Bootstrap prediction in univariate volatility models

with leverage effect

Carlos César Trucíos-Maza, Luiz K. Hotta∗
Department of Statistics, University of Campinas, IMECC-UNICAMP, Rua Sérgio
Buarque de Holanda 651, Cidade Universitária, Barão Geraldo, CEP 13083-859,
Campinas, SP, Brazil.

Abstract
The EGARCH and GJR-GARCH models are widely used in modeling volatil-
ity when a leverage effect is present in the data. Traditional methods of con-
structing 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 (PPR)
algorithm, developed to obtain prediction intervals for GARCH models, is
adapted to obtain prediction intervals of returns and volatility in EGARCH
and GJR-GARCH models. These adjustments 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 showed good performance
in Monte Carlo experiments. However, the performance, especially in volatil-
ity prediction, can be very poor in the presence an additive outlier near the
forecasting origin. The algorithms are applied to the daily return’s series of
the São Paulo Stock Market Index.
Keywords: Interval prediction, Volatility interval prediction, interval
prediction and outlier, interval prediction in EGARCH model, interval
prediction in GJR-GARCH model, skew distribution.

∗
Corresponding author. Tel.: +55 19 35216081
Email addresses: ctrucios@gmail.com (Carlos César Trucíos-Maza),
hotta@ime.unicamp.br (Luiz K. Hotta)

January 15, 2013

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, sales forecasting, etc. 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 important [13]. Nevertheless, authors of
textbooks on time series analysis and forecasting generally devote little at-
tention to prediction intervals and give little guidance on how to calculate
them [2, 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 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 us-
ing the bootstrap procedure, which does not require any assumption on the
distribution of the innovations [4]. In recent decades, several works have pro-
posed bootstrap procedures to construct intervals for time series prediction.
In a seminal work in this area, [16] constructed intervals for autoregressive
models using bootstrap replicates and fixing the latest p observations, where
p is the autoregressive order. In the financial time series field, we can mention
the work of [10] proposing a bootstrap procedure to obtain intervals for fore-
casting in ARCH processes; [15], who obtained prediction intervals for ARCH
models; [14], who extended the procedure presented by [12] to ARIMA mod-
els for predicting volatility and return densities in GARCH processes; [3],
who proposed new methods for prediction intervals of return and volatil-
ity in ARCH and GARCH models; [9], who used bootstrap sub sampling
for interval prediction in GARCH models; and [8], who developed bootstrap
interval prediction for ARFIMA 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 [14] (PRR) for GARCH models. This method in-
corporates the uncertainty of the estimation in the forecasting interval, since

2
the parameters are estimated at each bootstrap replication. The method
also does not depend on the (conditional) innovation distribution. This pa-
per 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 model. Section 4 presents the
results obtained by simulation; and Section 5 presents an application of the
proposed procedures to the series of daily returns of the São Paulo Stock
Market Index (Ibovespa). Section 6 concludes.

2. The EGARCH and GJR-GARCH models

GARCH models [1] 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
[11] and the GJR-GARCH models [6].

Definition 2.1 (Univariate EGARCH model). An EGARCH (p,q) process,

{rt }, is defined as:

rt = σt εt
q p
X X (2.1)
log(σt2 ) =ω+ g(εt−i ) + 2
βj log(σt−j ),
i=1 j=1

where

g(εt−i ) = αi εt−i + γi (|εt−i | − E(|εt−i |)), (2.2)

and ω, αi , βj , γi , are real numbers, and εt ∼ IID(0, 1) (independent and
identically distributed random variables with zero mean and unit variance).
The parameters also must satisfy the usual stationary restrictions. An ad-
vantage of this model is that there is no restriction imposed to ensure that
the variance is positive.

3
Definition 2.2 (GJR-GARCH model). An GJR-GARCH(p,q) process, {rt },
is defined as:

rt = σt εt
q p q
X X X (2.3)
σt2 =ω+ 2
αi rt−i + 2
βj σt−j + 2
γi rt−i I(rt−i < 0),
i=1 j=1 i=1

where, I(.) is the indicator function, εt ∼ IID(0, 1), ω > 0 and αi , βi , γi

are non-negative real numbers satisfying similar conditions as in the GARCH
model.

3. Bootstrap prediction intervals

This section presents an adaptation of the PRR algorithm proposed by
[14] for the GARCH process for application in EGARCH and GJR-GARCH
models. These modifications keep the original idea of PRR, i.e., incorpo-
rating 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) processes in a straightforward way.

• Step 1: Estimate the model with the observed data. Denote the esti-
mates as: θ̂ = (ω̂, α̂, β̂, γ̂), and calculate the centered residuals. Denote
by F̂T the empirical distribution of the centered residuals.

