Available online at www.sciencedirect.

com

ScienceDirect
Mathematics and Computers in Simulation 120 (2016) 91103
www.elsevier.com/locate/matcom

Original articles

Bootstrap prediction in univariate volatility models

with leverage effect
Carlos Trucos, Luiz K. Hotta
Department of Statistics, University of Campinas, IMECC-UNICAMP, Rua Sergio Buarque de Holanda 651, Cidade Universitaria,
Barao Geraldo, CEP 13083-859, Campinas, SP, Brazil

Received 3 July 2013; received in revised form 30 June 2015; accepted 6 July 2015
Available online 13 July 2015

Abstract

The EGARCH and GJR-GARCH models are widely used in modeling volatility when a leverage effect is present in the data.
Traditional methods of constructing prediction intervals for time series normally assume that the model parameters are known,
and the innovations are normally distributed. When these assumptions are not true, the prediction interval obtained usually has the
wrong coverage. In this article, the Pascual, Romo and Ruiz (PRR) algorithm, developed to obtain prediction intervals for GARCH
models, is adapted to obtain prediction intervals of returns and volatilities in EGARCH and GJR-GARCH models. These adjust-
ments have the same advantage of the original PRR algorithm, which incorporates a component of uncertainty due to parameter
estimation and does not require assumptions about the distribution of the innovations. The adaptations show good performance in
Monte Carlo experiments. However, the performance, especially in volatility prediction, can be very poor in the presence of an
additive outlier near the forecasting origin. The algorithms are applied to the daily returns series of the GBP/USD exchange rates.
c 2015 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.

Keywords: Interval prediction; Volatility interval prediction; Interval prediction and outlier; Interval prediction in EGARCH model; Interval
prediction in GJR-GARCH model

1. Introduction

The prediction of future values is a key objective in time series analysis, and it is of interest in many areas of
knowledge, such as economics, finance, production planning, and sales forecasting. The importance stems from the
fact that it is advantageous to know the likely evolution of the series in the future.
Generally, these predictions are given as point estimates, although the prediction interval is even more impor-
tant [17]. Nevertheless, authors of textbooks on time series analysis and forecasting generally devote little attention to
prediction intervals and give little guidance on how to calculate them [5, p. 479]. Also, in general prediction intervals
are calculated under the assumption that the model is known and errors are normally distributed.
In the financial time series literature there is little work on procedures to obtain prediction intervals for return and
volatility in the GARCH family. Moreover, some stylized facts like (conditional) innovation distribution with heavy

Corresponding author. Tel.: +55 19 35216081.

E-mail addresses: ctrucios@gmail.com (C. Trucos), hotta@ime.unicamp.br (L.K. Hotta).

http://dx.doi.org/10.1016/j.matcom.2015.07.001
0378-4754/ c 2015 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
92 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103

tail and asymmetry, and the leverage effect affect the coverage of the traditional intervals, making the predictions
inadequate and generally leading to a greater risk than is desirable.
An alternative to solve this problem is to obtain prediction intervals using the bootstrap procedure, which does
not require any assumption on the distribution of the innovations [7]. In recent decades, several works have proposed
bootstrap procedures to construct intervals for time series prediction. In a seminal work in this area, [21] constructed
intervals for autoregressive models using bootstrap replicates and fixing the latest p observations, where p is the
autoregressive order, [20] also constructed prediction intervals for autoregressive models using bootstrap replicates
but start the bootstrap with a randomly chosen block of size p from the observed series. In the financial time series
field, we can mention the work of [14] proposing a bootstrap procedure to obtain intervals for forecasting in ARCH
processes; [19], who obtained prediction intervals for ARCH models; [18], who extended the procedure presented
by [16] to ARIMA models for predicting volatility and return densities in GARCH processes; [6], who proposed new
methods for prediction intervals of return and volatility in ARCH and GARCH models; [13], who used bootstrap sub
sampling for interval prediction in GARCH models; among others.
One bootstrap method for prediction intervals for volatility models that has shown good results, and that appears
to be generalizable to other models is the method proposed by [18] (PRR) for GARCH models. This method incorpo-
rates the uncertainty of the estimation in the forecasting interval, since the parameters are estimated at each bootstrap
replication. The method also does not depend on the (conditional) innovation distribution. This paper proposes an
adaptation of the PRR algorithm developed for prediction intervals for GARCH models, to get prediction intervals for
EGARCH and GJR-GARCH models. The paper also studies the effect of additive outliers on the proposed prediction
intervals.
The paper is organized as follows: Section 2 introduces the volatility models, Section 3 presents the bootstrap
procedures for obtaining prediction intervals for EGARCH and GJR-GARCH models. Section 4 presents the results
obtained by simulation; and Section 5 presents an application of the proposed procedures to the daily returns series of
the GBP/USD exchange rates. Section 6 concludes.

2. The EGARCH and GJR-GARCH models

GARCH models [3] have been widely used in modeling volatility. Based on this model, other models have been
proposed to incorporate other stylized facts, such as the leverage effect. In this sense, we mention the EGARCH [15]
and GJR-GARCH models [10]. Given its popularity in empirical applications, in this paper we focus on the
EGARCH(1,1) and GJR-GARCH(1,1) models.

Definition 2.1 (Univariate EGARCH Model). An EGARCH(1,1) process, {rt }, is defined as:

rt = t t ,
(2.1)
log(t2 ) = + t1 + (|t1 | E|t1 |) + log(t1
2
),

where , , , , are real numbers, and t I I D(0, 1) (independent and identically distributed random variables
with zero mean and unit variance). Particularly, when || < 1, the EGARCH model is stationary if t comes from
Gaussian or Generalized Error Distribution (GED) with shape parameter >1. Nevertheless, if t comes from Student-t
or a GED distribution with shape parameter 1, the EGARCH model is stationary if ||. An advantage of this
model is that there is no restriction imposed to ensure that the variance is positive.

