# Forecasting Realized Volatility with Linear and Nonlinear Univariate Models

**ABSTRACT** In this paper we consider a nonlinear model based on neural networks as well as linear models to forecast the daily volatility of the S&P 500 and FTSE 100 futures. As a proxy for daily volatility, we consider a consistent and unbiased estimator of the integrated volatility that is computed from high frequency intra-day returns. We also consider a simple algorithm based on bagging (bootstrap aggregation) in order to specify the models analyzed.

**0**Bookmarks

**·**

**95**Views

- Citations (28)
- Cited In (0)

- [Show abstract] [Hide abstract]

**ABSTRACT:**Barron (1993) obtained a deterministic approximation rate (in L2-norm) of r-m for a class of single hid- den layer feedforward artificial neural networks (ANN) with r hidden units and sigmoid activation func- tions when the target function satisfies certain smoothness conditions. Hornik, Stinchcombe, White, and Auer (HSWA, 1994) extended Barron's result to a class of ANNs with possibly non-sigmoid activation approximating the target function and its derivatives simultaneously. Recently Makovoz (1996) obtained an improved degree of approximation rate r-o+/a for Barron's ANNs with sigmoid activation func- tion where d is the dimension of the domain of the target function.05/2002; - SourceAvailable from: Ole E. Barndorff-Nielsen[Show abstract] [Hide abstract]

**ABSTRACT:**This paper shows how to use realized kernels to carry out efficient feasible inference on the ex post variation of underlying equity prices in the presence of simple models of market frictions. The weights can be chosen to achieve the best possible rate of convergence and to have an asymptotic variance which equals that of the maximum likelihood estimator in the parametric version of this problem. Realized kernels can also be selected to (i) be analyzed using endogenously spaced data such as that in data bases on transactions, (ii) allow for market frictions which are endogenous, and (iii) allow for temporally dependent noise. The finite sample performance of our estimators is studied using simulation, while empirical work illustrates their use in practice. Copyright 2008 The Econometric Society.Econometrica 01/2008; 76(6):1481-1536. · 3.82 Impact Factor - SourceAvailable from: Fulvio Corsi[Show abstract] [Hide abstract]

**ABSTRACT:**The paper proposes an additive cascade model of volatility components defined over different time periods. This volatility cascade leads to a simple AR-type model in the realized volatility with the feature of considering different volatility components realized over different time horizons and thus termed Heterogeneous Autoregressive model of Realized Volatility (HAR-RV). In spite of the simplicity of its structure and the absence of true long-memory properties, simulation results show that the HAR-RV model successfully achieves the purpose of reproducing the main empirical features of financial returns (long memory, fat tails, and self-similarity) in a very tractable and parsimonious way. Moreover, empirical results show remarkably good forecasting performance. Copyright The Author 2009. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oupjournals.org, Oxford University Press.Journal of Financial Econometrics 01/2009; 7(2):174-196. · 1.17 Impact Factor

Page 1

doi: 10.1111/j.1467-6419.2010.00640.x

FORECASTING REALIZED VOLATILITY

WITH LINEAR AND NONLINEAR

UNIVARIATE MODELS

Michael McAleer

Erasmus University Rotterdam and National Chung Hsing

University

Marcelo C. Medeiros

Pontifical Catholic University

Abstract. In this paper, we consider a nonlinear model based on neural networks

as well as linear models to forecast the daily volatility of the S&P 500 and

FTSE 100 futures. As a proxy for daily volatility, we consider a consistent

and unbiased estimator of the integrated volatility that is computed from high-

frequency intraday returns. We also consider a simple algorithm based on bagging

(bootstrap aggregation) in order to specify the models analysed in this paper.

Keywords. Bagging; Financial econometrics; Neural networks; Nonlinear models;

Realized volatility; Volatility forecasting

1. Introduction

Modelling and forecasting the conditional variance, or volatility, of financial time

series has been one of the major topics in financial econometrics. It is widely known

that the daily returns of financial assets, especially of stocks, are difficult, if not

impossible, to predict, although the volatility of the returns seems to be relatively

easier to forecast. Therefore, it is hardly surprising that financial econometrics and,

in particular, the modelling of financial volatility, has played such a central role in

modern pricing and risk management theories.

There is, however, an inherent problem in using models where the volatility

measure plays a central role. The conditional variance is latent, and hence is

not directly observable. It can be estimated, among other approaches, by the

(generalized) autoregressive conditional heteroskedasticity, or (G)ARCH, family

of models proposed by Engle (1982) and Bollerslev (1986), stochastic volatility

models (see, for example, Taylor, 1986) or exponentially weighted moving averages,

as advocated by the Riskmetrics methodology (see McAleer (2005) for a recent

exposition of a wide range of univariate and multivariate, conditional and stochastic,

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main

Street, Malden, MA 02148, USA.

