Iran. Econ. Rev. Vol. 24, No. 1, 2020. pp.

299-311

Modeling Gold Volatility: Realized GARCH Approach

Esmaiel Abounoori*1, Mohammad Amin Zabol2

Accepted: 2018, September 29 Received: 2018, November 4

Abstract
F orecasting the volatility of a financial asset has wide implications in
finance. Conditional variance extracted from the GARCH
framework could be a suitable proxy of financial asset volatility. Option
pricing, portfolio optimization, and risk management are examples of
implications of conditional variance forecasting. One of the most recent
methods of volatility forecasting is Realized GARCH (RGARCH) that
considers a simultaneous model for both realized volatility and
conditional variance at the same time. In this article, we estimate
conditional variance with GARCH, EGARCH, GIR-GARCH, and
RGARCH with two realized volatility estimators using gold intraday
data. We compared models, for in-sample fitting; by the log-likelihood
value and used MSE and QLIKE lose functions to evaluate predicting
accuracy. The results show that the RGARCH method for GOLD
outperforms the other methods in both ways. So, using the RGARCH
model in practical situations, like pricing and risk management would
tend to better results.
Keywords: Realized GARCH, Gold, GARCH Models, Volatility.
JEL Classification: G10, G15, G17.

1. Introduction
Examples of volatility prediction applications include option pricing,
optimal portfolio selection, and risk management. Variance, standard
deviation, conditional variance, or all of these indicators, is, in fact, a
proxy for volatility and a tool for its estimation. Today, in many types
of research, the conditional variance estimated from GARCH family
models has used a reasonable estimate of volatility.
In conventional GARCH models, only daily stock returns are used
to predict daily conditional variance. Since the information obtained
from daily returns is lower than the different criteria derived from
1. Department of Economics, Semnan University, Semnan, Iran (Corresponding Author:
esmaiel.abounoori@semnan.ac.ir).
2. Department of Economics, Semnan University, Semnan, Iran (m.zabol@semnan.ac.ir).
300/ Modeling Gold Volatility: Realized GARCH Approach …

intraday data, the information set of conventional GARCH models is

limited. In addition, since GARCH models are based on moving
average with decreasing weight, these models act a bit slow to react to
volatility changes (Andersen et al., 2003). Thus, there was a tendency
to introduce the intraday criteria in the framework of GARCH models.
Engel (2002) proposed to include realized variance as an exogenous
variable in GARCH models. The disadvantage of this innovation is
that only one-day ahead conditional variance prediction is possible.
Recently, Hansen et al. (2012) presented a model called “realized
GARCH”, which provides a framework for modeling the realized
volatility and conditional variance at the same time in one model.
In this paper, using the Gold five-minutes intra-day data from April
2012 to April 2018, we compare the realized GARCH model with
some conventional GARCH models such as GARCH, EGARCH, and
GJR-GARCH. The comparison will be in two ways. First, we consider
how well data has been fitted in the models. Then, we examine the
accuracy of the prediction of the conditional variance of the sample by
using the rolling window approach and using a loss function to select
the most accurate model.

2. Literature Review
The GARCH models are widely used in finance. Concerning the case
of option valuation, the latest studies have been done by Badzko et al.
(2015) and Huang et al. (2017); the latest study showed that using the
GARCH model for the S&P index is more appropriate than another
volatility method in the case of option pricing. Regarding portfolio
optimization, GARCH is used by Ðrnovkiewicz et al. (2016) and
Sahamköping et al. (2018).
In order to explain how the GARCH models emerge, we should put
aside the heteroscedasticity assumption of the linear regression. Engel
(1982) introduces a particular type of heteroscedasticity in which the
variance of the innovation term is a function of the lags of squared
innovations. The introduction of this kind of heterogeneity variance
provided a very important tool for economists and especially
researchers in the field of financial econometrics to measure and
estimate the conditional variance of a series. Consider an AR(1)
process for asset returns as in equation (1).
 Iran. Econ. Rev. Vol. 24, No.1, 2020 /301

rt  0  1rt 1   t (1)

Where rt represents the asset returns and εt is an iid with zero means.
In this case, the conditional variance of εt may vary over time and be
a function of the previous innovations. This model was originally
presented by Engel (1982). The reason for introducing this model was
that although it is observed, εt are independent, but their square is
related to each other. Engel suggested the following equation for the
conditional variance εt , which is known as ARCH (p).

