Department of Statistics

Master's Thesis

Modelling and Forecasting Volatility of Gold Price with

Other Precious Metals Prices by Univariate GARCH
Models

Yuchen Du①

Supervisor: Lars Forsberg

①
[email protected]
Modelling and Forecasting Volatility of Gold Price with Other
Precious Metals Prices by Univeriate GARCH Models

Yuchen Du , June, 2012

Abstract

This paper aims to model and forecast the volatility of gold price with the help of
other precious metals. The data applied for application part in the article involves
three financial time series which are gold, silver and platinum daily spot prices. The
volatility is modeled by univariate Generalized Autoregressive Conditional
Heteroskedasticity (GARCH) models including GARCH and EGARCH with different
distributions such as normal distribution and student-t distribution. At the same time,
comparisons of estimation and forecasting the volatility between GARCH family
models have been done.

Keywords: gold price, volatility, precious metal, GARCH, EGARCH, volatility

forecasting
CONTENT

1.Introduction ................................................................................................................. 1

1.1 Background ...................................................................................................... 1

1.2 Brief Introduction to the methodology ............................................................ 2

2.Methodology ............................................................................................................... 3

2.1 Standard GARCH Model ................................................................................. 3

2.2 Exponential GARCH (EGARCH) Model ....................................................... 4

2.3 Univariate GARCH Models with error terms.................................................. 5

2.4 Distributions of Error Terms ........................................................................... 6

2.5 Root Mean Square Error (RMSE) ................................................................... 8

3. Data ............................................................................................................................ 9

3.1 Returns ............................................................................................................. 9

3.2 Plots ............................................................................................................... 10

3.3 Data Analysis ................................................................................................. 11

4. Estimation ................................................................................................................ 13

4.1 GARCH Models ............................................................................................ 13

4.2 EGARCH Models .......................................................................................... 15

5. Forecast .................................................................................................................... 16

6. Conclusion ............................................................................................................... 18

Reference ..................................................................................................................... 20

Appendix ...................................................................................................................... 21
1.Introduction

1.1 Background

During the end of the World War II, in 1944, forty-four countries around the
world signed the well-known Bretton Woods Agreement to establish a new
currency system based on gold. They specified the value of the US Dollar is
connected to gold and other currencies are connected to the US Dollar. This
system indicated where gold standing, not only just as an invest good. This metal
has been the only commodity that served as currency in international financial
business for centuries. Gold is so ideal that a number of currencies were based on
the value of it. For residents, it is an excellent investment good since it could
protect them from assets devaluation. For a country, it is an important reserve
asset; the situation is somewhat the same as the other precious metals. The
best-known precious metals are gold and silver. While both have industrial uses,
they are better known for their usage in art, jewellery as well. Other precious
metals include the platinum groupmetals: palladium, osmium, iridium, and
platinum which is the most widely traded one in this group. Historically, precious
metals have commanded much higher prices than common industrial metals, like
②
copper and iron.

During the past few years, especially after the chaos of the US dollar system
caused by the financial crisis, investors who have been able to master the
reasonable investment of the precious metals markets have gotten remarkable
returns. These metals could play an essential role in a portfolio because of their
ability to act as a store of value and their stability against inflation as well.
Therefore, the modeling and forecasting of the volatility of gold price is
indispensable.

②
''Precious Metal", http://www.wikipedia.org/ . See more in <http://en.wikipedia.org/wiki/Precious_metal.>

1
With the common sense that higher profit comes with higher risk, modeling
volatility in asset returns is concerned by financial investors. Basically, volatility
is widely-used for measuring assets risk; investors want guarantees for
investing in risky assets for their investment decision. The demand of assisting
investors to handle the volatility of commodities prices need to be supplied
by some creative and relative financial instruments.

1.2 Brief Introduction to the methodology

Although it is an urgent need of modelling volatility, there were still some

limitations and difficulties. As common financial time series, the usual commodity
spot price data always have properties like volatility clustering, fat-tailness, excess
kurtosis and skewness. The widely-used classical linear regression model usually
considers the residuals have homoscedasticity which means random disturbances
have equal variance. That is the exact difficulty using traditional linear tools.