Page 2

FORECASTING REALIZED VOLATILITY WITH LINEAR/NONLINEAR MODELS7

models of volatility, and Asai et al. (2006) for a review of the growing literature

on multivariate stochastic volatility models). However, as observed by Bollerslev

(1987), Malmsten and Ter¨ asvirta (2004) and Carnero et al. (2004), among others,

most of the latent volatility models fail to describe satisfactorily several stylized

facts that are observed in financial time series.

An empirical fact that standard latent volatility models fail to describe in an

adequate manner is the low, but slowly decreasing, autocorrelations in the squared

returns that are associated with high excess kurtosis of returns. Correctly describing

the dynamics of the returns is important in order to obtain accurate forecasts of

the future volatility which, in turn, is important in risk analysis and management.

In this sense, the assumption of Gaussian standardized returns has been refuted in

many studies, and heavy-tailed distributions have instead been used. See Jondeau

et al. (2007) for a nice discussion on the application of non-Gaussian distributions

in finance.

The search for an adequate framework for the estimation and prediction of

the conditional variance of financial assets returns has led to the analysis of high-

frequency intraday data. Merton (1980) noted that the variance over a fixed interval

can be estimated arbitrarily, although accurately, as the sum of squared realizations,

provided the data are available at a sufficiently high sampling frequency. More

recently, Andersen and Bollerslev (1998) showed that expost daily foreign exchange

volatility is best measured by aggregating 288 squared 5-minute returns. The 5-

minute frequency is a tradeoff between accuracy, which is theoretically optimized

using the highest possible frequency, and microstructure noise that can arise through

the bid–ask bounce, asynchronous trading, infrequent trading and price discreteness,

among other factors (see Madhavan, 2000; Biais et al., 2005, for very useful

surveys).

Ignoring the remaining measurement error, which can be problematic, the ex

post volatility essentially becomes ‘observable’. Andersen and Bollerslev (1998)

and Patton (2008) used this new volatility measure to evaluate the out-of-sample

forecasting performance of GARCH models. As volatility becomes ‘observable’,

it can be modelled directly, rather than being treated as a latent variable. Based

on the theoretical results of Barndorff-Nielsen and Shephard (2002), Andersen

et al. (2003) and Meddahi (2002), several recent studies have documented the

properties of realized volatilities constructed from high-frequency data. However,

microstructure effects introduce a severe bias on the daily volatility estimation.

Zhang et al. (2005), Bandi and Russell (2006), Hansen and Lunde (2006) and

Barndorff-Nielsen et al. (2008), among others, have discussed various solutions to

the inconsistency problem.

In this paper, we consider the forecasting of stock market volatility via nonlinear

models based on a neural network (NN) version of the heterogeneous autoregressive

model (HAR) of Corsi (2009). As in Hillebrand and Medeiros (2009) we evaluate

the benefits of bagging (bootstrap aggregation) in forecasting daily volatility as

well as the inclusion of past cumulated returns over different horizons as possible

predictors. As the number of predictors can get quite large, the application of

bagging is recommended as a device to improve forecasting performance.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 3

8MCALEER AND MEDEIROS

The remainder of the paper is organized as follows. In Section 2, we briefly

discuss the main concepts in construction realized volatility measures. In Section 3,

the models considered in this paper are presented, whereas in Section 4 we

describe the bagging methodology to specify the models and construct forecasts.

The empirical results are presented in Section 5. Section 6 concludes the

paper.

2. Realized Volatility

Suppose that, during day t, the logarithmic prices of a given asset follow a

continuous time diffusion process, as follows:

dp(t + τ) = μ(t + τ)dτ + σ(t + τ)dW(t + τ)

where p(t + τ) is the logarithmic price at time t + τ, μ(t + τ) is the drift

component, σ(t + τ) is the instantaneous volatility (or standard deviation) and

W(t + τ) is a standard Brownian motion.

Andersen et al. (2003) and Barndorff-Nielsen and Shephard (2002) showed

that daily returns, r(t) = p(t) − p(t − 1), are Gaussian conditionally on Ft≡

F{μ(t + τ − 1),σ(t + τ − 1)}τ=1

the sample paths of μ(t + τ − 1) and σ(t + τ − 1), 0 ≤ τ ≤ 1, such that

??1

The term IVt=?1

object of interest as a measure of the true daily volatility.