• Step 2: Generate a bootstrap series rt∗ , t = 1, ..., T using the following

recursion:

σ̂t∗2 = ω̂ + α̂rt−1
∗2 ∗2
+ β̂ σ̂t−1 ∗2
+ γ̂rt−1 ∗
I(rt−1 < 0))
∗ ∗ ∗
(3.4)
rt = εt σ̂t ,

where ε∗t ∼ i.i.d F̂T . Adjust the GJR-GARCH (1,1) model to the boot-
strap sequence RT∗ to obtain the bootstrap estimates θˆ∗ = (ω̂ ∗ , α̂∗ , β̂ ∗ , γ̂ ∗ ).

4
 • Step 3: Calculate forecasts of returns and volatilities h steps ahead,
h = 1, 2, ... using the following recursion:
σ̂T∗2 (h) = ω̂ ∗ + α̂∗ rT∗2 (h − 1) + β̂ ∗ σ̂T∗2 (h − 1) + γ̂ ∗ rT∗2 (h − 1)I(rT∗ (h − 1) < 0)
rT∗ (h) = ε∗T (h)σ̂T∗ (h),
(3.5)
where ε∗T (h) ∼ i.i.d F̂T , rT∗ = rT and
T −2
ω∗ ω∗ X
σT∗2 (0) = ∗ ∗
− (α + γ P ) β ∗i
1 − α∗ − β ∗ − γ ∗ P 1 − α∗ − β ∗ − γ ∗ P i=0
T −2
X T −2
X
+ α∗ β ∗i rT2 −i−1 + γ ∗ β ∗i rT2 −i−1 I(rT −i−1 < 0),
i=0 i=0
(3.6)
where P = P rob(rt < 0).
• Step 4: Repeat steps 2 and 3, B times to obtain B bootstrap replicates
∗(1) ∗(B)
(rT (h), ..., rT (h)) and
∗(1) ∗(B)
(σT (h), ..., σT (h)). Then a 100(1−γ)% prediction interval for rT∗ (h)
is given by:

[Q∗r,B ( γ2 ), Q∗r,B (1 − γ2 )],

∗b (a)≤a)
#(rT
where: Q∗r,B (a) = G∗−1 ∗
r,B (a), and Gr,B (a) = B
. Similarly, obtain
a prediction interval for the volatility.
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∗ (h)
 ∗  
 rT (h − 1) 
log(σ̂T∗2 (h)) ωˆ∗ + αˆ∗ T∗ γˆ∗ ˆ∗ ∗2
= +  σ ∗ (h − 1)  − M + β log(σT (h − 1))
 
σT (h) T
rT∗ (h) = ∗ ∗
εT (h)σ̂T (h),
(3.7)

5
 where ε∗T (h) ∼ i.i.d F̂T , M = E(|εT |), rT∗ = rT and

T −2
ω∗ X
∗j rT −1−j
log(σ̂T∗2 (0)) = + α ∗
β
1 − β∗ i=0
σT∗ −1−j
T −2
 
T −2
(3.8)
r 
∗j  T −1−j 
X X
+γ β  ∗  − γE(|εT |) β ∗j
 σT −1−j 
i=0 i=0

3.3. Algorith for DCC GARCH model

Consider rT a sequence two dimensional of T observations generated by
the process DCC(1,1)-GARCH(1,1). The algorithm is described for process
DCC(1,1)-GARCH(1,1) with K = 2 dimensions, but it is easily generalized
to a DCC(m,n)-GARCH(p,q) process with more than K = 2 dimensions in
a straightforward way.

• Step 1: Obtain estimates of the process parameters DCC-GARCH(1,1)

θ = (ω1 , ω2 , α1 , α2 , β1 , β2 , a, b) give by θ̂ = (ωˆ1 , ωˆ2 , αˆ1 , αˆ2 , βˆ1 , βˆ2 , â, b̂).
Calculate the vectors of residues centered ˆt − ¯ˆ, where ˆt = Dt−1 rt and
Dt is a diagonal matrix with components Dkkt = σ̂kt ;
−1/2
• Step 2: Obtain the residues standardized εt = ¯t R̂t and denote by
F̂T the empirical distribution of the residues standardized;
1/2
• Step 3: Obtain the residues bootstrap ∗t = ε∗t R̂t∗ t = 1, ..., T where
,∗−1/2 ∗ ,∗−1/2
∗
R̂t = Qt Qt Qt , Qt = Q̄ (1 − â − b̂) + â(t(∗t−1 )∗t−1 ) + b̂Q∗t−1 ,
∗ ∗
,∗1/2
Q̄∗ = Corr(), Q∗1 = R1∗ = Q̄∗ , ∗1 = ε∗1 and Qt is a diagonal matrix
,∗1/2 p ∗
with components Qkkt = qkkt ;