"Models of Autoregressive Conditional Heteroskedasticity (ARCH) form the most

popular way of parameterizing this dependence", Teräsvirta (2006), p.2.

The autoregressive conditional heteroskedasticity (ARCH) which is the first model

of conditional heteroskedasticity was proposed by Engle (1982). Then the
generalized autoregressive conditional heteroskedasticity (GARCH) model was
proposed by Bollerslev and Taylor (1986) separately. Since it was brought into use,
GARCH model shows better capability of financial time series analysis in some
empirical studies. Thereafter, many GARCH family models came up, such as
exponential GARCH which was brought up by Nelson (1991), Asymmetric Power
ARCH which was carried out by Ding, Z., Engle, R.F. and Granger, C.W.J., (1993).
Among the worldwide introduced GARCH family models, Teräsvirta (2006)
thought that the overwhelmingly most popular GARCH model in applications has
been the GARCH (1,1) model.

2
2. Methodology

Since Engle (1982) introduced Autoregressive Conditional Heteroskedasticity which

allows the conditional variance to change over time as a function of past errors and
Bollerslev (1986) generalized it which allows for a much more flexible lag structure.
Nelson (1991) brought out a new model with a new form of heteroskedasticity, called
Exponential GARCH (EGARCH).

2.1 Standard GARCH Model

Let {rt } denote a financial time series, presenting returns in this paper.

rt  E (rt | t 1 )   t

Here, t 1 is an information set contained all the information through

time period t  1 , following Bollerslev (1986). We can see here the

conditional mean and variance both change with t 1 .

The function of the conditional mean t ( ) ,

E(rt | t 1 )  t ( )  0  1rt 1     p rt  p  1 t 1    q t q

which can also be considered as an AutoRegressive MovingAverage

(ARMA)(p,q) process.

Then an AutoRegressive Conditional Heteroskedasticity - ARCH model

proposed by Engle (1982) can be regarded as the variance function,

Var (rt | t 1 )  E ( t2 | t 1 )  h t ( )

Here, there is an assumption that the conditional distribution of  t is supposed

3
to be normal distribution with zero mean and variance equals to ht .

The ARCH(q) process function form is like this:

 t | t 1 ~ N (0,h t )
q
h t      i t2i
i 1

The Generalized ARCH (GARCH) process introduced by Bollerslev (1986) is

given below with the same assumption.

 t | t 1 ~ N (0,h t )
q p
ht      i t2i    j ht  j
i 1 j 1

where p  0, q  0， 0  0, i  0, i  1,..., q, i  0, j  1,..., p

p
We can call  h
j 1
j t j part as GARCH terms. For p=0 which means there is no

GARCH terms, the process becomes an ARCH(q) process; for p=q=0,  t is

simply white noise.

As Teräsvirta said, usually the simplest GARCH(1,1) model works very

well:

 t |  t -1 ~ N (0, h t )
ht     t21  ht 1

with   0,   0,   0 .

2.2. Exponential GARCH (EGARCH) Model

The Exponential GARCH(EGARCH) Model was introduced by Nelson (1991) in

order to model asymmetric variance effects. It has some special properties which

4
makes it especially efficient in an asset pricing context. The basic frame for the mean
and variance is similar to standard GARCH model, the conditional variance function
part is very unique.

Given the EGARCH(p,q) process, with similarity of the conception in standard

GARCH,

rt  E (rt | t 1 )   t

 t | t 1 ~ N (0, ht )

p q
ln ht   0   j g j ( Zt  j )    j ln ht  j
j 1 j 1

where g (Zt  j )  Zt  j   (| Zt  j |  E(| Zt  j |)), j  1,..., q and Z t could be a standard

normal variable. The part relative to Z t allows the sign and the magnitude effects

affect the volatility separately which is the very property making EGARCH may be
more suitable for price return data.