In practical applications, prices are observed at discrete and irregularly spaced

intervals and there are many ways to sample the data. Suppose that on a given

day t, we partition the interval [0, 1] and define the grid of observation times

{τ1,···,τn}, 0 = τ1< τ2··· < τn= 1. The length of the ith subinterval is given by

δi= τi− τi−1. The most widely used sampling scheme is calendar time sampling,

where the intervals are equidistant in calendar time, that is δi= 1/n. Let pt,i,i =

1,...,n, be the ith log price observation during day t, such that rt,i= pt,i− pt,i−1

is the ith intra-period return of day t. Realized variance is defined as

0 ≤ τ ≤ 1,t = 1,2,3,... (1)

τ=0, the σ-algebra (information set) generated by

?1

0σ(t + τ − 1)dτ is known as the integrated variance, which

is a measure of the day-t ex post volatility. The integrated variance is typically the

rt| Ft∼ N

0

μ(t + τ − 1)dτ,

0

σ(t + τ − 1)dτ

?

RVt=

n

?

i=2

r2

t,i

(2)

Realized volatility is the square-root of (2).

Under regularity conditions, including the assumption of uncorrelated intraday

returns, realized variance RV2

that RVt

→IVt. However, when returns are serially correlated, realized variance

is a biased and inconsistent estimator of integrated variance. Serial correlation

may be the result of market microstructure effects such as bid–ask bounce and

discreteness of prices (Campbell et al., 1997; Madhavan, 2000; Biais et al., 2005).

tis a consistent estimator of integrated variance, such

p

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 4

FORECASTING REALIZED VOLATILITY WITH LINEAR/NONLINEAR MODELS9

These effects prevent very fine sampling partitions. Realized volatility is therefore

not an error-free measure of volatility.

The search for asymptotically unbiased, consistent and efficient methods for

measuring realized volatility in the presence of microstructure noise has been one

of the most active research topics in financial econometrics over the last few years.

Although early references in the literature, such as Andersen etal. (2001), advocated

the simple selection of an arbitrary lower frequency (typically 5–15 minutes) to

balance accuracy and the dissipation of microstructure bias, a procedure that is

known as sparse sampling, recent articles have developed estimators that dominate

this procedure.

Recently, Barndorff-Nielsen et al. (2008), hereafter BHLS (2008), proposed the

flat-top kernel-based estimator

RV(BHLS)

t

= RVt+

H

?

h=1

k

?h − 1

H

?

(ˆ γh+ ˆ γ−h)(3)

where k(x) for x ∈ [0,1] is a non-stochastic weight function such that k(0) = 1

and k(1) = 0, RVtis defined as in (2) and

n

n − h

BHLS (2008) discussed different kernels and provided all the technical details.

ˆ γh=

n−h

?

j=1

rt,jrt,j+h

3. The Models

Let yt be the square-root of the logarithm of a consistent and unbiased estimator

for the integrated variance of day t, such as the estimator in (3), and call it the daily

‘realized volatility’.1Define daily accumulated logarithm returns over an h-period

interval as

rh,t=

h−1

?

i=0

rt−i

(4)

where rt is the daily return at day t. Furthermore, define the average log realized

volatility over h days as

yh,t=1

h

h−1

?

i=0

yt−i

(5)

3.1 The Linear Heterogeneous Autoregressive Model

The linear HAR model proposed by Corsi (2009) is defined as

?

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

yt= β0+

ιi∈I

βiyιi,t−1+ εt= β0+ β?xt−1+ εt

(6)

Page 5

10MCALEER AND MEDEIROS

where xt−1= (yι1,t−1,..., yιp,t−1)?, I = (ι1,ι2,...,ιp) is a set of p indices with

0 < ι1< ι2< ··· < ιp< ∞ and i = 1,..., p. Throughout the paper, εt is a

zero-mean and uncorrelated process with finite, but not necessarily constant

variance (Corsi et al., 2008). Corsi (2009) advocated the use of I = (1,5,22).

His specification builds on the HARCH model proposed by M¨ uller et al. (1997).

This type of specification captures long-range dependence by aggregating the log

realized volatility over the different time scales in I (daily, weekly and monthly).

Hillebrand and Medeiros (2009) consider more lags than 1, 5 and 22, as well

as dummy variables for weekdays and macroeconomic announcements and past

cumulated returns over different horizons as defined in (3). Hence,

?

where dt is a vector of n dummy variables as described above, xt−1 is defined

as in (6), rt−1= (rκ1,t−1,...,rκq,t−1)?, k = (κ1,κ2,...,κq)?is a set of q indices

with 0 < κ1< κ2< ··· < κq< ∞ and i = 1,...,κ. The final set of variables in

the model was determined by a bagging strategy as a flexible choice of the lag

structure imposes high computational costs.

yt= δ?dt+

ιi∈I

βiyιi,t−1+

?