• Step 4: Generate a bootstrap series rT∗ = (r1T

∗ ∗
, r2T ∗
), where rkT is calcu-
lated using the following recursion:
∗2
σkt ∗2
= ωˆk + αˆk rk(t−1) + βˆk σk(t−1)
∗2
(3.9)
rkt = ∗kt σkt
∗
,

• Step 5: Compute Q∗T (0), RT∗ (0) and HT∗ (0). Ht∗ = σt∗ Rt∗ σt∗ where Rt∗ =
,∗1/2 ,∗1/2
Qt Q∗t Qt , Q∗t = Q̄∗ + aˆ∗ (t(et−1 )et−1 ) + bˆ∗ Q∗t−1 , et = rt Dt∗−1 , Q̄∗ =

6
 Corr(ˆ), R1∗ = R̂1∗ , Q∗1 = Q̂∗1 and σt∗ = c(σ1t
∗ ∗
, σ2t ∗
) and σkt is calculated
using the following recursion:

∗2 ωˆk∗
σkT =
1 − αˆ∗ − βˆ∗
k k
T −2 (3.10)
ωˆk∗
βˆk∗j (rk(T −j−1) −
X
+ αˆk∗ );
j=0 1 − αˆ∗ − βˆ∗
k k

• Step 6: Calculate forecasts of returns,volatilities and covariance h steps

ahead, h = 1, 2, ... using the following recursion: rT∗ (h) = ∗T (h)DT∗ (h)
HT∗ (h) = σT∗ (h)RT∗ σT∗ (h), where σT∗ (h) = c(σ1T
∗ ∗
(h), σ2T (h)) and is cal-
culated using the following recursion:
∗
σkT (h) = ωˆk∗ + αˆk∗ rkT
∗
(h − 1) + βˆk∗ σkT
∗
(h − 1),
∗ ∗
(3.11)
rkT (h) = T (h)σkT (h),

RT∗ (h) = QT (h)Q∗T (h)QT (h) Q∗T (h) = Q̄∗ (1−aˆ∗ −bˆ8 )+aˆ∗ (t(∗T (h))∗T (h))+
,∗1/2 ,∗1/2

∗1/2
bˆ∗ Q∗T (h − 1) ∗T (h) = ε∗T (h)RT (h) where ε∗T (h) ∼ i.i.d F̂T ∗T (0) =
∗−1/2
r T DT

• Step 7: Repeat steps 4 - 6, B times and there by obtain B boot-

∗(1) ∗(B) ∗(l) ∗(l) ∗(1)
strap replicates (rT (h), ..., rT (h)), where yT (h) is rT (h), σT (l)
∗(1)
or CovT (l) respectively. The limits of prediction for rT (h), σT (h)
and CovT (h) are defined as the quantiles of bootstrap rT∗ (h), σT∗ (h) e
CovT∗ (h).

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

EGARCH(1,1):

log(σt2 ) = −0.3474 − 0.1420εt + 0.2195(|εt−1 | − E(|εt−1 |)) + 0.9496log(σt−1

2
)
(4.12)

7
GJR - GARCH(1,1):
2
σt2 = 0.01 + 0.05ε2t−1 + 0.15ε2t−1 I(εt−1 < 0) + 0.83σt−1 , (4.13)

where εt are innovations. The skew-normal distribution is denoted by SN(λ),

where (λ) is the skewness parameter. In the symmetric distribution, λ = 0.
The skew-t distribution is denoted by ST(λ, ν), where (λ) is the skewness
parameter and ν is the degrees of freedom. The distribution is parameter-
ized as in Fernández and Steel [5], such that, when λ = 1, the distribution
is symmetric. In the EGARCH(1,1) model we consider innovations with
SN(0.7) and ST(0.8, 6) distributions, while for the GJR-GARCH(1,1) model
we consider innovations with SN(0.7) and ST (0.8, 7) distributions. For each
model, we ran 1, 000 replications, each involving three steps. For the k-th
replication, the steps are:

1. Step 1: Generate a series with size T = 1000. That is generate {rt , t =

1, · · · , 1000} and {σt , t = 1, · · · , 1000}.
2. Step 2: Run steps 1 to 4 given in Section 3.2 for the EGARCH(1,1)
model, and in Section 3.1 for the GJR-GARCH(1,1) model to find the
h = 1, 3, 5 steps-ahead 95% prediction intervals for the returns and
volatilities.
3. Step 3: Generate 1, 000 sets of future values {rTi +j , j = 1, · · · , 5, i =
1, · · · , 1000} and {σTi +j , j = 1, · · · , 5, i = 1, · · · , 1000} considering
that the previous observations are the same as given in Step 1. For
h = 1, 3, 5, denote by pr,h,k , pr,h,b,k and pr,h,a,k the proportions of the
values {rTi +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 pσ,h,k , pσ,h,b,k
and pσ,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.