σ2t  a 0  a1ε 2t 1  a pε 2t p (2)

The number of the lag of an ARCH model, p, is large, leading to an

increase in the number of estimated parameters. As a result, Bollerslev
(1986) suggested the following model overcome this problem.

p q
σ  a 0  a ε
2
t
2
i t i  b jσ 2t  j (3)
i 1 j1

Where ai and bj are assumed to be positive to ensure that the variance

is positive. This model is known as the Generalized ARCH model,
GARCH (p, q). If q is zero, then this model will be reduced to
ARCH(p).
Since in the GARCH model, εt appear in squared form in the
equation, the sign of these shocks does not affect the conditional
variance. Meanwhile, it has been observed that negative shocks or bad
news increase the variance more than the positive shocks or good
news. Also, the parameters of a GARCH model should be restricted to
ensure that the variance stays positive. For this purpose, Nelson
(1991) introduced the EGARH model as follows.

p
ε t  i  γ i ε t i q
ht  a 0  a i  b jh t  j
i 1 σ t i j1 (4)
h t  lnσ 2
t
302/ Modeling Gold Volatility: Realized GARCH Approach …

In this model, when εt is positive, the total shock effect is(1 +

γt )εt , and if there is bad news, the total shock effect will be as large
as(1 − γt )εt . If bad news is supposed to have a higher variance, we
expect γ to be negative. Apart from the fact that the effects of good
and bad news are considered differently invariance, this model has the
advantage of the GARCH model, which variance will always be
positive for any coefficients.
Another way to consider the effect of good and bad news on the
variance is by using a dummy variable as follows:

p p q
σ  a 0  a ε
2
t
2
i t i  γ iS ε 2
t i t i  b jσ 2t  j (5)
i 1 i 1 j1

In which St−i is a dummy variable that is zero if εt−1 is positive,

and 1 if εt−1 is negative. In this case, the effect of a positive shock is
ai ε2t−i and the effect of a negative shock is(ai +γi )ε2t−i . By assumption
of the effect of bad news on the variance, we expect γi to be positive.
This model is known as GJR introduced by Glosten, Jagannathan, and
Runkle (1993).
After these initial models, a number of models and explanations
were introduced for conditional variance modeling. Some researchers
tried to use another explanatory variable apart from the squared
innovation in the model, which became known as GARCH-X models.
In 2003, Engle first used the realized variance criteria to explain the
conditional variance. This effort, although was an improvement of the
GARCH models, but actually, it was the same GARCH-X
specification. In fact, the realized variance was added as an exogenous
variable to the GARCH model.
An important problem of using realized volatility in a GARCH-X
framework is that the models only can predict variance for the next
period. Engel and Gallo (2006) introduced the first model in which a
structure was presented for realized volatility so that an additional
latent volatility process was considered for each realized measure in
the model. This model is known as the Multiplicative Error Model.
Another model concerning realized volatility as an endogenous
variable is the HEAVY model presented in Shepard and Sheffard
 Iran. Econ. Rev. Vol. 24, No.1, 2020 /303

(2010), which as a mathematical view is a nested MEM model. Unlike

traditional GARCH models, these models operate on the basis of
multivariate latent volatility processes. For example, three latent
volatility processes are used in a MEM model, but the HEAVY model
includes at least two latent volatility processes.
In one of the most recent works done in this regard, Hansen et al.
(2012), introduced the third equation into a GARCH model to
simulate the realized volatility in an Endogenous way. They consider
conditional variance as a function of realized volatility as follows:

rt  h t ε t (6)

h t  ω  βh t 1  γx t 1 (7)
x t  ξ  φh t  τ  ε t   u t (8)

In which 𝑥𝑡 is the realized volatility, ℎ𝑡 denotes the conditional

variance, and τ (.) is the leverage function. Equation (6) and (7)
represent the mean and conditional variance, respectively as in
conventional GARCH models replacing equation (8) in equation (7).
Adding regression (8) distinguishes realized GARCH models from
conventional GARCH models. This equation is known as the
measurement equation because it relates the realized measure to the
hidden volatility (Hansen et al., 2012). Leverage function in the
simplest form could be zero but Hansen (2012) suggest𝜏(𝜀𝑡 ) =
𝜆1 (𝜀𝑡 ) + 𝜆2 (𝜀𝑡2 − 1), so this function can capture asymmetric
behavior of shocks too. This model is known as the RGARCH model.
Tian et al (2015) used this method and traditional GARCH methods to
estimate the daily volatility of the short-term interest rate in the euro–
yen market. The results have indicated that the RGARCH model has
better performance than traditional GARCH models regarding the
prediction of the conditional variance. Sharma & Vipul (2016)
investigate the variance predict the performance of the RGARCH
model for 16 stock indices in a 14-year period. They indicated that the
results are sensitive to the performance decision criteria. Wei et al.
(2017) construct a new realized GARCH model by introducing the
perturbation of leveraged parameter in the volatility equations of the
304/ Modeling Gold Volatility: Realized GARCH Approach …

