggestions for future research.

2. BOOTSTRAP PREDICTION INTERVALS IN SS MODELS

2.1. SS models and the Kalman filter

Consider the following time-invariant SS model,
yt ¼ Zat þ d þ et ; ð1aÞ

at ¼ Tat1 þ c þ Rgt ; t ¼ 1; . . . ; T ð1bÞ

where yt is a univariate time series, at is the m  1 vector of unobservable state
variables, et is a serially uncorrelated disturbance with zero mean and variance H
and gt is a g  1 vector of serially uncorrelated disturbances with zero mean and
covariance matrix Q. The disturbances et and gt are uncorrelated with each other
in all time periods Finally, the initial state vector, a0 , has mean a0 and covariance
matrix P0.
The Kalman filter allows one to estimate the state vector, atþ1 , and its MSE
based on the information available at time t. These estimates are given by
atþ1jt ¼ Tatjt1 þ c þ Kt Ft1 vt ; ð2aÞ

Ptþ1jt ¼ TPtjt1 T0  Kt Ft1 K0t þ RQR0 ; ð2bÞ

Ó 2009 The Authors

Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
 BOOTSTRAP PREDICTION INTERVALS IN SS MODELS 169
where vt ¼ yt  d  Zatjt1 is the innovation, Ft ¼ ZPtjt1Z0 þH is its variance
and Kt ¼ TPtjt1Z0 .
Although the SS model in eqn (1) has several disturbances, the IF has a unique
disturbance. The IF is given by eqn (2a) together with
yt ¼ Zatjt1 þ d þ vt : ð3Þ
Assuming that future prediction errors are Gaussian, the k-step-ahead prediction
intervals for yTþk are given by
 pffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffi 
~yT þkjT  z1a=2 FT þkjT ; ~yT þkjT þ z1a=2 FT þkjT ; ð4Þ
where z1a/2 is the (1  (a/2))-percentile of the standard normal distribution,
and ~yT þkjT and FTþkjT are the k-step-ahead prediction of yTþk and its MSE,
given by
X
k1
yT þkjT ¼ ZTk aT þ Z
e Tj c þ d; k ¼ 1; 2; . . . ; ð5aÞ
j¼0
k1 h
X  0 i
FT þkjT ¼ ZðTk ÞPT ðTk Þ0 Z0 þ Z ðT j ÞRQR0 Tj Z0 þ H; k ¼ 1; 2; . . . : ð5bÞ
j¼0

In practice, the unknown parameters involved in eqn (5) are substituted by

consistent estimates, usually quasi-maximum likelihood (QML) estimates because
of their well-known asymptotic properties (see, e.g. Durbin and Koopman, 2001).
The intervals in eqn (4), where ~yT þkjT and FTþkjT are obtained by substituting the
unknown parameters by their QML estimates, are called standard (ST). Note that
the ST prediction intervals underestimate, in general, the variability of the
forecast error because they do not take into account the uncertainty caused by
parameter estimation. Moreover, these intervals could have inaccurate coverages
when the prediction errors are not Gaussian.
To overcome these problems, Wall and Stoffer (2002) propose a bootstrap
procedure based on the IF. Following Thombs and Schucany (1990), they
propose the use of the backward SS representation to generate bootstrap
replicates of the series with fixed last observations. These replicates are used to
incorporate the uncertainty caused by parameter estimation in the density of the
prediction errors. However, unlike Thombs and Schucany (1990), they propose to
obtain bootstrap densities of the prediction errors instead of bootstrapping
directly future observations of the series of interest. The bootstrap densities of the
prediction errors are obtained in two steps. First, they construct bootstrap
prediction errors that incorporate the uncertainty due to the fact that, when
predicting, future innovations are equal to zero while in fact they are not. These
bootstrap replicates do not incorporate the uncertainty caused by parameter
estimation. Therefore, they obtain another set of bootstrap prediction errors that
incorporate the variability attributable to parameter estimation through the use of
bootstrap-estimated parameters instead of the original estimates. These bootstrap
replicates assume that future innovations are zero. Finally, combining both sets of
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
170 A. RODRIGUEZ AND E. RUIZ

bootstrap prediction errors, they estimate the density of the conditional forecast
errors that can be used for constructing the corresponding bootstrap prediction
intervals. They are given by
h i
~yT þkjT þ Qa=2;d  ; ~yT þkjT þ Q1a=2;d  ð6Þ
k k

where Qa=2;d is the a/2-percentile of the empirical conditional bootstrap

k
distribution of the k-step-ahead prediction errors of yTþk.