4.1. Prediction interval for EGARCH 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, left and right wrong

8
coverages and the mean of the interval’s length for return and volatility
prediction intervals.
In both cases, the bootstrap procedure performed well, obtaining coverage
proportions near to the nominal values. The proportions of right and the left
error coverages are close for the return, while there is a little more failure
in the left error for the volatility. 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 EGARCH 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, skew Normal or skew-t.
The table also presents the results for the one-step-ahead intervals given
by the estimator asymptotic distribution [1]. The analysis of the results sug-
gests that for 95% confidence intervals the bootstrap procedure presents bet-
ter results for all innovation distributions (normal, skew-normal and skew-t).
For the 99% confidence interval, no procedure is better than the other, but we
must take into account that the analysis is based only in 1, 000 replications.
For the volatility, for any innovation distribution, the coverage standard de-
viation decreases when the prediction horizon increases.
The effect of additive outliers, as defined in [7], is also considered. For a
sample size of 1, 000 the outlier is added at the 500-th or the 999-th position.
Let y1 , · · · , y1000 be a series generated without outliers. When we have an
additive outlier of size δ in the 999-th observation, the observed series is
y1 , · · · , y998 , y999 + δ ∗ sign(y999 ), y1000 . In the simulation we considered δ
equal to three times the sample standard deviation of {y1 , · · · , y1000 }. The
distribution of the innovation is the Student-t with 6 degrees of freedom.
The simulation tests the effect of the outlier when the outlier is ignored. 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 nil in
the proportion of the coverage of the return prediction interval, while there
is little influence for the coverage of the volatility intervals for the one-step-
ahead prediction, and almost none for longer forecast. 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. 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 proportion of the coverage of the volatility is
much smaller than the nominal values. This happens because the outlier has
a strong bias effect on the point volatility forecast.

9
Table 1: Estimated coverage of the 95% (top panel) and 99% (bottom panel) return
bootstrap prediction intervals. EGARCH(1,1) model.
Hori- Innov. Average Std. Average Average Average
zon distr. coverage errors below int. above int. length
Normal 0.9470 0.0156 0.0263 0.0267 0.1300
Asymptotic 0.9437 0.0128 0.0230 0.0333 0.1303
SN 0.9480 0.0152 0.0255 0.0265 0.1306
1 Asymptotic 0.9527 0.0117 0.0382 0.0092 0.1324
ST 0.9484 0.0159 0.0252 0.0264 0.1372
Asymptotic 0.9477 0.0131 0.0361 0.0162 0.1350
Normal 0.9473 0.0153 0.0261 0.0267 0.1315
2 SN 0.9480 0.0154 0.0255 0.0266 0.1318
ST 0.9486 0.0155 0.0252 0.0262 0.1384
Normal 0.9469 0.0156 0.0263 0.0268 0.1322
SN 0.9484 0.0146 0.0255 0.0261 0.1332
3
ST 0.9489 0.0155 0.0249 0.0262 0.1400
Normal 0.9476 0.0150 0.0259 0.0266 0.1333
SN 0.9486 0.0144 0.0252 0.0262 0.1344
4
ST 0.9485 0.0160 0.0253 0.0262 0.1404
Normal 0.9468 0.0148 0.0264 0.0268 0.1339
5 SN 0.9480 0.0145 0.0256 0.0264 0.1353
ST 0.9484 0.0161 0.0253 0.0263 0.1416
Normal 0.9881 0.0066 0.0055 0.0064 0.1716
Asymptotic 0.9884 0.0034 0.0050 0.0066 0.1712
SN 0.9881 0.0065 0.0055 0.0065 0.1715
1 Asymptotic 0.9876 0.0049 0.0118 0.0006 0.1740
ST 0.9881 0.0068 0.0054 0.0066 0.2094
Asymptotic 0.9807 0.0066 0.0150 0.0043 0.1774
Normal 0.9882 0.0064 0.0055 0.0064 0.1768
2 SN 0.9884 0.0064 0.0055 0.0062 0.1775
ST 0.9888 0.0062 0.0051 0.0061 0.2165
Normal 0.9882 0.0065 0.0055 0.0063 0.1799
3 SN 0.9888 0.0060 0.0052 0.0061 0.1830
ST 0.9888 0.0062 0.0050 0.0063 0.2215
Normal 0.9885 0.0062 0.0053 0.0062 0.1844
4 SN 0.9886 0.0061 0.0051 0.0063 0.1881
ST 0.9888 0.0063 0.0051 0.0061 0.2254
Normal 0.9881 0.0063
10 0.0055 0.0064 0.1865
5 SN 0.9884 0.0060 0.0054 0.0062 0.1921
ST 0.9886 0.0067 0.0053 0.0062 0.2278
Table 2: Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility
bootstrap prediction intervals. EGARCH(1,1) model.
Hori- Innov. Average Std. Average Average Average
zon distr. coverage errors below int. above int. length
Normal 0.9420 0.2339 0.0180 0.0400 0.0060
1 SN 0.9470 0.2241 0.0230 0.0300 0.0058
ST 0.9400 0.2376 0.0270 0.0330 0.0077
Normal 0.9526 0.0824 0.0170 0.0304 0.0153
2 SN 0.9500 0.0853 0.0216 0.0284 0.0172
ST 0.9483 0.1031 0.0243 0.0273 0.0191
Normal 0.9529 0.0606 0.0169 0.0302 0.0196
3 SN 0.9512 0.0625 0.0202 0.0286 0.0221
ST 0.9481 0.0798 0.0237 0.0282 0.0250
Normal 0.9522 0.0491 0.0173 0.0305 0.0230
SN 0.9517 0.0498 0.0193 0.0290 0.0260
4
ST 0.9483 0.0664 0.0233 0.0283 0.0294
Normal 0.9509 0.0447 0.0186 0.0305 0.0257
SN 0.9512 0.0433 0.0198 0.0290 0.0292
5
ST 0.9479 0.0586 0.0234 0.0287 0.0330
Normal 0.9770 0.1500 0.0080 0.0150 0.0081
1 SN 0.9870 0.1133 0.0040 0.0090 0.0078
ST 0.9780 0.1468 0.0110 0.0110 0.0105
Normal 0.9870 0.0487 0.0052 0.0078 0.0218
SN 0.9889 0.0320 0.0037 0.0073 0.0254
2
ST 0.9855 0.0635 0.0077 0.0068 0.0330
Normal 0.9882 0.0332 0.0043 0.0075 0.0282
SN 0.9896 0.0211 0.0032 0.0072 0.0326
3
ST 0.9869 0.0464 0.0063 0.0068 0.0430
Normal 0.9885 0.0242 0.0037 0.0078 0.0330
SN 0.9893 0.0183 0.0033 0.0074 0.0379
4
ST 0.9882 0.0334 0.0050 0.0068 0.0503
Normal 0.9887 0.0194 0.0036 0.0077 0.0370
5 SN 0.9892 0.0168 0.0036 0.0073 0.0427
ST 0.9885 0.0269 0.0046 0.0069 0.0561