κj∈k

λjrκj,t−1+ εt= δ?dt+ β?xt−1+ λ?rt−1+ εt(7)

3.2 The Nonlinear HAR Model

McAleer and Medeiros (2008) proposed an extension of the linear HAR model by

incorporating smooth transitions. The resulting model is called the multiple-regime

smooth transition HAR model and is defined as

M

?

where ztis a transition variable, dtand εtare defined as before, and

yt= δ?dt+ β?

0xt−1+

i=1

β?xt−1f [γi(zt− ci)] + εt

(8)

f [γi(zt− ct)] =

1

1 + e−γi(zt−ci)

(9)

is the logistic function. The authors also presented a modelling cycle based on

statistical arguments to select the set of explanatory variables as well as the number

of regimes, M.

Hillebrand and Medeiros (2009) put forward a nonlinear version of the HAR

model based on NN. Their specification is defined as follows:

yt= β?

0wt−1+

m

?

i=1

βif (γ?

iwt−1) + εt

(10)

where wt−1= (d?

function as in (9).

As first discussed in Kuan and White (1994), the model defined by equa-

tion (10) may alternatively have a parametric or a non-parametric interpretation.

In the parametric interpretation, the model can be viewed as a kind of smooth

transition regression where the transition variable is an unknown linear combination

t, x?

t−1, r?

t−1)?, εtis defined as above, and f (γ?

iwt−1) is the logistic

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 6

FORECASTING REALIZED VOLATILITY WITH LINEAR/NONLINEAR MODELS11

of the explanatory variables in wt−1(van Dijk et al., 2002). In this case, there is

an optimal, fixed number M of logistic transitions that can be understood as the

number of limiting regimes (Medeiros and Veiga, 2000; Trapletti et al., 2000;

Medeiros et al., 2006). On the other hand, for M → ∞, the NN model is a

representation of any Borel-measurable function over a compact set (Hornik et al.,

1989, 1994; Chen and Shen, 1998; Chen and White, 1998; Chen et al., 2001).

For large M, this representation suggests a non-parametric interpretation as series

expansion, sometimes referred to as sieve approximator. In this paper, we adopt

the non-parametric interpretation of the NN model and show that it approximates

typical nonlinear behaviour of realized volatility well.

As model (10) is, in principle, more flexible than model (8) we will consider

only the NN-HAR model in our empirical experiment.

4. Bagging Linear and Nonlinear HAR Models

4.1 What is Bagging?

The idea of bagging was introduced in Breiman (1996), studied more rigorously

in B¨ uhlmann and Yu (2002), and introduced to econometrics in Inoue and Kilian

(2004). Bagging is motivated by the observation that in models where statistical

decision rules are applied to choose from a set of predictors, such as significance

in pre-tests, the set of selected regressors is data dependent and random. Bootstrap

replications of the raw data are used to re-evaluate the selection of predictors, to

generate bootstrap replications of forecasts, and to average over these bootstrapped

forecasts. It has been shown in a number of studies that bagging reduces the

mean squared error of forecasts considerably by averaging over the randomness

of variable selection (Lee and Yang, 2006; Inoue and Kilian, 2008). Applications

include, among others, financial volatility (Huang and Lee, 2007; Hillebrand and

Medeiros, 2009), equity premium (Huang and Lee, 2008) and employment data

(Rapach et al., 2010).

4.2 Bagging the Linear HAR Model

Selecting the regressors in the flexible HAR model (7) involves a number of

decisions, such as the choice of significance levels for t-tests. As in Inoue and

Kilian (2004), we expect that the application of bagging will improve the forecasting

performance of the flexible HAR model.

Using the same notation as in Section 3, set wt−1= (d?

p + q + n, and write (7) as

yt= θ?wt−1+ εt

The bagging forecast for model (11) is constructed in steps as follows:

t, x?

t−1, r?

t−1)?∈ RJ, J =

(11)

Proposal 1: Bagging the linear HAR model.

(1) Arrange the set of tuples (yt,wt−1)?,t = 1,...,T, in the form of a matrix X

of dimension T × J.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 7

12MCALEER AND MEDEIROS

(2) Construct

w?∗

where the block size m is chosen to capture possible dependence in the error

term of the realized volatility series, such as conditional variance (‘volatility

of volatility’).