2.2. A new procedure

We propose to construct bootstrap prediction intervals directly approximating the
conditional distribution of yTþk by the distribution of bootstrap replicates that
incorporate simultaneously the variability caused by parameter estimation and
the uncertainty because of unknown future innovations without using the
backward filter. The new procedure consists of the following steps.
Step 1. Estimate the parameters of model (1) by QML, h, ^ and obtain the
s
standardized innovations f^vt ; 1  t  T g.
Step 2. Obtain a sequence of bootstrap standardized innovations
f^vs
t ; 1  t  T þ Kg via random draws with replacement from the stan-
dardized innovations, ^vst .
Step 3. Compute a bootstrap replicate f^yt ; 1  t  T g by means of the IF in
eqns (3) and (2a) using ^vs ^
t and the estimated parameters, h. Estimate the cor-
responding bootstrap parameters, h . ^ 

Step 4. Run the Kalman filter with ^ h and the original observations and obtain
a bootstrap replicate of the state vector at time T which incorporates the
uncertainty caused by parameter estimation, ^aT jT 1 .
Step 5. Obtain conditional bootstrap k-step-ahead predictions, f^yT þkjT ;
1  k  Kg, by the following expressions
X
k1 X
k1
^ k ^
aT þkjT ¼ T
^ aT jT 1 þ ^ k1j c þ
T ^ k1j K
T ^  F^ 1^v ;
T þj T þj T þj
j¼0 j¼0

X
k1
^yT þkjT ¼ Z
^ T
^ k a
^T jT 1 þ Z
^ ^ k1j c þ d
T
j¼0

X
k 1
^
þZ ^ k1j K
T ^  F^ 1^v þ ^v ; k ¼ 1; . . . ;
T þj T þj T þj T þk
j¼0

^ ^
where ^vT ¼ yT  Z aT jT 1 and the hat in top of the matrices means that they are
obtained by substituting the parameters by their corresponding bootstrap
estimates.

Steps 2 to 5 are repeated B times, obtaining B bootstrap replicates of ^yT þkjT .

Note that the empirical distribution of ^yT þkjT incorporates both the variability
caused by unknown future innovations and the variability caused by parameter
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
 BOOTSTRAP PREDICTION INTERVALS IN SS MODELS 171
estimation in just one step. The procedure above, denoted as state-space
bootstrap (SSB), has three advantages over the Wall and Stoffer procedure.
First, it does not require the use of the backward representation. Second, it is
simpler as a unique set of bootstrap replicates of future observations is required
instead of two as in the WS procedure. Third, unlike the WS procedure, we do not
fix ^aT jT 1 ¼ ^aT jT 1 in all bootstrap replicates of future observations. This value
depends on the estimated parameters and, consequently, it should be allowed to
vary among bootstrap replicates in order to incorporate the uncertainty caused by
parameter estimation. However, note that all realizations of ^yT þkjT pass through
yT, making the predictions conditional on the available sample.
Finally, the SSB bootstrap prediction intervals are constructed directly by the
percentile method1 as follows
h i
Qa=2;^y  ; Q1a=2;^y  ð7Þ
T þkjT T þkjT

where Qa=2;^y  is the a/2-percentile of the empirical distribution ^yT þkjT .