11
Table 3: Estimated coverage of the 95% (top panel) and 99% (bottom panel) return
bootstrap prediction intervals. EGARCH(1,1) model with outlier at the k-th observation.
Hori- Outlier Average Std. Average Average Average
zon position coverage errors below int. above int. length
500 0.9471 0.0169 0.0261 0.0268 0.1337
1 999 0.9763 0.0209 0.0120 0.0118 0.1897
500 0.9483 0.0168 0.0253 0.0264 0.1365
3 999 0.9743 0.0202 0.0126 0.0131 0.1857
500 0.9479 0.0168 0.0254 0.0266 0.1379
5 999 0.9723 0.0196 0.0136 0.0141 0.1810
500 0.9885 0.0065 0.0054 0.0062 0.2076
1 999 0.9954 0.0056 0.0022 0.0023 0.2968
500 0.9889 0.0065 0.0049 0.0063 0.2179
3 999 0.9951 0.0053 0.0022 0.0026 0.2982
500 0.9888 0.0062 0.0051 0.0061 0.2217
5 999 0.9948 0.0053 0.0025 0.0027 0.2941

Table 4: Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility
bootstrap prediction intervals. GJR-GARCH(1,1) model with outlier at the k-th observa-
tion.
Hori- Outlier Average Std. Average Average Average
zon position coverage errors below int. above int. length
500 0.9390 0.2395 0.0290 0.0320 0.0079
1 999 0.4100 0.4921 0.5890 0.0010 0.0182
500 0.9484 0.0737 0.0220 0.0295 0.0219
3 999 0.5016 0.4063 0.4893 0.0091 0.0333
500 0.9474 0.0549 0.0227 0.0300 0.0287
5 999 0.5913 0.3540 0.3977 0.0110 0.0403
500 0.9800 0.1401 0.0090 0.0110 0.0107
1 999 0.5150 0.5000 0.4840 0.0010 0.0244
500 0.9881 0.0355 0.0049 0.0069 0.0362
3 999 0.5933 0.4205 0.4043 0.0024 0.0551
500 0.9886 0.0225 0.0040 0.0073 0.0468
5 999 0.6851 0.3498 0.3121 0.0028 0.0668

12
4.2. Prediction interval for GJR-GARCH model
Here the simulation done in Section 4.1 is repeated for the GJR-GARCH
model. The results without outliers are presented in Tables 5 and 6, and with
outliers in Tables 7 and 8. The conclusions are the same as before. 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 prediction 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.