(3) Compute the ith bootstrap one-step ahead forecast as

⎧

⎩

S∗w∗

sample, with the jth diagonal element given by

⎧

⎩

(LS) estimator given by

?

t=1

(4) Compute the average forecast over the bootstrap samples:

bootstrapsamples of theform

{(y∗

(i)1,w?∗

(i)0),...,(y∗

(i)T,

(i)T−1)},i = 1,..., B, by drawing blocks of m rows of X with replacement,

ˆ y∗

(i)t |t−1=

⎨

0 if |tj| < c

otherwise

∀j

ˆθ

?˜ w∗

(i)t−1

where tj is the t-statistic for the null hypothesis H0:θj= 0, ˜ w∗

t−1, S∗is a diagonal selection matrix, which depends on the bootstrap

(i)t−1=

S∗

jj=

⎨

1 if |tj| ≥ c

otherwise

∀j

0

c is apre-specifiedcritical valueofthe test, andˆθ is the ordinaryleastsquares

ˆθ =

T

?

˜ w∗

(i)t−1˜ w?∗

(i)t−1

?−1

T

?

t=1

˜ w?∗

(i)t−1y∗

t

ˆ yt |t−1=1

B

B

?

i=1

ˆ y∗

(i)t |t−1

We choose a block size of m = T1/3for the bootstrap procedure described above.

This allows for dependence in the error term of equation (11). The critical value c

is set equal to 1.96, corresponding to a two-sided test at the 96% confidence level.

4.3 Bagging Nonlinear HAR Models

There are two main problems in specifying model (10): the selection of variables

in the vector x and the number of hidden units M. There are many approaches in

the literature to tackle these problems. For example, when model (10) is seen as a

variant of parametric smooth transition models, Medeiros et al. (2006) proposed a

methodology based on statistical arguments to variable selection and determination

of M. However, this approach is not directly applicable here, as we advocate

model (10) as a semi-parametric specification. On the other hand, as shown in

Hillebrand and Medeiros (2009), Bayesian regularization (BR; MacKay, 1992) is

a viable alternative, which is equivalent to penalized quasi-maximum likelihood.

However, relaying on a single specification of the model may deliver a very poor

out-of-sample performance.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 8

FORECASTING REALIZED VOLATILITY WITH LINEAR/NONLINEAR MODELS13

5001000150020002500

-6

-4

-2

0

2

4

observation

returns

S&P 500

500 1000150020002500

-1.5

-1

-0.5

0

0.5

1

observation

log realized volatility (BHLS)

Figure 1. Upper Panel: Daily Returns for the S&P 500 Index. Lower Panel: Daily Log

Realized Volatility Computed Via the Method Described in BHLS (2008) and Using the

Tukey–Hanning Kernel. We Use High-frequency Tick-by-tick on S&P 500 Futures from

2 January 1996 to 29 March 2007.

In this paper, we do not specify either the elements of x or the number of hidden

units, M. In turn, in each bootstrap sample, we randomly select M from a uniform

distribution on the interval [0, 20], and the elements of x are selected as the ones

with significant coefficients in the linear HAR case. The bagging procedure can be

summarized as follows:

Proposal 2: Bagging the NN-HAR model.

(1) Repeat steps (1) and (2) in Proposal 1.

(2) For each bootstrap sample, first remove insignificant regressors by pre-testing

as in step (3) of Proposal 1. Then, estimate the NN-HAR model randomly

selecting M from a uniform distribution on the interval [0, 20]. Compute the

ith bootstrap one-step ahead forecast and call it ˆ y∗

(3) Compute the average forecast over the bootstrap samples:

(i)t |t−1.

ˆ yt |t−1=1

B

B

?

i=1

ˆ y∗

(i)t |t−1

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 9

14MCALEER AND MEDEIROS

5001000 1500 200025003000

-5

-4

-3

-2

-1

0

1

2

3

4

5

observation

returns

FTSE 100

500 100015002000 25003000

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

observation

log realized volatility (BHLS)

Figure 2. Upper Panel: Daily Returns for the FTSE Index. Lower Panel: Daily Log

Realized Volatility Computed Via the Method Described in BHLS (2008) and Using the

Tukey–Hanning Kernel. We Use High-frequency Tick-by-tick on FTSE 100 Futures from

2 January 1996 to 28 December 2007.

5. Empirical Results

We use high-frequency tick-by-tick on S&P 500 futures from 2 January 1996 to 29

March 2007 (2796 observations) and FTSE 100 futures from 2 January 1996 to 28

December 2007 (3001 observations). In computing the daily realized volatilities,

we employ the realized kerned estimator with the modified Tukey–Hanning kernel

of BHLS (2008). As it is a standard practice in the literature, we focus on the

logarithm of the daily realized volatilities. Figures 1 and 2 illustrate the data. The

last 1000 observations are left out the estimation sample in order to evaluate the

out-of-sample performance of different models.