realized GARCH model. The empirical analysis of the high-frequency

data of the Shanghai Stock Exchange 50 index shows that using their
new model can improve the prediction accuracy of a measure of risk
to a certain extent. Huang et al. (2017) conduct an extensive empirical
analysis on S&P500 index options using realized GARCH and the
results show that their computationally fast formula outperforms
competing methods in terms of pricing errors, both in-sample and out-
of‐sample. Jiang et al. (2018) apply realized GARCH models by
introducing several risk measures of intraday returns into the
measurement equation, to model the daily volatility of E-mini S&P
500 index futures returns. The empirical results show that realized
GARCH models using the generalized realized risk measures provide
better volatility estimation for the in-sample and substantial
improvement in volatility forecasting for the out-of-sample.
Although we expect that the RGARCH model outperforms other
GARCH models, but there is some evidence that it depends on the
type of asset. In this paper, we intend to examine the performance of
the RGARCH model in predicting the conditional variance of GOLD
in comparison with other GARCH family models.

3. Methodology
In order to estimate a realized GARCH model, intraday data should be
used so the realized volatility which is needed in this procedure could
be calculated. So, we have used five-minutes gold trades in forex
(XAU/USD)1. The data has been gathered from April 2012 to April
2018.
In GARCH models, there is a mean equation, which in this paper is
considered as an AR(1) prose.

rt  0  1rt 1   t (9)

In which rt is the daily return on financial assets. This

specification of the model makes it possible to assume that the return
is related to a constant amount, along with a coefficient of return on

1. The data is available in many forex source such as link below:

https://www.dukascopy.com/swiss/english/marketwatch/historical/
 Iran. Econ. Rev. Vol. 24, No.1, 2020 /305

the trading day before and the shock of that day. We also consider
equations (10) to (12) for the estimation of the traditional GARCH
models as the following;

GARCH(1,1) σ 2t  a 0  a1ε 2t 1  b1σ 2t 1 (10)

EGARCH(1,1) ε t 1  γε t 1
ht  a 0  a1  b1h t 1 , ht  ln  t2 (11)
σ t 1
GJR-GARCH(1,1) σ 2  a  a ε 2  γS ε 2  b σ 2 (12)
t 0 1 t 1 t 1 t 1 1 t 1

And RGARCH as follows;

h t  ω  βh t 1   x t 1 , h t  ln  t2
x t  ξ  φh t  τ  ε t   u t (13)
τ  ε t   1ε t  2 (εt2 1)

In which 𝑥𝑡 is the realized volatility, ℎ𝑡 denotes the conditional

variance, and τ (.) is the leverage function that relates the volatility to
variance equation which distinguishes realized GARCH models from
conventional GARCH models. This equation is also known as the
measurement equation because it relates the realized measure to the
hidden volatility (Hansen et al.. 2012). We used two proxies of
realized volatility which are RV and BV. So we mention RV-GARCH
and BV-GARCH as the realized GARCH models with using RV and
BV proxies of the realized volatility respectively. These two proxies
can be calculated as follows:

M
RVt   r i ,t
2
(14)
i 1

M
M
BVt ()   2  rt ,i rt ,i 1
( M  1) i  2
(15)

Where 𝑟𝑖,𝑡 represents the ith intraday return in day t, M denotes the
number of trades in a day and 𝜇 is equal to√2/𝜋 ≅ 0.79788. One
306/ Modeling Gold Volatility: Realized GARCH Approach …

important difference between RV and BV is that the first is not robust

to jumps but the latter is.
A good model not only should fit data well but also it should have
more accurate performance in predicting out of sample volatility. So
in this paper both in-sample and out of sample modeling performance
has been investigated. For in-sample fitting, as the method of
estimation is ML, the Log L value is a reasonable measure to compare
models.
To compare the predictive performance of conventional GARCH
models to the realized GARCH model, the rolling window technique
of size 500 has been used. Therefore we used the first 500 data to
predict the conditional variance of day 501 and compared it with the
actual volatility of the same day. Then we omit the first day and add
day 501 to predict volatility and compare it with actual one of day 502
and so on.
we used the method of Hansen et al. (2005) to estimate the actual
volatility as follow:

n
n 1  (rt  ˆ ) 2
 t2  cˆ.RV t , cˆ  t 1
n
(16)
n 1
 RV
t 1
t

Where n is the number of the trading day, 𝑥𝑡 is the daily return and 𝜇̂
denotes the average of daily returns in n days. After defining the
actual volatility, a loose function should be used in order to rank the
models by the accuracy of predicted volatility. Patto (2011) showed
that among 9 loose functions to rank the volatility, only MSE and
Qlike are robust in the possible existence of proxy error and these two
only should be used. These equations are as follow:

MSE  E  l1,k,t  , l1,k,t  (σ 2t  σ̂ 2t ) 2 (17)

σ 2t
QLIKE  E  l2,k,t  , l2,k,t  log  σ̂ 2t   (18)
σ̂ 2t

The lower lose function, the better model and among all models,
 Iran. Econ. Rev. Vol. 24, No.1, 2020 /307

the model with the lowest value in the loose function is the most
accurate model.

4. Empirical Results and Discussion

In order to estimate the model, we have used five-minutes gold trades
in forex (XAU/USD) from April 2012 to April 2018. Table 1 shows
the daily GOLD return in which indicates that 90% of the returns are
between -1.3% and 1.3%. Kurtosis and Skewness calculated from the
winsorized gold return at 1st and 99th percentiles1 suggest that the
center part of the returns distribution doesn’t differ much from the
normal distribution.

Table 1: Descriptive Statistics of Daily GOLD Return in Full Sample

5th 95th
n smallest mean median largest std. dev. Skewness* Kurtosis*
percentile percentile
1941 -9.50% 1.30% 0 0 1.30% 4.90% 0.86% -0.04 3.97
Note: Kurtosis and Skewness are calculated from the winsorized gold return at 1 st
and 99th percentiles.

The predictive performance of GARCH models should be

investigated in both in the sample and out of sample data. Table 2
represents the log-likelihood of GARCH models. As the logL for RV-
GARCH is bigger than other models, it means the RV-GARCH model
fits the gold data better.

Table 2: LogL Value of the Models Estimated in Full Sample

MODEL Log L Rank
GARCH -1685.792 4
EGARCH -1685.858 5
GJR-GARCH -1685.789 3
RV-GARCH -1676.404 1
BV-GARCH -1677.098 2

For out of sample performance evaluation of GARCH models,

one-step-ahead conditional variances of the models were predicted
and the value of loose functions for each model was calculated. The

1. Winsorised the data means data less than 1% taken as 1th percentile and the data more than
99% is taken as 99th percentile this will remove the outlier data.
308/ Modeling Gold Volatility: Realized GARCH Approach …

result is shown in table 3. As it could be seen, the top two models are
BV-GARCH and RV-GARCH Respectively. İt implies that for
predicting Gold volatility, the most accurate model is BV-GARCH. İn
figure 1 time-series of realized daily variance and conditional variance
predicted one-step ahead for the best model, BV-GARCH, is depicted.

Table 3: The Value of Lose Function

MODEL MSE (*1000000) Rank Qlike Rank
GARCH 6.07505 3 -4084.77 3
EGARCH 6.10756 5 -4073.43 4
GJR-GARCH 6.07724 4 -4051.65 5
RV-GARCH 6.06677 2 -4095.48 2
BV-GARCH 5.88260 1 -4102.83 1
.0015
.001
.0005

2014-01-01 2015-01-01 2016-01-01 2017-01-01 2018-01-01

REALIZED BV_GARCH

Therefore, we observed that the realized GARCH models (BV-

GARCH and RV-GARCH, respectively) had the highest performance
in terms of both in-sample and out-of-sample prediction of conditional
variance (volatility). Since over-estimation and under-estimation of
risk, would costs to the investor, the use of the RGARCH models can
minimize this cost.
 Iran. Econ. Rev. Vol. 24, No.1, 2020 /309

5. Conclusion
In this paper, using the Gold five-minutes intra-day data from April
2012 to April 2018, we estimate the conditional variance of GARCH,
EGARCH, and GJR-GARCH as well as the RGARCH model using
two RV and BV proxies for intra-day realized volatility. A good
model not only should fit data well but also it should have accurate
performance in predicting out of sample volatility. So in this paper
both in-sample and out of sample modeling performance has been
investigated. We compared models, for in-sample fitting, by the log-
likelihood value and used MSE and QLIKE lose functions to evaluate
predicting accuracy. The results show that the RGARCH method for
GOLD outperforms the other methods in both ways.
Therefore, the use of RGARCH models instead of conventional
GARCH models provides a more accurate estimate for the conditional
variance as a proxy of volatility which is a key factor in many risk
management and portfolio management.