5. Application
Figure 1 presents the daily returns of the Ibovespa series from 4th January
2000 to 30th April 2012 (T = 3054 observations). The sample mean and
standard deviation are equal to 0.00042 and 0.01930, respectively.
0.10
0.05
ibovespa returns

0.00
−0.05
−0.10

2000 2002 2004 2006 2008 2010 2012

Year

Figure 1: Daily Ibovespa returns series from 4th January 2000 to 30th April 2012

13
Table 5: Estimated coverage of the 95% (top panel) and 99% (bottom panel) return
bootstrap prediction intervals. GJR-GARCH(1,1) model.
Hori- Innov. Average Std. Average Average Average
zon distr. coverage errors below int. above int. length
Normal 0.9475 0.0153 0.0255 0.0270 1.7720
Asymptotic 0.9477 0.0134 0.0261 0.0262 1.7180
SN 0.9479 0.0145 0.0255 0.0266 1.8720
1 Asymptotic 0.9521 0.0114 0.0390 0.0089 1.9048
ST 0.9476 0.0149 0.0255 0.0269 1.8285
Asymptotic 0.9412 0.0120 0.0427 0.0160 1.7917
Normal 0.9473 0.0161 0.0258 0.0269 1.7774
2 SN 0.9474 0.0145 0.0257 0.0269 1.8857
ST 0.9478 0.0149 0.0256 0.0266 1.8433
Normal 0.9471 0.0158 0.0260 0.0270 1.7844
3 SN 0.9480 0.0145 0.0255 0.0265 1.9035
ST 0.9476 0.0151 0.0258 0.0266 1.8556
Normal 0.9476 0.0155 0.0260 0.0264 1.7968
4 SN 0.9476 0.0146 0.0256 0.0268 1.9194
ST 0.9476 0.0152 0.0255 0.0269 1.8723
Normal 0.9477 0.0155 0.0255 0.0268 1.8059
5 SN 0.9485 0.0145 0.0250 0.0265 1.9354
ST 0.9479 0.0150 0.0255 0.0266 1.8885
Normal 0.9879 0.0069 0.0054 0.0067 2.3370
Asymptotic 0.9891 0.0049 0.0054 0.0055 2.2578
SN 0.9880 0.0066 0.0055 0.0065 2.4561
1 Asymptotic 0.9871 0.0049 0.0123 0.0006 2.5033
ST 0.9879 0.0065 0.0054 0.0067 2.7363
Asymptotic 0.9787 0.0067 0.0163 0.0051 2.3547
Normal 0.9882 0.0068 0.0055 0.0063 2.3813
2 SN 0.9880 0.0062 0.0055 0.0065 2.5287
ST 0.9881 0.0063 0.0054 0.0064 2.8114
Normal 0.9882 0.0065 0.0055 0.0063 2.4288
3 SN 0.9886 0.0061 0.0052 0.0062 2.6153
ST 0.9881 0.0065 0.0055 0.0064 2.8734
Normal 0.9884 0.0063 0.0053 0.0063 2.4740
4 SN 0.9887 0.0060 0.0051 0.0063 2.6887
ST 0.9879 0.0065 0.0055 0.0066 2.9236
Normal 0.9883 0.0065
14 0.0055 0.0063 2.5000
5 SN 0.9884 0.0063 0.0053 0.0063 2.7497
ST 0.9880 0.0066 0.0054 0.0066 2.9951
Table 6: Estimated coverage of the 95% (top panel) and 99% (bottom panel) volatility
bootstrap prediction intervals. GJR-GARCH(1,1) model.
Hori- Innov. Average Std. Average Average Average
zon distr. coverage errors below int. above int. length
Normal 0.9310 0.2536 0.0310 0.0380 0.0817
1 SN 0.9450 0.2281 0.0230 0.0320 0.0812
ST 0.9450 0.2281 0.0170 0.0380 0.1007
Normal 0.9426 0.1147 0.0289 0.0285 0.1958
2 SN 0.9515 0.0974 0.0213 0.0272 0.2398
ST 0.9536 0.0978 0.0170 0.0294 0.2440
Normal 0.9419 0.0911 0.0282 0.0299 0.2484
3 SN 0.9505 0.0762 0.0216 0.0279 0.3093
ST 0.9523 0.0774 0.0183 0.0294 0.3257
Normal 0.9401 0.0777 0.0295 0.0303 0.2883
4 SN 0.9492 0.0651 0.0225 0.0283 0.3593
ST 0.9497 0.0704 0.0203 0.0301 0.3835
Normal 0.9385 0.0708 0.0309 0.0306 0.3191
5 SN 0.9482 0.0583 0.0236 0.0282 0.4044
ST 0.9480 0.0653 0.0217 0.0302 0.4299
Normal 0.9790 0.1435 0.0060 0.0150 0.1092
1 SN 0.9890 0.1044 0.0020 0.0090 0.1090
ST 0.9830 0.1293 0.0040 0.0130 0.1348
Normal 0.9866 0.0529 0.0062 0.0072 0.3059
2 SN 0.9892 0.0375 0.0042 0.0067 0.3848
ST 0.9895 0.0410 0.0033 0.0072 0.4521
Normal 0.9856 0.0435 0.0070 0.0073 0.3770
3 SN 0.9889 0.0295 0.0041 0.0070 0.4747
ST 0.9890 0.0355 0.0034 0.0076 0.5748
Normal 0.9852 0.0377 0.0072 0.0075 0.4342
4 SN 0.9888 0.0253 0.0042 0.0069 0.5555
ST 0.9884 0.0322 0.0041 0.0075 0.6630
Normal 0.9844 0.0349 0.0080 0.0076 0.4831
5 SN 0.9887 0.0238 0.0045 0.0068 0.6292
ST 0.9880 0.0282 0.0042 0.0078 0.7348