In this paper, we consider the following competing models: the standard HAR

model with average volatility over 1, 5 and 22 days as regressors (see equation (6));

the flexible HAR model where cumulated returns over 1 to 200 days and average

past volatility over 1 to 60 days are initially included as possible regressors; the

NN-HAR model estimated with BR and the same set of regressors as the flexible

HAR model; and finally, the NN-HAR model estimated by nonlinear LS. Bagging

is applied to all models apart from the standard HAR specification.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 10

FORECASTING REALIZED VOLATILITY WITH LINEAR/NONLINEAR MODELS15

Table 1. Forecasting Results: Main Statistics.

ModelRMSE MAEMeanSDMax. Min.

S&P 500

−0.038

−0.043

−0.096

−0.041

FTSE 100

−0.011

−0.015

−0.094

−0.016

Flexible HAR w/ bagging

NN-HAR (BR) w/ bagging

NN-HAR (LS) w/ bagging

HAR (1, 5, 22) w/o bagging

0.228

0.229

0.247

0.237

0.180

0.179

0.195

0.186

0.225

0.225

0.228

0.233

1.326

1.305

1.208

1.268

−0.853

−0.865

−0.870

−0.896

−0.900

−0.882

−1.000

−0.912

Flexible HAR w/ bagging

NN-HAR (BR) w/ bagging

NN-HAR (LS) w/ bagging

HAR (1, 5, 22) w/o bagging

0.264

0.266

0.292

0.270

0.198

0.198

0.224

0.202

0.264

0.266

0.277

0.268

1.745

1.720

1.570

1.694

The table shows the RMSE and the MAE as well as the mean, the standard deviation, the maximum

and the minimum one-step-ahead forecast error for the following models: the standard HAR model;

the flexible HAR model where cumulated returns over 1 to 200 days and average past volatility over

1 to 60 days are initially included as possible regressors; the NN-HAR model estimated with BR

and the same set of regressors as the flexible HAR model; and the NN-HAR model estimated by

nonlinear LS. Bagging is applied to all models, apart from the standard HAR specification.

Table 2. Forecasting Results: Diebold–Mariano Test.

Model Squared errors Absolute errors

S&P 500

Flexible HAR w/ bagging

NN-HAR (BR) w/ bagging

NN-HAR (LS) w/ bagging

4.52e-5

2.89e-4

0.001

1.36e-4

3.23e-4

0.004

FTSE 100

Flexible HAR w/ bagging

NN-HAR (BR) w/ bagging

NN-HAR (LS) w/ bagging

0.011

0.144

5.68e-11

0.006

0.016

1.30e-10

The table shows the p-value of the modified Diebold–Mariano test of equal predictive accuracy of

different models with respect the benchmark standard HAR model. The test is applied to the squared

errors as well as to the absolute errors. The following models are considered : the flexible HAR

model where cumulated returns over 1 to 200 days and average past volatility over 1 to 60 days are

initially included as possible regressors; the NN-HAR model estimated with BR and the same set of

regressors as the flexible HAR model; and the NN-HAR model estimated by nonlinear LS. Bagging

is applied to all models, apart from the benchmark standard HAR specification.

The forecasting results are presented in Tables 1 and 2. Table 1 shows the

root mean squared error (RMSE) and the mean absolute error (MAE) as well

as the mean, the standard deviation, the maximum and the minimum one-step-

ahead forecast error for the four models considered in the empirical exercise. From

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 11

16MCALEER AND MEDEIROS

the table it is clear that the flexible linear HAR model and the nonlinear HAR

model estimated with BR (NN-HAR (BR)) are the two best models. However, the

performance of the standard HAR specification is not much worse. On the other

hand, the NN-HAR model without BR seems to be the worst model among the

four competing ones. One possible explanation is that without BR, the NN-HAR

model can be overparametrized when M is large, leading to a very poor in-sample

estimates and out-of-sample-performance. In this case, bagging will not help. The

results are similar for the S&P 500 and the FTSE 100.

Table 2 presents the p-value of the modified Diebold–Mariano test of equal

predictive accuracy of different models with respect the benchmark standard HAR

model. The test is applied to the squared errors as well as to the absolute errors.

It is clear from the table that both the flexible linear HAR and the NN-HAR (BR)

models have superior out-of-sample performance than the standard HAR model in

the case of the S&P 500 index. For the FTSE 100, the NN-HAR (BR) model has a

statistically superior performance than the standard HAR specification only when

the absolute errors are considered.