References
Andersen, T. G., Bollerslev, T., Diebold, F. X., & Labys, P. (2003).
Modeling and Forecasting Realized Volatility. Econometrica, 7(2),
579-625.

Badescu, A., Elliott, R. J., & Ortega, J. P. (2015). Non-Gaussian

GARCH Option Pricing Models and Their Diffusion
Limits. European Journal of Operational Research, 247(3), 820-830.

Barndorff-Nielsen, O. E. (2004). Power and Biopower Variation with

Stochastic Volatility and Jumps. Journal of Financial
Econometrics, 2(1), 1-37.

Bollerslev, T. (1986). Generalized Autoregressive Conditional

Heteroskedasticity. Journal of Econometrics, 31(3), 307-327.

Engle, R. (2002). New Frontiers for Arch Models. Journal of Applied

Econometrics, 17(5), 425-446.

Engle, R. F. (1982). Autoregressive Conditional Heteroskedasticity

with Estimates of the Variance of United Kingdom Inflation.
Econometrica, 50(4), 987-1007.
310/ Modeling Gold Volatility: Realized GARCH Approach …

Engle, R. F., & Gallo, G. M. (2006). A Multiple Indicators Model for

Volatility Using Intra-Daily Data. Journal of Econometrics, 131(1-
2), 3-27.

Glosten, L. R., Jagannathan, R., & Runkle, D. E. (1993). On the Relation

between the Expected Value and the Volatility of the Nominal Excess
Return on Stocks. Journal of Finance, 48(5), 1779-1801.

Hansen, P. R., & Lunde, A. (2005). A Forecast Comparison of

Volatility Models: Does Anything Beat a GARCH (1,1)? Journal of
Applied Econometrics, 20(7), 873-889.

Hansen, P. R., Huang, Z., & Shek, H. H. (2012). Realized GARCH: A

Joint Model for Returns and Realized Measures of Volatility. Journal
of Applied Econometrics, 27(6), 877-906.

Huang, Z., Wang, T., & Hansen, P. R. (2017). Option Pricing with the
Realized GARCH Model: An Analytical Approximation
Approach. Journal of Futures Markets, 37(4), 328-358.

---------- (2017). Option Pricing with the Realized GARCH Model: An

Analytical Approximation Approach. Journal of Futures
Markets, 37(4), 328-358.

Jiang, W., Ruan, Q., Li, J., & Li, Y. (2018). Modeling Returns
Volatility: Realized GARCH Incorporating Realized Risk
Measure. Physica A: Statistical Mechanics and its Applications, 500,
249-258.

Nelson, D. B. (1991). Conditional Heteroskedasticity in Asset

Returns: A Nee Approach. Econometrica, 59(2), 347-370.

Patton, A. J. (2011). Volatility Forecast Comparison Using Imperfect

Volatility Proxies. Journal of Econometrics, 160(1), 246-256.

Ranković, V., Drenovak, M., Urosevic, B., & Jelic, R. (2016). Mean-
univariate GARCH VaR Portfolio Optimization: Actual Portfolio
Approach. Computers & Operations Research, 72, 83-92.
 Iran. Econ. Rev. Vol. 24, No.1, 2020 /311

Sahamkhadam, M., Stephan, A., & Östermark, R. (2018). Portfolio

Optimization Based on GARCH-EVT-Copula Forecasting
Models. International Journal of Forecasting, 34(3), 497-506.

Sharma, P. (2016). Forecasting Stock Market Volatility Using

Realized GARCH Model: International Evidence. The Quarterly
Review of Economics and Finance, 59, 222-230.

Shephard, N., & Sheppard, K. (2010). Realizing the Future:

Forecasting with High-Frequency-Based Volatility (Heavy)
Models. Journal of Applied Econometrics, 25(2), 197-231.

Wei, Z. Y., Luo, Y. F., Yu, D. Y., & Wang, A. F. (2017). The
Measure of Risk for SSE 50 Index Based on Realized NGARCH
Model. Journal of Chongqing University of Technology (Natural
Science), 5, Retrieved from
http://en.cnki.com.cn/Article_en/CJFDTotal-CGGL201705030.htm.