15
Table 7: 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
k-th observation.
Hori- Outlier Average Std. Average Average Average
zon position coverage errors below int. above int. length
500 0.9387 0.0956 0.0252 0.0261 1.7328
1 999 0.9859 0.0145 0.0069 0.0071 2.7278
500 0.9390 0.0956 0.0249 0.0261 1.7580
3 999 0.9837 0.0151 0.0080 0.0083 2.6649
500 0.9388 0.0957 0.0252 0.0260 1.7738
5 999 0.9810 0.0159 0.0094 0.0096 2.5990
500 0.9789 0.0986 0.0050 0.0061 2.6252
1 999 0.9975 0.0035 0.0011 0.0013 4.1331
500 0.9792 0.0987 0.0048 0.0059 2.7410
3 999 0.9971 0.0037 0.0013 0.0015 4.1738
500 0.9791 0.0986 0.0051 0.0058 2.8078
5 999 0.9965 0.0043 0.0017 0.0018 4.1580

Table 8: 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 k-th
observation.
Hori- Outlier Average Std. Average Average Average
zon position coverage errors below int. above int. length
500 0.9150 0.2790 0.0440 0.0310 0.1000
1 999 0.2610 0.4394 0.7390 0.0000 0.2460
500 0.9207 0.1496 0.0417 0.0276 0.2776
3 999 0.3345 0.3679 0.6611 0.0044 0.4906
500 0.9223 0.1290 0.0398 0.0279 0.3631
5 999 0.4207 0.3472 0.5732 0.0061 0.6115
500 0.9730 0.1622 0.0070 0.0100 0.1340
1 999 0.4120 0.4924 0.5880 0.0000 0.3238
500 0.9730 0.1148 0.0103 0.0067 0.4791
3 999 0.4664 0.4235 0.5324 0.0012 0.8345
500 0.9732 0.1064 0.0100 0.0068 0.6165
5 999 0.5478 0.3838 0.4506 0.0015 1.0388

16
 Table 9: Descriptive statistics of the daily Ibovespa return series.
Mean D.P Min. 25% Med. 75% Max. Asym. Kurt.
0.0004 0.0193 -0.1210 -0.0102 0.0011 0.0116 0.1368 -0.1271 6.779

To evaluate the prediction intervals of returns, we count the proportion

of true values inside the intervals. For empirical data, the volatility is not
observed. In this case, the estimated volatility was considered as the true
value, as in Pascual et al. [14].
The GARCH, EGARCH and GJR-GARCH models with several orders
were fitted to the series using normal, skew-normal and skew-t distributions
for the innovation term. The best two adjustments, according to the AIC
and BIC, were the EGARCH(1,1) and GJR-GARCH(1,1) models, both with
skew-t distribution. Because the analysis of the standardized residuals of
both models passed the usual diagnostic analysis, in the following, only these
models will be entertained.
At first, the models were fitted to the first T − 255 observations, h =
1, 2, 3, 4, 5 steps-ahead bootstrap prediction intervals were constructed for
the return and the (T − 254)-th observation was included. The models were
estimated, and then the were intervals constructed and checked as to whether
they included the true value. This procedure was repeated 250 times. Ta-
bles 10 and 11 present the estimated coverage of the 95% and 99% return
and volatility bootstrap prediction intervals. 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 larger than the nominal values,
especially for short horizons. For instance, all the one-step-ahead predic-
tion intervals covered the right value, indicating unnecessarily wide intervals.
When the intervals did not cover the true volatility values, the volatility was
always larger than the interval’s upper bound.