6. Conclusions

In this paper, we considered linear and nonlinear models to forecast daily realized

volatility: the standard HAR model with average volatility over 1, 5 and 22 days as

regressors; the flexible HAR model where cumulated returns over 1 to 200 days and

average past volatility over 1 to 60 days are initially included as possible regressors;

the NN-HAR model estimated with BR and the same set of regressors as the flexible

HAR model and finally, the NN-HAR model estimated by nonlinear LS. Both the

flexible HAR and the NN-HAR (BR) models outperformed the benchmark HAR

model. The NN-HAR model estimated with nonlinear LS was the worst model

among all the alternatives considered. Finally, it is important to mention that the

models considered in this paper might be used to construct out-of-sample value-at-

risk estimates.

Acknowledgements

The first author acknowledges the financial support of the Australian Research Council

and National Science Council, Taiwan. The second author thanks the CNPq/Brazil for

partial financial support.

Notes

1. In fact, there is an abuse of terminology here as ‘realized volatility’ specifically

refers to the square root of the sum of the squared intraday returns, which is a biased

and inconsistent estimator of the daily integrated volatility under the presence of

micro-structure noise. However, to simplify notation and terminology, we will refer

to any unbiased and consistent estimator as realized volatility.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 12

FORECASTING REALIZED VOLATILITY WITH LINEAR/NONLINEAR MODELS17

References

Andersen, T. and Bollerslev, T. (1998) Answering the skeptics: yes, standard volatility

models do provide accurate forecasts. International Economic Review 39: 885–906.

Andersen, T., Bollerslev, T., Diebold, F.X. and Ebens H. (2001) The distribution of

realized stock return volatility. Journal of Financial Economics 61: 43–76.

Andersen, T., Bollerslev, T., Diebold, F.X. and Labys, P. (2003) Modeling and forecasting

realized volatility. Econometrica 71: 579–625.

Asai, M., McAleer, M. and Yu, J. (2006) Multivariate stochastic volatility: a review.

Econometric Reviews 25: 145–175.

Bandi, F.M. and Russell, J.R. (2006) Separating market microstructure noise from

volatility. Journal of Financial Economics 79: 655–692.

Barndorff-Nielsen, O. and Shephard, N. (2002) Econometric analysis of realised volatility

and its use in estimating stochastic volatility models. Journal of the Royal Statistical

Society B 64: 253–280.

Barndorff-Nielsen, O., Hansen, P., Lunde, A. and Shephard, N. (2008) Designing realized

kernels to measure the ex-post variation of equity prices in the presence of noise.

Econometrica 76:1481–1536.

Biais, B., Glosten, L. and Spatt, C. (2005) Market microstructure: a survey of

microfoundations, empirical results, and policy implications. Journal of Financial

Markets 8: 217–264.

Bollerslev, T. (1986) Generalized autoregressive conditional heteroskedasticity. Journal

of Econometrics 21: 307–328.

Bollerslev, T. (1987) A conditionally heteroskedastic time series model for speculative

prices and rates of return. Review of Economics and Statistics 69: 542–547.

Breiman, L. (1996) Bagging predictors. Machine Learning 36: 105–139.

B¨ uhlmann, P. and Yu, B. (2002) Analyzing bagging. Annals of Statistics 30: 927–

961.

Campbell, J., Lo, A. and Mackinlay, A. (1997) The Econometrics of Financial Markets.

Princeton, NJ: Princeton University Press.

Carnero, M.A., Pe˜ na, D. and Ruiz, E. (2004) Persistence and kurtosis in GARCH and

stochastic volatility models. Journal of Financial Econometrics 2: 319–342.

Chen, X. and Shen, X. (1998) Sieve extremum estimates for weakly dependent data.

Econometrica 66: 289–314.

Chen, X. and White, H. (1998) Improved rates and asymptotic normality for nonparametric

neural network estimators. IEEE Transactions on Information Theory 18: 17–

39.

Chen, X., Racine, J. and Swanson, N.R. (2001) Semiparametric ARX neural-network

models with an application to forecasting inflation. IEEE Transactions on Neural

Networks 12: 674–683.

Corsi, F. (2009) A simple approximate long memory model of realized volatility. Journal

of Financial Econometrics 7: 174–196.

Corsi, F., Mittnik, S., Pigorsch, C. and Pigorsch, U. (2008) The volatility of realized

volatility. Econometric Reviews 27: 46–78.

van Dijk, D., Ter¨ asvirta, T. and Franses, P.H. (2002) Smooth transition autoregressive

models – a survey of recent developments. Econometric Reviews 21: 1–47.

Engle, R.F. (1982) Autoregressive conditional heteroskedasticity with estimates of the

variance of United Kingdom inflation. Econometrica 50: 987–1007.