Definition 2.2 (GJR-GARCH Model). A GJR-GARCH(1,1) process, {rt }, is defined as:

rt = t t ,
(2.2)
t2 = + rt1
2
+ t1
2
+ rt1
2
I (rt1 < 0),

where, I () is the indicator function, t I I D(0, 1), > 0 and , , are non-negative real numbers for ensuring
positive t2 . The model is stationary if < 2(1 ).
 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103 93

3. Bootstrap prediction intervals

This section presents an adaptation of the PRR algorithm proposed by [18] for the GARCH process for application
in EGARCH and GJR-GARCH models. These modifications keep the original idea of PRR, i.e., incorporating the
element of uncertainty in parameter estimation and making no assumptions about the distribution of the innovations.

3.1. Algorithm for GJR-GARCH models

Consider RT a sequence of T observations generated by a GJR-GARCH(1,1) process. The algorithm is described

for a GJR-GARCH(1,1) process, but it is generalized to a GJR-GARCH(p,q) process in a straightforward way.
Step 1: Estimate the model with the observed data by the chosen method. Denote the estimates as: = (, ,
,
),
and calculate the residuals. Denote by FT the empirical distribution of the centered and re-scaled residuals.
Step 2: Generate a bootstrap series rt , t = 2, . . . , T using the following recursion:

t2 = + r 2
t1 + t1
2
+ rt1
2
I (rt1

< 0),
(3.3)
rt = t t ,

with 1 = 1 , r1 = 1 1 and where t i.i.d FT . Adjust the GJR-GARCH(1,1) model to the bootstrap sequence
RT to obtain the bootstrap estimates = ( , , , ) using the same method used in step 1.
Step 3: Calculate forecasts of returns and volatilities h steps ahead, h = 1, 2, . . . using the following recursion:
T2 (h) = + r T2 (h 1) + T2 (h 1) + r T2 (h 1)I (r T (h 1) < 0),
(3.4)
r T (h) = T (h) T (h),
where T (h) i.i.d FT , r T = r T and T2 (0) is obtained from the following recursion for t = 2, . . . , T.

t2 (0) = + rt1
2
+ t1
2
(0) + rt1
2
I (rt1 < 0), (3.5)
with 1 (0) = 1 .
(1) (B)
Step 4: Repeat steps 2 and 3, B times to obtain B bootstrap replicates (r T (h), . . . , r T (h)) and
(1) (B)
(T (h), . . . , T (h)). Then a 100(1 )% prediction interval for r T (h) is given by:
( ), Q (1 )],
[Q r,B 2 r,B 2
(a) = G 1 (a), and G (a) = #(r Tb (a)a)
where: Q r,B r,B r,B B . Similarly, obtain a prediction interval for the volatility.
Note that, T2 (0) in step 3 is different in each bootstrap replication, because of the use of the bootstrap parameters.
This allows to construct prediction intervals for volatilities one step ahead.

3.2. Algorithm for EGARCH models

Let RT be a sequence of T observations generated by the EGARCH(1,1) process. Steps 1, 2 and 4 are similar to
those described in the previous algorithm. Step 3 is modified as follows:
Step 3: Calculate forecasts of returns and volatilities h steps ahead, h = 1, 2, . . . using the following recursion:

r T (h) r T (h 1)

log( T (h)) = +
2
+ M + log( T2 (h 1)),

T (h) T (h 1) (3.6)
r T (h) = T (h) T (h),

where T (h) i.i.d FT , M = 2
, r T = r T and log( T2 (0)) is obtained from the following recursion for
t = 2, . . . , T,

rt1 rt1
log( t2 (0)) = +

+ |

| M + log( t1
2
(0)), (3.7)
t1 (0) t1 (0)
where log( 12 (0)) = log( 12 ).
94 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103

Table 1
Estimated coverage of the 95% (top panel) and 99% (bottom panel) return bootstrap prediction intervals. GJR-GARCH(1,1) model.

Horizon Innovation Average Std. Average Average Average Emp.Ave.

distribution coverage dev. below int. above int. length length
Normal 94.58 1.58 2.70 2.72 1.7267 1.7361
Student-t 94.48 1.55 2.72 2.81 1.6863 1.7088
1
SN 95.12 1.45 2.41 2.47 1.8471 1.8042
ST 95.09 1.61 2.45 2.46 1.9501 1.8927
Normal 94.55 1.51 2.69 2.77 1.7376 1.7497
Student-t 94.50 1.59 2.69 2.81 1.6966 1.7193
2
SN 95.20 1.40 2.37 2.43 1.8711 1.8232
ST 95.11 1.57 2.41 2.48 1.9669 1.9096
Normal 94.54 1.57 2.68 2.78 1.7458 1.7619
Student-t 94.49 1.55 2.70 2.81 1.7075 1.7365
3
SN 95.27 1.36 2.35 2.38 1.8921 1.8361
ST 95.10 1.58 2.40 2.50 1.9856 1.9204
Normal 94.48 1.58 2.72 2.80 1.7521 1.7687
Student-t 94.47 1.58 2.68 2.86 1.7144 1.7423
4
SN 95.27 1.42 2.36 2.37 1.9130 1.8501
ST 95.10 1.63 2.45 2.45 2.0020 1.9348
Normal 94.48 1.54 2.70 2.82 1.7594 1.7722
Student-t 94.45 1.61 2.71 2.84 1.7219 1.7481
5
SN 95.31 1.42 2.33 2.37 1.9271 1.8650
ST 95.12 1.64 2.42 2.46 2.0120 1.9482