By a more specific way, an EGARCH (p,q) process can also be presented like this,

 t | t 1 ~ N (0, ht )

 
p q
ln( ht )      j ln( ht  j )  i |  t i | / ht i  E (|  t i | / ht i )   ( t i / ht i )
j 1 i 1

2.3 Univariate GARCH Models with extra error terms

In this paper, the aim is to model and forecast volatility of gold price with the help of
other precious metals, it is proper to treat the silver and platinum error terms as
regressors. The unique process is similar to generalized linear model in this paper.
Since we consider these three kind of precious metals, price data are correlated with

5
each other, we could add the terms we choose of the other two metals to the original
GARCH (1,1) Model based on its popularity and applicability. The new univariate
GARCH Model with extra error terms behaves like this,

rt  E (rt | t 1 )   t

 t | t 1 ~ N (0, ht )

hgold,t    0 gold

2
,t 1  hgold,t 1  1 silver ,t 1   2 platinum,t 1
2 2

Here, t 1 stands for a new information set which contains not only gold

information during time period t-1, but also both silver and platinum information
during that time.

The first equation for the returns and distribution of the residuals are proper to all

three series. 1 and  2 here stand for the special parameters of extra terms.

For a certain pattern of the univariate standard GARCH model case above, the
first-order model is the most popular EGARCH model in practice as well. Then it
would closely resemble the the situation of standard GARCH(1,1) model, the error
terms of silver and platinum return prices are added up to original EGARCH(1,1)
model respectively.

2.4 Distributions of Error Terms

2.4.1 Normal Distribution

Normal distribution is one of the most important and common distribution in

statistical analysis which is also known as Gaussian distribution.

6
The probability density function of normal distribution is presented as,

( t   )2
1
f ( t )  e 2 ht

2ht

where  stands for the mean and ht stands for the variance.

Normal distribution is a continuous and symmetric distribution with a bell-shape

curve. When   1, ht  0 , the distribution is considered as standard normal

distribution, interpreted as N(0,1).

2.4.2 Student-t Distribution

Student-t distribution, also known as t distribution is proposed by Gosset (1908). The

little difference to normal distribution is found by him based on a large number of
empirical applications. t distribution is also a symmetric, continuous and bell-shaped
distribution, all like normal distribution only with a fatter tail which is perfectly
corresponded to the financial time series. Estimation GARCH models with student-t
distribution were first proposed by Bollerslev (1987).

Let the conditional distribution of rt , t  1,..., T , be t distributed with variance ht|t 1

and degrees of freedom  .

rt  E (rt | t 1 )   t

 t | t 1 ~ f ( t | t 1 )

f ( t | t 1 ) denotes the conditional density function for  t

7
  1
( ) 1 1
f ( t | t 1 )  2
 (  2)ht  2
( ) (1  ( t
))( 1) / 2
2 ht (  2)

Where  is considered as a number of degree of freedom, (z ) is gamma function.

When   1 , t distribution turns into standard Cauchy distribution, having no mean.

When  is large, t distribution is could approximate standard normal distribution. ③

2.5 Root Mean Square Error (RMSE)

In this paper, we prefer a criterion called RMSE for goodness of forecasting.

The Root Mean Square Error is also known as Root Mean Square Deviation (RMSD).
It is used to measure the differences between the forecast values calculated from a
model or an estimator and the values actually observed generally.

Here is the function form of RMSE,

E t
2

RMSE  t 1
N

where Et  Yt  Ft , Yt is the actual value at period t, and Ft is the forecasting value

for period t. N here is the number of out-of-sample observations.

Finally, the square root of the average is taken. The criterion of RMSE requires that a
smaller value which reflects a better forecasting the model completed. But for more
accuracy, we cannot definitely judge which model is better with only comparing the
quantities of some criterions. Since there hardly is such significance concluded from
③
More details about t distribution could be seen in Statistical Inference text book 2nd Edition (Casella, Berger,
2002)

8
only criterion comparison.

3. Data

The price of gold is determined through trading in the gold and its derivatives markets,
however a procedure known as the Gold Fixing in London which sprang from
September 1919, provides a daily benchmark price to the industry. The afternoon
fixing was introduced in 1968 to provide a price when US markets are open ④ which is
caused by time differences. The other precious metals use the similar pricing method.
⑤
The data in this paper is downloaded from Kitco via wikiposit.org. All of them are
the afternoon fixing prices.