Hansen, P.R. and Lunde, A. (2006) Realized variance and market microstructure noise

(with discussion). Journal of Business and Economic Statistics 24: 127–218.

Hillebrand, E. and Medeiros, M.C. (2009) The benefits of bagging for forecast models of

realized volatility. Econometric Reviews, to be published.

Hornik, K., Stinchombe, A. and White, H. (1989) Multi-layer feedforward networks are

universal approximators. Neural Networks 2: 359–366.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

Page 13

18MCALEER AND MEDEIROS

Hornik, K., Stinchombe, A., White, H. and Auer, P. (1994) Degree of approximation

results for feedforward networks approximating unknown mappings and their

derivatives. Neural Computation 6: 1262–1274.

Huang, H. and Lee, T.-H. (2007) Forecasting using high-frequency financial time series.

Working Paper, University of California at Riverside.

Huang, H. and Lee, T.-H.(2008) To combine forecasts or to combine information.

Econometric Reviews, to be published.

Inoue, A. and Killian, L. (2004) Bagging time series models. Discussion Paper No. 4333,

Centre for Economic Policy Research (CEPR).

Inoue, A. and Killian, L. (2008) How useful is bagging in forecasting economic time

series? A case study of U.S. CPI inflation. Journal of the American Statistical

Association 103: 511–522.

Jondeau, E., Poon, S.-H. and Rockinger, M. (2007) Financial Modeling Under Non-

Gaussian Distributions. London: Springer.

Kuan, C.-M. and White, H. (1994) Artificial neural networks: an econometric perspective.

Econometric Reviews 13: 1–91.

Lee, T.-H. and Yang, Y. (2006) Bagging binary and quantile predictors for time series.

Journal of Econometrics 135: 465–497.

Mackay, D.J.C. (1992) A practical Bayesian framework for backpropagation networks.

Neural Computation 4: 448–472.

Madhavan, A. (2000) Market microstructure: a survey. Journal of Financial Markets 3:

205–258.

Malmsten, H. and Ter¨ asvirta, T. (2004) Stylized facts of financial time series and three

popular models of volatility. Working Paper Series in Economics and Finance 563,

Stockholm School of Economics.

McAleer, M. (2005) Automated inference and learning in modeling financial volatility.

Econometric Theory 21: 232–261.

McAleer, M. and Medeiros, M. (2008) A multiple regime smooth transition heterogeneous

autoregressive model for long memory and asymmetries. Journal of Econometrics

147: 104–119.

Meddahi, N. (2002) A theoretical comparison between integrated and realized volatility.

Journal of Applied Econometrics 17: 479–508.

Medeiros, M.C. and Veiga, A. (2000) A hybrid linear-neural model for time series

forecasting. IEEE Transactions on Neural Networks 11: 1402–1412.

Medeiros, M.C., Ter¨ asvirta, T. and Rech, G. (2006) Building neural network models for

time series: a statistical approach. Journal of Forecasting 25: 49–75.

Merton, R.C. (1980) On estimating the expected return on the market: an exploratory

investigation. Journal of Financial Economics 8: 323–361.

M¨ uller, U.A., Dacorogna, M.M., Dave, R.D., Olsen, R.B., Puctet, O.V. and von

Weizsacker, J. (1997) Volatilities of different time resolutions – analyzing the

dynamics of market components. Journal of Empirical Finance 4: 213–239.

Patton, A. (2008) Volatility forecast evaluation and comparison using imperfect volatility

proxies. Journal of Econometrics, to be published.

Rapach, D. and Strauss, J. (2007) Bagging or combining (or both)? An analysis based on

forecasting U.S. employment growth. Working Paper, Saint Louis University.

Rapach, D., Strauss, J. and Zhou, G. (2010) Out of sample equity premium prediction

and links to the real economy. Review of Financial Studies 23: 821–862.

Taylor, S.J. (1986) Modelling Financial Time Series. Chichester: Wiley.

Trapletti, A., Leisch, F. and Hornik, K. (2000) Stationary and integrated autoregressive

neural network processes. Neural Computation 12: 2427–2450.

Zhang, L., Mykland, P. and Ait-Sahalia, Y. (2005) A tale of two time scales: determining

integrated volatility with noisy high-frequency data. Journal of the American

Statistical Association 100: 1394–1411.

Journal of Economic Surveys (2011) Vol. 25, No. 1, pp. 6–18

C ?2010 Blackwell Publishing Ltd

#### View other sources

#### Hide other sources

- Available from Marcelo C Medeiros · Jul 28, 2014
- Available from puc-rio.br
- Available from SSRN