Normal 98.75 0.69 0.60 0.66 2.2732 2.2685

Student-t 98.71 0.65 0.60 0.69 2.5070 2.5015
1
SN 98.94 0.62 0.48 0.57 2.4378 2.3506
ST 98.87 0.65 0.54 0.58 2.8814 2.7557
Normal 98.73 0.70 0.58 0.70 2.3175 2.3137
Student-t 98.72 0.69 0.58 0.71 2.5447 2.5550
2
SN 98.99 0.58 0.47 0.54 2.5276 2.4154
ST 98.90 0.61 0.52 0.57 2.9757 2.8355
Normal 98.74 0.68 0.57 0.69 2.3528 2.3532
Student-t 98.70 0.65 0.61 0.69 2.5889 2.5936
3
SN 99.06 0.54 0.44 0.50 2.6148 2.4636
ST 98.90 0.67 0.51 0.59 3.0770 2.8998
Normal 98.72 0.70 0.58 0.69 2.3965 2.3883
Student-t 98.70 0.66 0.58 0.71 2.6302 2.6288
4
SN 99.05 0.55 0.45 0.50 2.6826 2.5260
ST 98.95 0.60 0.51 0.55 3.1549 2.9674
Normal 98.71 0.71 0.58 0.71 2.4294 2.4211
Student-t 98.71 0.68 0.58 0.70 2.6670 2.6584
5
SN 99.05 0.56 0.45 0.50 2.7340 2.5741
ST 98.94 0.66 0.50 0.56 3.2173 3.0107

4. Simulation study
Here we analyze the performance of the suggested algorithm through Monte Carlo simulation. We consider the
GJR-GARCH(1,1) and EGARCH(1,1) models with normal, Student-t, skew-normal and skew Student-t (skew-t)
distributions for the innovations. The models are:
GJR-GARCH(1,1):

t2 = 0.01 + 0.05rt1
2 2
+ 0.15rt1 I (rt1 < 0) + 0.83t1
2
, (4.8)
EGARCH(1,1):

log(t2 ) = 0.35 0.15t1 + 0.20(|t1 | E(|t1 |)) + 0.90log(t1

2
) (4.9)
 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103 95

Table 2
Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility bootstrap prediction intervals. GJR-GARCH(1,1) model.

Horizon Innovation Average Std. Average Average Average Emp.Ave.

distribution coverage dev. below int. above int. length length
Normal 94.10 23.57 1.30 4.60 0.0820
Student-t 91.80 27.45 1.60 6.60 0.1017
1
SN 93.80 24.13 4.90 1.30 0.1034
ST 93.60 24.49 3.70 2.70 0.1730
Normal 95.43 8.18 1.25 3.33 0.2008 0.1526
Student-t 94.77 10.12 1.91 3.32 0.2119 0.1552
2
SN 93.71 13.76 4.18 2.11 0.3158 0.1821
ST 94.47 13.03 3.22 2.31 0.4549 0.2021
Normal 95.10 6.32 1.52 3.37 0.2579 0.2129
Student-t 94.42 8.82 2.08 3.50 0.2779 0.2263
3
SN 94.38 10.15 3.51 2.11 0.4168 0.2532
ST 94.94 10.11 2.67 2.39 0.6212 0.2964
Normal 94.82 5.60 1.73 3.45 0.2982 0.2563
Student-t 94.26 7.87 2.15 3.60 0.3277 0.2776
4
SN 94.71 8.15 3.20 2.09 0.4917 0.3032
ST 95.11 8.61 2.43 2.45 0.7446 0.3631
Normal 94.61 5.26 1.92 3.47 0.3340 0.2921
Student-t 94.05 7.43 2.25 3.70 0.3652 0.3172
5
SN 94.84 7.06 3.05 2.12 0.5566 0.3453
ST 95.09 7.82 2.41 2.50 0.8397 0.4175

Normal 98.60 11.75 0.10 1.30 0.1090

Student-t 97.70 15.00 0.60 1.70 0.1361
1
SN 98.90 10.44 0.80 0.30 0.1382
ST 98.10 13.66 1.20 0.70 0.2325
Normal 98.89 3.45 0.25 0.86 0.3362 0.2400
Student-t 98.59 5.69 0.57 0.84 0.4198 0.2884
2
SN 98.53 6.82 1.01 0.46 0.5587 0.2925
ST 98.67 6.38 0.74 0.59 0.9642 0.3914
Normal 98.82 2.86 0.29 0.89 0.4253 0.3140
Student-t 98.57 4.57 0.53 0.90 0.5524 0.3869
3
SN 98.61 5.21 0.91 0.48 0.7222 0.3824
ST 98.79 4.66 0.60 0.61 1.3218 0.5242
Normal 98.78 2.69 0.33 0.89 0.5030 0.3749
Student-t 98.45 4.18 0.57 0.98 0.6450 0.4598
4
SN 98.72 4.13 0.81 0.47 0.8830 0.4530
ST 98.73 4.24 0.64 0.63 1.6225 0.6247
Normal 98.73 2.39 0.37 0.90 0.5717 0.4259
Student-t 98.39 3.78 0.59 1.02 0.7284 0.5193
5
SN 98.82 3.46 0.72 0.47 1.0189 0.5183
ST 98.76 3.90 0.61 0.63 1.8298 0.7088

where t are innovations. The parameters have been chosen to resemble the parameters values often estimated when
the models are fitted to real series. The skew-normal distribution is denoted by SN() and the skew-t distribution
is denoted by ST(, ), where () is the skewness parameter and is the degrees of freedom. The distributions
are parameterized as in [8], such that, when = 1, the distribution is symmetric. In the GJR-GARCH(1,1) and
EGARCH(1,1) models we consider innovations with SN(0.8) and ST(0.8, 7) distributions. For each model, we ran
1000 replications, each involving three steps. For the kth replication, the steps are:
1. Step 1: Generate a series with size T = 1000. That is generate {rt , t = 1, . . . , 1000} and {t , t = 1, . . . , 1000}.
2. Step 2: Run steps 14 given in Section 3.1 for the GJR-GARCH(1,1) model, and in Section 3.2 for the
EGARCH(1,1) model to find the h = 1, 2, 3, 4, 5 steps-ahead 100(1 )% prediction intervals for the returns
and volatilities.
96 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103