3.1 Returns

Let { yt } be the financial time series of the daily price of some financial assets. The

return on the tth day is defined as

rt  log( yt )  log( yt 1 )

where rt is defined as return of the tth variable during time period t.

Sometimes the returns are then multiplied by 100 so that they can be treated as
percentage changes in the price making some sense in economics. Furthermore, this
procedure may decrease numerical errors as the raw returns could be very small
numbers and lead to large rounding errors in some calculations, following Cryer and
Chan (2008).

④
"Gold fixing".http://www.wikipedia.org/. More details here <http://en.wikipedia.org/wiki/Gold_fixing>

⑤
The data is downloaded from <http://wikiposit.org/uid?KITCO>

9
3.2 Plots

Figure 1: Plot of original daily data

Figure 1 shows the plot of the original database used in this paper. It is a plot of daily
spot prices including 3 kinds of precious metals, gold, silver and platinum during the
time period Oct. 7th. 1996 to Mar. 11th. 2011. We can see that gold price appears to be
increasing in general trend, meanwhile silver and platinum prices have more severe up
and down than gold price.

Figure 2: Plot of return data

Figure 2 shows the plot of the results of log-difference. The analysis, consist of
estimation and forecast, we will do needs return data instead of the raw price data
through the way that using the first-order log-difference result multiplied by one
hundred of each daily afternoon fixing prices.

10
3.3 Data Analysis

Table 1 presents some summary statistics of the return data. All these three time series
have the same amount of observations. The mean of each is very close to zero, as well
as the median and a not very large standard deviation. Among them, only gold return
data does have a no-negative skewness and each of them has excess kurtosis. The
ARCH-LM test here is a methodology to test for the lag length of ARCH errors using
the Lagrange multiplier test which was proposed by Engle (1982)⑥. This test should
be done before we apply GARCH models to the data. The p-values for this test are all
very small that the null hypothesis, say the dataset has no ARCH effects should be
rejected. Based on this test, it is secured and proper that we can fit this data to a
GARCH model.

Table 1: Summary Statistics

Gold Returns Silver Returns Platinum Returns

Min. -7.97200 -16.08000 -17.28000

Max. 7.00600 18.28000 11.73000

Median 0.01829 0.05153 0.00000

Mean 0.03281804 0.03916584 0.04063242

Std.Dev 1.105005 1.957008 1.593869

Skewness 0.01957904 -0.3763566 -0.5944549

Kurtosis 8.451281 10.75905 13.11895

ARCH-LM Test p-value < 2.2e-16 p-value < 2.2e-16 p-value < 2.2e-16

Note: The data is collected since 1996-10-07 to 2011-03-11 and each series has
3624 observations in total.

⑥
More details about ARCH-LM Test could be seen in Appendix.

11
The Figure 3 demonstrates the autocorrelation function results. As a financial
time series, having the autocorrelation means it can be predictable because
future values depend on current and past value and this dependence could
last for long term. Teräsvirta (2006) said that observations in return series of
financial assets observed at weekly and higher frequencies are not
independent. While observations in these series are uncorrelated or nearly
uncorrelated, the series contain higher-order dependence. This can explain
the autocorrelation function results below precisely.

Figure 3: Autocorrelation Function of Returns, Returns Square and absolute value of

Return data

We can see that gold return data is not correlated; silver and platinum
either. However, return square data shows that both gold and silver data

12
are highly correlated and also have a long-term memory. The return
square data shows autocorrelation since it is formed as square which is
exactly corresponded to the form of conditional variance. The absolute
values of return data turns out higher correlation than both two before.

4. Estimation

We apply univariate GARCH models and EGARCH models to the data for
estimation distinguishingly with different error term distributions.

4.1 GARCH Models

Table 2 below presents the results of estimation of univariate GARCH model, only
from log-likelihood and AIC values of the models we can see that when we both
add silver and platinum error terms, the model has a smaller AIC value and larger
log-likelihood. From the point of view, different distributions make a difference on
estimation indeed. With a student-t distribution, the models all provide smaller AIC
and larger log-likelihood. Besides, the parameter of silver term often turns out to be
very small, sometimes not significant which suggests that we could model the
volatility without silver term.