T þkjT

2.3. Finite-sample properties

To analyse the finite-sample properties of the SSB prediction intervals, we
consider the local level model given by
yt ¼ lt þ et ; et  i.i.d.ð0; r2e Þ; ð8aÞ
lt ¼ lt1 þ gt ; gt  i.i.d.ð0; qr2e Þ; ð8bÞ
where lt is the level of the series that evolves over time following a random walk
and q is known as the signal-to-noise ratio.
Simulation results are based on R ¼ 1000 replicates of series of sizes T ¼
50, 100 and 500. The parameters of the model have been chosen to cover a wide
range of situations from cases in which the noise is large relative to the signal, i.e.
q is small, to cases in which q is large. In particular, we consider q ¼ f0.1, 1, 2g.
With respect to the disturbances, we consider two distributions, Gaussian and a
centred and re-scaled chi-square with 1 degree of freedom,2 v2ð1Þ . For each
ðrÞ ðrÞ
simulated series, fy1 ; . . . ; yT g, r ¼ 1, 2, . . . , R, we first generate B ¼ 1000
ðrÞ
observations of yT þk for prediction horizons k ¼ 1, 5 and 15, and then obtain
95% prediction intervals computed using the ST intervals in eqn (4), the Wall and
Stoffer intervals in eqn (6) and the SSB intervals in eqn (7). Finally, we compute
the coverage of each of these intervals as well as the length and the percentage of
observations left out on the right and on the left of the limits of the prediction
intervals.3
Table I reports the Monte Carlo averages of these quantities when both
disturbances are Gaussian. Although, in this case, the three procedures have very
similar properties, the average coverage of the ST intervals is smaller than the
coverage of the SSB intervals when the sample size is small (T ¼ 50) and the
prediction horizon is 5 and 15. This result can be explained by the fact that ST
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
172 A. RODRIGUEZ AND E. RUIZ

TABLE I
Monte Carlo Average Coverages, Length and Percentage of Observations Left Out on the
Right and on the Left of the Prediction Intervals for yTþk Constructed Using ST, Wall
and Stoffer (WS) and SSB when et is N(0, 1), gt is N(0, q) and the Nominal Coverage is 95%

Mean coverage Mean coverage in tails Mean length

ST WS SSB
k ST WS SSB Below/Above Below/Above Below/Above ST WS SSB
T ¼ 50
q ¼ 0.1 1 0.927 0.935 0.936 0.036/0.037 0.030/0.035 0.031/0.033 4.530 4.597 4.774
5 0.927 0.940 0.943 0.036/0.037 0.029/0.031 0.028/0.029 5.182 5.285 5.539
15 0.915 0.928 0.940 0.042/0.042 0.035/0.037 0.030/0.031 6.460 6.633 7.052
q¼1 1 0.936 0.923 0.928 0.029/0.035 0.036/0.041 0.036/0.035 6.157 6.250 6.280
5 0.927 0.921 0.938 0.035/0.039 0.037/0.042 0.032/0.031 9.722 9.718 10.274
15 0.914 0.909 0.934 0.041/0.045 0.043/0.047 0.033/0.033 15.258 15.194 16.469
q¼2 1 0.938 0.930 0.930 0.032/0.029 0.036/0.034 0.036/0.034 7.424 7.56 7.433
5 0.926 0.924 0.931 0.037/0.036 0.038/0.038 0.034/0.034 12.849 12.880 13.088
15 0.918 0.915 0.930 0.041/0.041 0.042/0.042 0.035/0.035 20.889 20.830 21.632
T ¼ 100
q ¼ 0.1 1 0.945 0.941 0.943 0.025/0.030 0.031/0.028 0.026/0.031 4.569 4.576 4.618
5 0.945 0.942 0.948 0.025/0.030 0.030/0.028 0.024/0.029 5.206 5.238 5.334
15 0.938 0.938 0.945 0.029/0.033 0.032/0.030 0.026/0.030 6.498 6.575 6.743
q¼1 1 0.944 0.940 0.939 0.028/0.028 0.030/0.029 0.030/0.031 6.271 6.314 6.278
5 0.939 0.937 0.942 0.031/0.030 0.032/0.031 0.029/0.029 9.874 9.873 10.120
15 0.934 0.932 0.940 0.033/0.033 0.034/0.034 0.030/0.030 15.547 15.521 16.165
q¼2 1 0.945 0.937 0.939 0.028/0.027 0.032/0.030 0.031/0.030 7.476 7.537 7.460
5 0.939 0.938 0.939 0.030/0.030 0.031/0.031 0.031/0.031 13.137 13.155 13.210
15 0.935 0.935 0.937 0.032/0.032 0.032/0.033 0.031/0.031 21.509 21.539 21.758
T ¼ 500
q ¼ 0.1 1 0.946 0.948 0.945 0.027/0.027 0.025/0.028 0.028/0.027 4.592 4.577 4.582
5 0.946 0.947 0.946 0.026/0.028 0.025/0.028 0.027/0.027 5.217 5.206 5.223
15 0.946 0.945 0.945 0.026/0.029 0.026/0.029 0.027/0.028 6.515 6.477 6.511
q¼1 1 0.948 0.948 0.947 0.029/0.023 0.027/0.025 0.027/0.025 6.339 6.335 6.314
5 0.948 0.947 0.947 0.027/0.025 0.027/0.026 0.028/0.025 10.075 10.049 10.073
15 0.947 0.946 0.947 0.027/0.026 0.027/0.027 0.027/0.026 15.956 15.919 15.944
q¼2 1 0.947 0.945 0.947 0.027/0.026 0.029/0.026 0.027/0.026 7.563 7.546 7.540
5 0.948 0.948 0.948 0.027/0.027 0.027/0.025 0.027/0.027 13.418 13.446 13.387
15 0.947 0.948 0.947 0.027/0.026 0.026/0.026 0.026/0.027 22.066 22.112 22.051