Table 3
Estimated coverage of the 95% (top panel) and 99% (bottom panel) return bootstrap prediction intervals. GJR-GARCH(1,1) model with one
additive outlier at the kth observation. Distribution of the innovations: Gaussian.
Horizon Outlier Average Std. Average Average Average Emp.Ave.
position coverage dev. below int. above int. length length
500 94.63 1.63 2.68 2.70 1.7286 1.7361
1
999 99.00 1.38 0.51 0.49 2.6644 1.7361
500 94.58 1.57 2.66 2.76 1.7398 1.7497
2
999 98.83 1.48 0.57 0.60 2.6284 1.7497
500 94.58 1.63 2.64 2.77 1.7480 1.7619
3
999 98.66 1.56 0.65 0.68 2.5891 1.7619
500 94.50 1.65 2.71 2.79 1.7532 1.7687
4
999 98.53 1.62 0.73 0.74 2.5554 1.7687
500 94.51 1.58 2.68 2.81 1.7606 1.7722
5
999 98.40 1.65 0.79 0.81 2.5232 1.7722

500 98.82 0.70 0.56 0.62 2.3047 2.2685

1
999 99.87 0.28 0.07 0.06 3.5700 2.2685
500 98.82 0.69 0.53 0.65 2.3494 2.3137
2
999 99.84 0.31 0.08 0.08 3.5774 2.3137
500 98.83 0.70 0.53 0.65 2.3919 2.3532
3
999 99.80 0.35 0.10 0.11 3.5770 2.3532
500 98.81 0.69 0.54 0.65 2.4309 2.3883
4
999 99.77 0.37 0.11 0.12 3.5991 2.3883
500 98.80 0.71 0.54 0.66 2.4668 2.4211
5
999 99.74 0.39 0.12 0.14 3.5965 2.4211

Table 4
Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility bootstrap prediction intervals. GJR-GARCH(1,1) model one additive
outlier at the kth observation. Distribution of the innovations: Gaussian.
Horizon Outlier Average Std. Average Average Average Emp. Ave.
position coverage dev. below int. above int. length length
500 93.80 24.13 2.30 3.90 0.0865
1
999 27.20 44.52 72.80 0.00 0.3217
500 94.43 10.36 2.38 3.19 0.2025 0.1526
2
999 30.33 38.83 69.44 0.23 0.5561 0.1526
500 94.06 8.53 2.64 3.30 0.2599 0.2129
3
999 34.22 36.99 65.45 0.33 0.6746 0.2129
500 93.90 7.44 2.75 3.35 0.3010 0.2563
4
999 38.41 35.79 61.18 0.42 0.7562 0.2563
500 93.72 6.84 2.89 3.38 0.3361 0.2921
5
999 42.98 34.68 56.52 0.50 0.8207 0.2921

500 98.60 11.75 0.40 1.00 0.1150

1
999 39.70 48.95 60.30 0.00 0.4231
500 98.71 4.84 0.47 0.82 0.3442 0.2400
2
999 41.93 44.48 58.02 0.05 0.9552 0.2400
500 98.54 4.45 0.68 0.78 0.4445 0.3140
3
999 45.80 42.36 54.14 0.07 1.1826 0.3140
500 98.52 3.76 0.69 0.79 0.5293 0.3749
4
999 50.05 40.18 49.86 0.09 1.3606 0.3749
500 98.47 3.39 0.72 0.81 0.5996 0.4259
5
999 54.94 37.70 44.95 0.11 1.5128 0.4259
 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103 97

Table 5
Estimated coverage of the 95% (top panel) and 99% (bottom panel) return bootstrap prediction intervals. EGARCH(1,1) model.

Horizon Innovation Average Std. Average Average Average Emp. Ave

distribution coverage dev. below int. above int. length length
Normal 94.73 1.55 2.59 2.68 3.7460 3.7388
Student-t 94.88 1.59 2.53 2.59 3.8918 3.8395
1
SN 94.82 1.61 2.55 2.63 3.7465 3.7232
ST 94.72 1.65 2.58 2.70 3.8318 3.8040
Normal 94.70 1.52 2.61 2.69 3.7558 3.7577
Student-t 94.83 1.53 2.55 2.63 3.9014 3.8610
2
SN 94.84 1.50 2.53 2.63 3.7693 3.7449
ST 94.78 1.57 2.56 2.66 3.8494 3.8225
Normal 94.75 1.47 2.59 2.67 3.7736 3.7702
Student-t 94.82 1.53 2.52 2.66 3.9040 3.8708
3
SN 94.84 1.52 2.56 2.60 3.7849 3.7626
ST 94.89 1.60 2.52 2.60 3.8760 3.8241
Normal 94.75 1.43 2.58 2.67 3.7864 3.7807
Student-t 94.83 1.49 2.56 2.61 3.9089 3.8774
4
SN 94.81 1.50 2.54 2.65 3.7915 3.7712
ST 94.78 1.56 2.58 2.64 3.8738 3.8453
Normal 94.80 1.40 2.57 2.63 3.8028 3.7883
Student-t 94.85 1.45 2.53 2.62 3.9194 3.8902
5
SN 94.83 1.44 2.54 2.63 3.8077 3.7853
ST 94.81 1.55 2.58 2.61 3.8915 3.8528