6. Conclusion
This paper adapts the PRR algorithm proposed by [14] to GARCH mod-
els, 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 procedures also have a good

17
Table 10: Empirical coverage of the return and volatility prediction bootstrap prediction
intervals for the Ibovespa return series. Nominal coverage 95%.
Hori- Volatility Average Std. Average Average
zon model coverage errors below int. above int.

Return
EGARCH 0.9640 0.1867 0.0200 0.0160
1 GJR 0.9640 0.1867 0.0200 0.0160
EGARCH 0.9560 0.2055 0.0200 0.0240
3 GJR 0.9680 0.1764 0.0120 0.0200
EGARCH 0.9680 0.1764 0.0200 0.0120
5 GJR 0.9680 0.1764 0.0160 0.0160
Volatility
EGARCH 1.0000 0.0000 0.0000 0.0000
1 GJR 1.0000 0.0000 0.0000 0.0000
EGARCH 0.9760 0.1534 0.0000 0.0240
3 GJR 0.9800 0.1403 0.0000 0.0200
EGARCH 0.9840 0.1257 0.0000 0.0160
5 GJR 0.9760 0.1534 0.0000 0.0240

18
Table 11: Empirical coverage of the return and volatility prediction bootstrap prediction
intervals for the Ibovespa return series. Nominal coverage 99%.
Hori- Volatility Average Std. Average Average
zon model coverage errors below int. above int.

Return
EGARCH 0.9880 0.1091 0.0080 0.0040
1 GJR 0.9920 0.0893 0.0080 0.0000
EGARCH 0.9880 0.1091 0.0080 0.0040
3 GJR 0.9920 0.0893 0.0080 0.0000
EGARCH 0.9840 0.1257 0.0080 0.0080
5 GJR 0.9880 0.1091 0.0080 0.0040
Volatility
EGARCH 1.0000 0.0000 0.0000 0.0000
1 GJR 1.0000 0.0000 0.0000 0.0000
EGARCH 0.9960 0.0632 0.0000 0.0040
3 GJR 0.9880 0.1091 0.0000 0.0120
EGARCH 0.9840 0.1257 0.0000 0.0160
5 GJR 0.9840 0.1257 0.0000 0.0160

19
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.

References
[1] T. Bollerslev, Generalized autoregressive conditional heteroskedasticity,
J Econom. 31 (1986) 307–327.

[2] 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.
475–494.

[3] B. Chen, Y.R. Gel, N. Balakrishna, B. Abraham, Computationally effi-

cient bootstrap prediction intervals for returns and volatilities in ARCH
and GARCH processes, J Forecast. 30 (2011) 51–71.

[4] J. Fan, Q. Yao, Nonlinear Time Series: Nonparametrics and Parametric

Methods, Springer, New York, 2003.

[5] C. Fernández, M.F.J. Steel, On Bayesian modeling of fat tails and skew-
ness, J Am Stat Assoc. 93 (1998) 359–371.

[6] 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) 1779–1801.

[7] L.K. Hotta, R.S. Tsay, Outliers in GARCH processes, in: T.M. W. Bell,
S.H. Holan (Eds.), Economic Time Series: Modeling and Seasonality,
Chapman and Hall/CRC Press, London, 2012, pp. 337–358.

[8] G.d.C. Lana, Intervalos de previsão em modelos ARFIMA utilizando

a metodologia bootstrap, Master dissertation, Universidade Federal de
Minas Gerais, 2012.

[9] Y.H. Lee, Sub-sample bootstrap prediction intervals in GARCH models,

Eur J Soc Sci. 22 (2011) 188–192.

[10] J.A. Miguel, P. Olave, Bootstrapping forecast intervals in ARCH models,

Test 8 (1999) 345–364.

20
[11] D.B. Nelson, Conditional heteroskedasticity in asset returns: A new
approach, Econometrica 59 (1991) 347–370.

[12] L. Pascual, J. Romo, E. Ruiz, Bootstrap predictive inference for ARIMA

processes, J Time Ser Anal. 25 (2004) 449–465.

[13] L. Pascual, J. Romo, E. Ruiz, Bootstrap prediction intervals for power-

transformed time series, Int J Forecast. 21 (2005) 219–235.

[14] L. Pascual, J. Romo, E. Ruiz, Bootstrap prediction for returns and

volatilities in GARCH models, Comput Stat Data Anal. 50 (2006) 2293–
2312.

[15] J.J. Reeves, Bootstrap prediction intervals for ARCH models, Int J Fore-
cast. 21 (2005) 237–248.

[16] L.A. Thombs, W.R. Schucany, Bootstrap prediction intervals for autore-
gression, J Am Stat Assoc. 85 (1990) 486–492.

21