intervals do not incorporate the uncertainty caused by parameter estimation.

Comparing the two bootstrap intervals, we can observe that both are similar, with
the average coverages of SSB being closer to the nominal in small samples and/or
when the prediction horizon increases. This result is illustrated in Figure 1 that
plots kernel estimates of the ST, Wall and Stoffer and SSB densities for 15-step-
ahead predictions for one particular series generated by each of the three models
considered together with the empirical density. Note that when the signal-to-noise
ratio is small, i.e. q ¼ 0.1 and T ¼ 50, the SSB density seems to be more similar to
the empirical density than the other densities.
Table II, that reports the results when et is v2ð1Þ and gt is Gaussian, shows that,
although the mean coverage of the ST intervals is close to the nominal, they are
not capable of dealing with the asymmetry in the distribution of et. The average
coverage in the left tail is smaller than that in the right tail. The difference between
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
 BOOTSTRAP PREDICTION INTERVALS IN SS MODELS 173
Empirical SSB ST WS
q = 0.1 q=1 q=2
0.08
0.1
0.2 0.08 0.06
T = 50

0.06 0.04
0.1 0.04
0.02
0.02
0 0
−5 0 5 −10 0 10 −20 −10 0 10 20

0.08
0.1
0.2 0.06
T = 100

0.04
0.05
0.1
0.02

0 0 0
−5 0 5 −10 0 10 −20 −10 0 10 20

0.08
0.1
0.2 0.06
T = 500

0.04
0.05
0.1
0.02

0 0 0
−5 0 5 −10 0 10 −20 −10 0 10 20

Figure 1. Kernel estimates densities of yTþk for k ¼ 15. Normal case.

the coverage in both tails is larger in the model with q ¼ 0.1, where the signal is
relatively small with respect to the non-Gaussian noise. Note that the inability of
the ST intervals to deal with the asymmetry in the distribution of et is larger the
larger the sample size. On the other hand, the coverages of the Wall and Stoffer
and SSB intervals are rather similar, with SSB being again slightly closer to the
nominal for almost all models and sample sizes considered. Both bootstrap
intervals are capable of coping with the asymmetry of the distribution of et.
Consequently, according to the results reported in Table II, using the much
simpler SSB method does not imply a worse performance of the prediction
intervals. Figure 2 illustrates these results by plotting the kernel density of the
simulated yTþ1 together with the ST, Wall and Stoffer and SSB densities
obtained with a particular series generated by each of the models and sample
sizes considered. This figure also illustrates the lack of fit of the ST density when
q ¼ 0.1 and 1. On the other hand, the shapes of the Wall and Stoffer and SSB
densities are similar, with SSB being always closer to the empirical.