Normal 98.80 0.70 0.56 0.64 4.9343 4.8577

Student-t 98.82 0.67 0.54 0.63 5.7941 5.6045
1
SN 98.80 0.70 0.56 0.64 4.9285 4.8506
ST 98.79 0.67 0.57 0.64 5.7147 5.5414
Normal 98.79 0.66 0.55 0.65 4.9843 4.9275
Student-t 98.79 0.67 0.55 0.65 5.8104 5.6750
2
SN 98.84 0.67 0.54 0.61 5.0057 4.9161
ST 98.79 0.68 0.57 0.65 5.7925 5.6214
Normal 98.83 0.64 0.55 0.62 5.0465 4.9784
Student-t 98.85 0.61 0.52 0.63 5.8913 5.7134
3
SN 98.84 0.65 0.55 0.60 5.0712 4.9745
ST 98.84 0.66 0.54 0.62 5.8223 5.6470
Normal 98.85 0.59 0.53 0.62 5.0911 5.0075
Student-t 98.84 0.62 0.54 0.62 5.9051 5.7344
4
SN 98.84 0.63 0.54 0.62 5.1039 5.0146
ST 98.82 0.65 0.55 0.63 5.8908 5.7090
Normal 98.83 0.63 0.54 0.63 5.1188 5.0468
Student-t 98.85 0.63 0.52 0.64 5.9404 5.7524
5
SN 98.86 0.61 0.54 0.60 5.1519 5.0364
ST 98.84 0.64 0.56 0.60 5.9272 5.7378

3. Step 3: Generate 1000 sets of future values {r Ti + j , j = 1, . . . , 5, i = 1, . . . , 1000} and {Ti + j , j = 1, . . . , 5, i =

1, . . . , 1000} considering that the previous observations are the same as given in step 1. For h = 1, 2, 3, 4, 5,
denote by pr,h,k , pr,h,b,k and pr,h,a,k the proportions of the values {r Ti +h , i = 1, . . . , 1000} which are inside the
prediction interval found in step 2, below the lower limit of the prediction interval and above the prediction interval,
respectively. Define similarly p,h,k , p,h,b,k and p,h,a,k for the volatilities.

We consider the cases where the series has no outliers and where the series has an additive outlier at the end and
in the middle of the series. Intervals with 95% and 99% nominal coverage are constructed. For sake of comparison
we also compute the 95% and 99% average empirical length of interval for returns and volatilities calculated as the
98 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103

Table 6
Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility bootstrap prediction intervals. EGARCH(1,1) model.

Horizon Innovation Average Std. Average Average Average Emp. Ave.

distribution coverage dev. below int. above int. length length
Normal 95.10 21.60 1.00 3.90 0.1770
Student-t 92.90 25.70 2.30 4.80 0.2487
1
SN 95.20 21.39 1.90 2.90 0.1779
ST 92.50 26.35 2.60 4.90 0.2505
Normal 95.89 7.62 1.15 2.95 0.3414 0.2475
Student-t 95.97 8.57 1.27 2.76 0.4091 0.2556
2
SN 95.37 9.27 1.69 2.94 0.3613 0.2707
ST 95.69 9.00 1.39 2.92 0.4357 0.2861
Normal 95.85 6.04 1.11 3.03 0.4123 0.3337
Student-t 95.96 6.92 1.19 2.85 0.4848 0.3526
3
SN 95.49 7.14 1.52 2.99 0.4382 0.3618
ST 95.55 7.46 1.33 3.12 0.5190 0.3932
Normal 95.69 5.41 1.25 3.06 0.4622 0.3960
Student-t 95.71 6.13 1.25 3.04 0.5332 0.4183
4
SN 95.30 6.15 1.60 3.10 0.4913 0.4280
ST 95.29 6.74 1.43 3.28 0.5728 0.4648
Normal 95.50 5.01 1.38 3.11 0.4998 0.4402
Student-t 95.57 5.82 1.31 3.12 0.5680 0.4663
5
SN 95.06 5.72 1.78 3.16 0.5300 0.4771
ST 95.12 6.31 1.55 3.34 0.6131 0.5187

Normal 98.40 12.55 0.50 1.10 0.2500

Student-t 98.00 14.01 0.40 1.60 0.3775
1
SN 98.60 11.75 0.50 0.90 0.2467
ST 98.20 13.30 0.60 1.20 0.3585
Normal 99.12 3.24 0.19 0.70 0.4975 0.3367
Student-t 99.27 2.96 0.16 0.57 0.6701 0.3988
2
SN 99.03 3.41 0.25 0.73 0.5308 0.3793
ST 99.06 3.79 0.23 0.70 0.7134 0.4649
Normal 99.12 2.31 0.11 0.77 0.5938 0.4558
Student-t 99.25 2.05 0.13 0.62 0.7765 0.5276
3
SN 99.00 2.71 0.24 0.75 0.6369 0.5027
ST 99.04 2.96 0.19 0.77 0.8389 0.6125
Normal 99.12 1.79 0.12 0.76 0.6622 0.5357
Student-t 99.20 1.60 0.12 0.67 0.8452 0.6101
4
SN 98.98 2.56 0.24 0.78 0.7104 0.5926
ST 99.00 2.38 0.18 0.82 0.9117 0.7058
Normal 99.10 1.62 0.15 0.75 0.7149 0.5956
Student-t 99.12 1.87 0.14 0.74 0.8895 0.6728
5
SN 98.95 2.06 0.24 0.81 0.7624 0.6541
ST 98.93 2.37 0.21 0.86 0.9664 0.7766

difference between the empirical quantiles of the simulated data. Note that for volatilities one-step-ahead we do not
present the empirical length because the volatility one-step-ahead is completely determined at time T.
In all situations, we consider the quasi maximum likelihood estimator (QMLE) with Gaussian distribution to
estimate the parameters.

4.1. Prediction interval for GJR-GARCH model

The prediction intervals for returns and volatilities were obtained for h = 1, 2, 3, 4, 5 steps forward for the 95%
and 99% nominal coverage. Tables 1 and 2 present the coverage mean, standard deviation, above and below wrong
coverages and the mean of the intervals length for return and volatility prediction intervals.
 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103 99

Table 7
Estimated coverage of the 95% (top panel) and 99% (bottom panel) return bootstrap prediction intervals. EGARCH(1,1) model with outlier at the
kth observation. Distribution of the innovations: Gaussian.
Horizon Outlier Average Std. Average Average Average Emp. Ave
position coverage dev. below int. above int. length length
500 94.83 1.59 2.53 2.64 3.7626 3.7388
1
999 97.47 2.18 1.26 1.27 4.5832 3.7388
500 94.79 1.55 2.56 2.65 3.7707 3.7577
2
999 97.21 2.06 1.37 1.42 4.4783 3.7577
500 94.86 1.50 2.52 2.62 3.7919 3.7702
3
999 97.03 1.98 1.47 1.50 4.4025 3.7702
500 94.86 1.44 2.52 2.62 3.8045 3.7807
4
999 96.90 1.83 1.53 1.57 4.3420 3.7807
500 94.89 1.42 2.52 2.59 3.8216 3.7883
5
999 96.74 1.77 1.62 1.64 4.2959 3.7883

500 98.89 0.71 0.50 0.60 5.0099 4.8577

1
999 99.55 0.64 0.21 0.23 6.1857 4.8577
500 98.89 0.66 0.50 0.61 5.0711 4.9275
2
999 99.51 0.62 0.23 0.26 6.0812 4.9275
500 98.93 0.62 0.50 0.57 5.1403 4.9784
3
999 99.49 0.60 0.25 0.27 6.0042 4.9784
500 98.95 0.59 0.47 0.58 5.1829 5.0075
4
999 99.47 0.52 0.26 0.27 5.9390 5.0075
500 98.92 0.61 0.49 0.59 5.1987 5.0468
5
999 99.41 0.55 0.28 0.31 5.8956 5.0468

Table 8
Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility bootstrap prediction intervals. EGARCH(1,1) model with outlier at
the kth observation. Distribution of the innovations: Gaussian.
Horizon Outlier Average Std. Average Average Average Emp. Ave.
position coverage dev. below int. above int. length length
500 94.30 23.20 2.50 3.20 0.1851
1
999 52.40 49.97 47.40 0.20 0.3700
500 95.55 8.99 1.69 2.75 0.3470 0.2475
2
999 56.57 42.92 42.79 0.64 0.4958 0.2475
500 95.44 7.02 175 2.81 0.4176 0.3337
3
999 62.60 37.33 36.58 0.81 0.5471 0.3337
500 95.31 6.22 1.86 2.84 0.4673 0.3960
4
999 68.38 31.78 30.65 0.97 0.5807 0.3960
500 95.13 5.68 1.96 2.91 0.5034 0.4402
5
999 73.80 26.45 25.10 1.10 0.6047 0.4402

500 98.60 11.75 0.40 1.00 0.2667

1
999 65.90 47.43 34.10 0.00 0.6421
500 99.17 3.64 0.24 0.59 0.5137 0.3367
2
999 69.42 40.18 30.45 0.13 0.8507 0.3367
500 99.25 1.88 0.12 0.63 0.6137 0.4558
3
999 76.14 32.85 23.68 0.18 0.9123 0.4558
500 99.21 1.50 0.13 0.66 0.6812 0.5357
4
999 81.58 26.32 18.21 0.20 0.9502 0.5357
500 99.12 1.57 0.21 0.67 0.7315 0.5956
5
999 86.07 20.69 13.70 0.23 0.9756 0.5956
100 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103

Table 9
Empirical coverage of the return and volatility prediction bootstrap prediction intervals for daily returns series of the GBP/USD exchange rates.
Nominal coverage 95%. * represents the average coverage of the MedRV point estimates.

Horizon Volatility Average Std. Average Average Average

model coverage deviation below int. above int. coverage*
Return
EGARCH 94.67 22.51 2.87 2.46
1
GJR 95.90 19.87 2.05 2.05
EGARCH 96.72 17.84 1.23 2.05
2
GJR 96.31 18.89 1.23 2.46
EGARCH 96.72 17.84 1.23 2.05
3
GJR 97.13 16.73 1.23 1.64
EGARCH 95.49 20.79 1.64 2.87
4
GJR 95.90 19.87 1.64 2.46
EGARCH 96.31 18.89 1.64 2.05
5
GJR 96.31 18.89 1.64 2.05
Volatility
EGARCH 72.95 44.51 19.26 7.79 40.57
1
GJR 71.31 45.32 23.36 5.33 31.56
EGARCH 81.56 38.86 14.75 3.69 53.28
2
GJR 76.64 42.40 21.31 2.05 40.16
EGARCH 84.84 35.94 12.70 2.46 57.79
3
GJR 79.92 40.14 18.44 1.64 46.72
EGARCH 87.70 32.91 11.07 1.23 66.39
4
GJR 80.33 39.83 18.44 1.23 52.87
EGARCH 88.52 31.94 10.66 0.82 69.67
5
GJR 82.38 38.18 17.21 0.41 58.20