3. EMPIRICAL APPLICATION

We illustrate the performance of the proposed procedure to construct bootstrap

prediction intervals by implementing it on the standardized quarterly mortgages
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
174 A. RODRIGUEZ AND E. RUIZ

TABLE II
Monte Carlo Average Coverages, Length and Percentage of Observations Left Out on the
Right and on the Left of the Prediction Intervals for yTþk Constructed Using ST, Wall
and Stoffer (WS) and SSB when et is v2ð1Þ , gt is N(0, q) and the Nominal Coverage is 95%

Mean coverage Mean coverage in tails Mean length

ST WS SSB
Case k ST WS SSB Below/Above Below/Above Below/Above ST WS SSB
T ¼ 50
q ¼ 0.1 1 0.941 0.940 0.942 0.010/0.049 0.030/0.030 0.027/0.031 4.513 4.909 4.734
5 0.943 0.934 0.946 0.013/0.044 0.039/0.027 0.027/0.026 5.221 5.507 5.596
15 0.930 0.919 0.950 0.025/0.045 0.053/0.029 0.027/0.023 6.572 6.665 7.329
q¼1 1 0.935 0.932 0.934 0.026/0.039 0.034/0.034 0.034/0.032 6.200 6.514 6.459
5 0.926 0.926 0.930 0.034/0.040 0.040/0.034 0.038/0.032 9.682 9.803 9.919
15 0.913 0.914 0.923 0.042/0.045 0.041/0.036 0.045/0.041 15.126 15.176 15.597
q¼2 1 0.937 0.933 0.932 0.028/0.035 0.034/0.034 0.035/0.033 7.378 7.714 7.575
5 0.927 0.924 0.927 0.035/0.038 0.040/0.036 0.038/0.035 12.805 12.897 12.957
15 0.919 0.917 0.923 0.040/0.041 0.043/0.040 0.040/0.037 20.839 20.880 21.236
T ¼ 100
q ¼ 0.1 1 0.947 0.939 0.943 0.006/0.048 0.033/0.028 0.027/0.029 4.552 4.773 4.710
5 0.946 0.937 0.942 0.010/0.043 0.037/0.026 0.031/0.027 5.196 5.356 5.414
15 0.939 0.929 0.944 0.021/0.040 0.045/0.026 0.032/0.024 6.491 6.597 6.912
q¼1 1 0.942 0.939 0.943 0.021/0.037 0.030/0.030 0.030/0.027 6.244 6.483 6.501
5 0.937 0.936 0.939 0.028/0.034 0.035/0.029 0.022/0.035 9.813 9.919 10.017
15 0.932 0.930 0.935 0.033/0.035 0.037/0.033 0.028/0.032 15.438 15.445 15.742
q¼2 1 0.947 0.944 0.945 0.022/0.031 0.028/0.028 0.027/0.028 7.507 7.685 7.686
5 0.942 0.941 0.943 0.028/0.030 0.031/0.028 0.030/0.027 13.220 13.307 13.399
15 0.938 0.937 0.940 0.030/0.031 0.033/0.030 0.032/0.028 21.659 21.752 21.941
T ¼ 500
q ¼ 0.1 1 0.948 0.940 0.950 0.006/0.045 0.033/0.026 0.023/0.027 4.575 4.697 4.707
5 0.948 0.939 0.947 0.011/0.041 0.036/0.025 0.028/0.025 5.184 5.272 5.314
15 0.946 0.937 0.946 0.019/0.035 0.039/0.024 0.031/0.024 6.455 6.507 6.631
q¼1 1 0.947 0.947 0.948 0.020/0.033 0.020/0.033 0.027/0.026 6.338 6.492 6.472
5 0.948 0.948 0.948 0.024/0.028 0.029/0.023 0.027/0.025 10.073 10.181 10.137
15 0.947 0.946 0.947 0.025/0.027 0.029/0.024 0.028/0.025 15.952 15.983 15.957
q¼2 1 0.944 0.945 0.944 0.026/0.030 0.027/0.028 0.029/0.026 7.554 7.648 7.636
5 0.947 0.947 0.948 0.026/0.027 0.028/0.025 0.028/0.024 13.369 13.447 13.466
15 0.947 0.947 0.948 0.026/0.027 0.028/0.025 0.027/0.025 21.968 21.992 22.085