In both cases, the bootstrap procedure performed well, obtaining coverage proportions near to the nominal values.
The proportions of above and below error coverages are close for the return. For both cases, return and volatility, the
interval with the largest length was with skew-t innovation.
The proposed bootstrap procedure to obtain prediction intervals for returns and volatilities in GJR-GARCH model
performs well and is not affected by the presence of asymmetry in innovation distribution. The estimated coverages
are close to the nominal for any innovation distribution: normal, Student-t, skew normal or skew-t.
The analysis of the results suggests that for 95% and 99% confidence intervals the bootstrap procedure presents
good results for all innovation distributions (normal, Student-t, skew-normal and skew-t).
For the volatility, in general, for any innovation distribution, the coverage standard deviation decreases when the
prediction horizon increases.
The effect of additive outliers, as defined in [11], is also considered. For a sample size of 1000 the outlier is added
at the 500th or the 999th position. Let y1 , . . . , y1000 be a series generated without outliers. When we have an addi-
tive outlier of size in the 999th observation, the observed series is y1 , . . . , y998 , y999 + sign(y999 ), y1000 . In
the simulation we considered equal to three times the sample standard deviation of {y1 , . . . , y1000 }. We selected a
small value because we want to show the effect of the outlier when the outlier is ignored. We show that even with such
outlier size, which is difficult to be detected, the effect is devastating. The distribution of the innovation is the Gaussian
distribution. The results are presented in Tables 3 and 4 for return and volatility, respectively. When the outlier occurs
in the middle series, its influence is almost null in the proportion of the coverage of returns and volatilities prediction
interval with a little increase in the length of the interval. However, when the outlier occurs near the end of the series,
both, for the return and the volatility, the length of the prediction intervals increases significantly. The increase in the
width of the intervals leads to a proportion of the coverage of the return larger than the nominal values, but the pro-
portion of the coverage of the volatility is much smaller than the nominal values, specially one-step-ahead, even with
the length of interval is larger. This happens because the outlier has a strong bias effect on the point volatility forecast.
 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103 101

Table 10
Empirical coverage of the return and volatility prediction bootstrap prediction intervals for the daily returns series of the GBP/USD exchange rates.
Nominal coverage 99%. * represents the average coverage of the MedRV point estimates.

Horizon Volatility Average Std. Average Average Average

model coverage deviation below int. above int. coverage*
Return
EGARCH 98.77 11.04 0.41 0.82
1
GJR 98.36 12.72 0.41 1.23
EGARCH 98.36 12.72 0.41 1.23
2
GJR 98.36 12.72 0.41 1.23
EGARCH 98.77 11.04 0.41 0.82
3
GJR 98.77 11.04 0.41 0.82
EGARCH 98.36 12.72 0.41 1.23
4
GJR 98.36 12.72 0.41 1.23
EGARCH 98.77 11.04 0.41 0.82
5
GJR 98.77 11.04 0.41 0.82
Volatility
EGARCH 84.02 36.72 12.30 3.69 53.28
1
GJR 75.82 42.91 20.90 3.28 42.21
EGARCH 89.34 30.92 8.61 2.05 66.80
2
GJR 81.15 39.19 18.03 0.82 58.61
EGARCH 92.62 26.19 6.56 0.82 72.95
3
GJR 83.20 37.47 16.80 0.00 60.25
EGARCH 92.62 26.19 7.38 0.00 78.69
4
GJR 83.61 37.10 16.39 0.00 65.57
EGARCH 94.67 22.51 5.33 0.00 81.56
5
GJR 84.84 35.94 15.16 0.00 66.39

4.2. Prediction interval for EGARCH model

Here the simulation done in Section 4.1 is repeated for the EGARCH model. The results without outliers are pre-
sented in Tables 5 and 6, and with outliers in Tables 7 and 8. The conclusions are very similar than before: in presence
of outlier the prediction interval for returns is larger than necessary, the performance of prediction intervals for the
volatilities is poor, the length of intervals increases significantly and in series uncontaminated by outlier the coverage
for returns and volatilities is close to the nominal value regardless of the innovation distribution used to simulate the
model. Thus, although the bootstrap does not take into account the distribution of the innovations, one must be careful
when there is an outlier near the prediction origin.
In summary, the presence of an additive outlier influences the construction of both the return and volatility pre-
diction intervals. When the outlier occurs at the end of the series, the error can be very large for the volatility, and
the analysis should be done carefully. Previous treatment of outliers can improve the performance of the bootstrap
prediction interval. Another solution can be to robustify the algorithm to obtain prediction intervals in presence of
additive outliers as in [22].

5. Application

Fig. 1 presents the daily returns series of the GBP/USD exchange rates in percentages from 5th January 2000 to
31st December 2013 (T = 3519 observations), the data is available at www.federalreserve.gov. The sample mean and
standard deviation are equal to 0.0004 and 0.6034, respectively. The high frequency data is provided by disktrading.
For empirical data, the volatility is not observed. In this case an approach is to use a realized measure as the bench-
mark value, as in [18]. [18] uses the realized volatility while we use the MedRV of [2], a robust realized measure. We
compute the MedRVs confidence intervals and analyze if the lower limit of the confidence interval is larger than the
upper limit of the bootstrap forecast interval or if the upper limit of the confidence intervals is smaller than the lower
102 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103

Fig. 1. Daily returns series of the GBP/USD exchange rates in percentage from 5th January 2008 to 31st December 2013.

limit of the bootstrap forecast interval. In this case, it is considered as above or below error coverage, respectively;
otherwise it is considered inside the interval.
To illustrate, the EGARCH(1,1) and GJR-GARCH(1,1) models are fitted using Gaussian quasi maximum like-
lihood. We construct 244 rolling windows and in each window h = 1, 2, 3, 4, 5 steps-ahead bootstrap prediction
intervals are computed for returns and volatilities. Tables 9 and 10 present the coverage of the 95% and 99% return
and volatility bootstrap prediction intervals. The coverage of the MedRV point estimates of the volatilities is given
in the last column of Tables 9 and 10. In general the results for EGARCH(1,1) model are better than for the GJR-
GARCH(1,1) model. The coverage of the return is approximately equal to the nominal values for 95% and 99% of
the cases. However, for the volatility, the coverage is smaller than the nominal values, mainly for short horizons. [18]
and [9] also reported a small coverage in their empirical application. This small coverage may be for any factors; for
example, presence of outliers in the sample. Studies on robust prediction intervals for GARCH process can be found
in [22].