change in the United StatesÕ home equity debt outstanding, unscheduled payments,
observed from the first quarter of 1991 to the second quarter of 2007 (Mortgages)
and measured in USD Billions.4 We use the observations up to the first quarter of
2001, T ¼ 61, to estimate the local level model, leaving the rest to evaluate the
out-of-sample forecast performance of the procedure. The QML estimates of the
parameters are r^2e ¼ 0:126 and ^q ¼ 0:671. These estimates are used in the Kalman
filter to obtain estimates of the innovations and their variances. Figure 3 plots the
correlogram and a kernel estimate of the density of the within-sample standardized
one-step-ahead errors. The correlations and partial correlations are not significant.
However, the density of the errors suggests that they are obviously far from
normality. Therefore, although the local level model seems appropriate to represent
the dependencies in the conditional mean of the Mortgages series, it is convenient to
implement a prediction procedure that takes into account the non-normality of the
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
 BOOTSTRAP PREDICTION INTERVALS IN SS MODELS 175
Empirical SSB ST WS
q = 0.1 q=1 q=2

0.6
0.3
0.2
0.4
T = 50

0.2
0.1
0.2 0.1

0 0 0
−4 −2 0 2 4 6 8 −5 0 5 −5 0 5 10

0.6
0.3
0.2
0.4
T = 100

0.2
0.1
0.2 0.1

0 0 0
−4 −2 0 2 4 6 8 −5 0 5 −5 0 5 10

0.6
0.3
0.2
0.4
T = 500

0.2
0.1
0.2 0.1

0 0 0
−4 −2 0 2 4 6 8 −5 0 5 −5 0 5 10

Figure 2. Kernel estimates densities of yTþk for k ¼ 1. v2ð1Þ case.

errors. We construct prediction intervals up to five-steps ahead using the ST, WS

and SSB procedures. The resulting intervals are plotted in Figure 4 together with the
observed values of the Mortgages series. First, observe that the two bootstrap
procedures generate very similar intervals which are wider than the ST intervals as
expected, given that they incorporate the uncertainty caused by parameter
estimation. Second, note that the observations corresponding to the second
quarter of 2006 and the first quarter of 2007, fall outside the ST prediction intervals.
However, both bootstrap intervals still contain these two values. It is important to
note that although bootstrap procedures are computationally intensive, in this
application with B ¼ 2000 bootstrap replicates, the SSB procedure requires
110 seconds using a MATLAB algorithm in an AMD Athlon 2.00 GHz
processor of a PC desktop with 2.00 Gb of RAM. The WS procedure requires
160 seconds. There is a reduction of 31% in the computer time required by the new
procedure proposed in this article.

4. CONCLUSIONS

This article proposes a new procedure to obtain bootstrap prediction intervals in

the context of state–space models. It is based on obtaining the density of future
observations in a single step that incorporates simultaneously the uncertainties
Ó 2009 The Authors
Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
176 A. RODRIGUEZ AND E. RUIZ

(a) Sample (ACF)

0.6
0.4
0.2
0
−0.2
−0.4
0 2 4 6 8 10 12 14 16 18 20
Lag
Sample partial (ACF)
0.6
0.4
0.2
0
−0.2
−0.4
0 2 4 6 8 10 12 14 16 18 20
Lag
(b) 0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
−4 −3 −2 −1 0 1 2 3 4 5

Figure 3. (a) Sample autocorrelations and partial autocorrelations of standardized one-step-ahead

errors. (b) Empirical density and histogram of the standardized one-step-ahead error.