6. Conclusion

This paper adapts the PRR algorithm proposed by [18] to GARCH models, to obtain prediction intervals for return
and volatility in EGARCH and GJR-GARCH models. The performances of the methods are evaluated by simulation.
We also considered symmetric and asymmetric distributions for the conditional innovation distribution. The proce-
dures also have a good performance in this case. However, when an additive outlier is inserted in the series, the effect
can be very large when the outlier is near the end of the series. Thus, the presence of an additive outlier must be treated
with care. Further researches considering the use of robust estimates of the model and the volatilities as considered
in [4] and test for the presence of outliers as in [11] are welcome. Robustification of the algorithm is also an important
further work.
The prediction intervals do not incorporate the model uncertainty. Two works for ARMA models which incorporate
model uncertainty are [12] and [1].

Acknowledgments

The authors acknowledge financial support from Sao Paulo Research Foundation (FAPESP), grants 2012/09596-
0 and 2013/00506-1 for the first and second authors, respectively. The authors also acknowledge support from
Laboratory EPIFISMA.

References

[1] A.M. Alonso, D. Pena, J. Romo, Introducing Model Uncertainty in Time Series Bootstrap, Statist. Sinica 14 (2004) 155174.
[2] T.G. Andersen, D. Dobrev, E. Schaumburg, Jump-robust volatility estimation using nearest neighbor truncation, J. Econometrics 169 (2012)
7593.
 C. Trucos, L.K. Hotta / Mathematics and Computers in Simulation 120 (2016) 91103 103

[3] T. Bollerslev, Generalized autoregressive conditional heteroskedasticity, J. Econometrics 31 (1986) 307327.

[4] M.A. Carnero, D. Pena, E. Ruiz, Estimating GARCH volatility in the presence of outliers, Econom. Lett. 114 (2012) 8690.
[5] C. Chatfield, Prediction intervals for time series forecasting, in: J.S. Armstrong (Ed.), Principles of Forecasting: a Handbook for Researchers
and Practitioners, Kluwer Academic Publishers, New York, 2001, pp. 475494.
[6] B. Chen, Y.R. Gel, N. Balakrishna, B. Abraham, Computationally efficient bootstrap prediction intervals for returns and volatilities in ARCH
and GARCH processes, J. Forecast. 30 (2011) 5171.
[7] J. Fan, Q. Yao, Nonlinear Time Series: Nonparametrics and Parametric Methods, Springer, New York, 2003.
[8] C. Fernandez, M.F.J. Steel, On Bayesian modeling of fat tails and skewness, J. Amer. Statist. Assoc. 93 (1998) 359371.
[9] D. Fresoli, E. Ruiz, The uncertainty of conditional returns, volatilities and correlations in DCC models., Comput. Statist. Data Anal. (2015)
in press, http://dx.doi.org/10.1016/j.csda.2015.03.017.
[10] L.R. Glosten, R. Jagannathan, D.E. Runkle, On the relation between the expected value and the volatility of the nominal excess return on
stocks, J. Finance. 48 (1993) 17791801.
[11] L.K. Hotta, R.S. Tsay, T.M.W. Bell, Outliers in GARCH processes, in: S.H. Holan (Ed.), Economic Time Series: Modeling and Seasonality,
Chapman and Hall/CRC Press, London, 2012, pp. 337358.
[12] L. Kilian, Accounting for lag order uncertainty in autoregressions: the endogenous lag order bootstrap algorithm, J. Time Series Anal. 19
(1998) 531548.
[13] Y.H. Lee, Sub-sample bootstrap prediction intervals in GARCH models, Eur. J. Soc. Sci. 22 (2011) 188192.
[14] J.A. Miguel, P. Olave, Bootstrapping forecast intervals in ARCH models, Test 8 (1999) 345364.
[15] D.B. Nelson, Conditional heteroskedasticity in asset returns: A new approach, Econometrica 59 (1991) 347370.
[16] L. Pascual, J. Romo, E. Ruiz, Bootstrap predictive inference for ARIMA processes, J. Time Series Anal. 25 (2004) 449465.
[17] L. Pascual, J. Romo, E. Ruiz, Bootstrap prediction intervals for power-transformed time series, Int. J. Forecast. 21 (2005) 219235.
[18] L. Pascual, J. Romo, E. Ruiz, Bootstrap prediction for returns and volatilities in GARCH models, Comput. Statist. Data Anal. 50 (2006)
22932312.
[19] J.J. Reeves, Bootstrap prediction intervals for ARCH models, Int. J. Forecast. 21 (2005) 237248.
[20] R.A. Stine, Estimating properties of autoregressive forecasts, J. Amer. Statist. Assoc. 82 (1987) 10721078.
[21] L.A. Thombs, W.R. Schucany, Bootstrap prediction intervals for autoregression, J. Amer. Statist. Assoc. 85 (1990) 486492.
[22] C. Trucos, L.K. Hotta, E. Ruiz, Robust bootstrap forecast densities for GARCH models: returns, volatilities and Value-at-Risk, 2015.
Manuscript.