4
Observed series SSB PI STD PI WS PI
3.5

2.5

1.5

0.5

0
2005:Q3 2005:Q4 2006:Q1 2006:Q2 2006:Q3 2006:Q4 2007:Q1

Figure 4. Prediction intervals for the out-of-sample forecasts of the Mortgage series.

Ó 2009 The Authors

Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
 BOOTSTRAP PREDICTION INTERVALS IN SS MODELS 177
caused by parameter estimation and error distribution. More importantly, our
bootstrap procedure does not rely on the backward representation. Consequently,
it is computationally very simple and can be extended to models without such
representation.
We show that our procedure, although much simpler, has slightly better finite-
sample properties than the bootstrap intervals of Wall and Stoffer (2002). As
expected, we also show that bootstrap intervals are more adequate than standard
intervals mainly in the presence of non-normal errors.
Finally, our proposed bootstrap procedure is implemented to obtain intervals
for future values of a series of Mortgages. In this case, the two bootstrap intervals
considered in this article are very similar. However, there is an important
improvement in terms of computer time when implementing our proposed
procedure.
When fitting state–space models to represent the dynamic evolution of a time
series, it is often of interest to obtain predictions not only of future values of the
series but also of future unobserved states. We are working on the adequacy of the
proposed bootstrap prediction intervals when implemented with this goal. An
issue left for further research is the implementation of the proposed procedure
when the system matrices are time-varying.

acknowledgements

Financial support from Project SEJ2006-03919 by the Spanish Government is

gratefully acknowledged. We are very grateful to the helpful comments of
Lorenzo Pascual and an anonymous referee. The usual disclaimer applies.

notes

1. We try alternative methods such as the bias-corrected and the acceleration

bias-corrected with similar results; see Efron (1987) for a definition of these
intervals.
2. We are particularly interested in dealing with this distribution because of its
relation with the linear transformation of the autoregressive stochastic
volatility model; see, for instance, Harvey et al. (1994). Results for other
distributions are similar and are not reported to save space. They are available
from the authors upon request.
3. All the results have been obtained using MATLAB, version 7.2, programs
developed by the first author.
4. The data have been downloaded from EcoWin.

Corresponding author: Esther Ruiz, Department of Statistics, Universidad Carlos

III de Madrid, Getafe, Spain. E-mail: [email protected]

Ó 2009 The Authors

Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2
178 A. RODRIGUEZ AND E. RUIZ

REFERENCES

Durbin, J. and Koopman, S. J. (2001) Time Series Analysis by State Space Methods. New York:
Oxford University Press.
Efron, B. (1987) Better bootstrap confidence intervals. Journal of the American Statistical Association
82, 171–85.
Harvey, A. C., Ruiz, E. and Shephard, N. G. (1994) Multivariate stochastic variance models. The
Review of Economic Studies 61, 247–64.
Pascual, L., Romo, J. and Ruiz, E. (2004) Bootstrap predictive inference for ARIMA processes.
Journal of Time Series Analysis 25, 449–65.
Pfeffermann, D. and Tiller, R. (2005) Bootstrap approximation to prediction MSE for state-space
models with estimated parameters. Journal of Time Series Analysis 26, 893–916.
Thombs, L. A. and Schucany, W. R. (1990) Bootstrap prediction intervals for autoregression. Journal
of the American Statistical Association 85, 486–92.
Wall, K. D. and Stoffer, D. S. (2002) A state space approach to bootstrapping conditional forecasts
in ARMA models. Journal of Time Series Analysis 23, 733–51.

Ó 2009 The Authors

Journal compilation Ó 2009 Blackwell Publishing Ltd.
JOURNAL OF TIME SERIES ANALYSIS Vol. 30, No. 2