13
 Table 2: Estimation of Standard GARCH Models with different distributions

+Plat +Plat
standard +Silver +Plat standard +Silver +Plat

+Silver +Silver

Normal distribution t distribution


0.026 0.022 0.026 0.021 0.012 0.012 0.012 0.013

(0.002) (0.002) (0.002) (0.002) (0.003) (0.003) (0.003) (0.003)

0.085 0.082 0.085 0.082 0.072 0.071 0.070 0.071

0
(0.005) (0.005) (0.005) (0.005) (0.009) (0.009) (0.009) (0.009)

0.894 0.895 0.895 0.895 0.922 0.922 0.918 0.924


(0.006) (0.007) (0.007) (0.007) (0.009) (0.010) (0.010) (0.009)

0.003 0.003 -0.001 -0.0005

1 0 0 0 0

(0.001) (0.001) (0.001) (0.0006)

-0.00028 -0.0001 0.001 0.002

2 0 0 0 0

(0.001) (0.001) (0.001) (0.001)

Log-likeli -4895 -4889 -4894 -4889 -4740 -4737 -4881 -4738

hood

AIC 2.781 2.780 2.782 2.779 2.694 2.694 2.713 2.694

Notes: 1 in the model equals to zero suggests that there is no platinum error term in it, 2 stands for

the parameter of silver error term. Both of them equal to zero means standard GARCH process.

14
4.2 EGARCH Models

Table 3: Estimation of Univariate EGARCH models

+plat +plati
standard +Silver +Plat standard +Silver +Plat

+silver +silver

Normal Distribution t distribution

 -0.124 -0.130 -0.131 -0.128 -0.109 -0.114 -0.113 -0.115

(0.007) (0.008) (0.008) (0.008) (0.012) (0.013) (0.012) （0.012）

0.169 0.158 0.167 0.162 0.150 0.146 0.146 0.153

0
(0.010) (0.010) (0.010) (0.010) (0.016) (0.017) (0.016) (0.0167

 0.052 0.061 0.057 0.058 0.042 0.048 0.047 0.045

(0.006) (0.006) (0.006) (0.006) (0.010) (0.011) (0.011) (0.011)

0.970 0.961 0.962 0.965 0.986 0.978 0.979 0.982



(0.002) (0.004) (0.003) (0.003) (0.003) (0.005) (0.004) (0.004)

0.001 0.002 0.001 0.002

1 0 0 0 0

(0.001) (0.001) (0.001) (0.001)

0.003 0.003 0.001 0.001

2 0 0 0 0

(0.001) (0.001) (0.0001) (0.001)

Log-likeli -4897 -4884 -4891 -4886 -4734 -4731 -4731 -4732

hood

AIC 2.783 2.778 2.781 2.778 2.691 2.691 2.691 2.692

Notes:  here is the parameter in the g ( Zt  j ) part, the same as the  in the

g (Zt  j )  Zt  j   (| Zt  j |  E(| Zt  j |)), j  1,..., q .

Following the same pattern of section 4.1, we can get the estimation results shown in

15
Table 3. Based on AIC and log-likelihood, we can get a similar result as univariate
standard GARCH models. When adding both platinum and silver return's error terms
since it has a smaller AIC value and larger log-likelihood. Only under the
assumption with t distribution, adding silver error term makes the model change a
little bit; it has a smaller AIC value.

4.3 Comparison

Compared to the univariate GARCH models, the average of all the six univariate
EGARCH models gives a better performance in both log-likelihood and AIC value.

With different distribution assumption, although the sample size is large enough that
we could approximate student-t distribution to standard normal distribution, the
estimation still is influenced by distribution difference.

Among all these models, the model based on EGARCH(1,1) model with silver terms
under the student-t distribution has both the smallest AIC and largest log-likelihood
value.

5. Forecast

Besides the estimation comparison of models, we can tell the goodness of fitting by
using that model forecasting and then compared to some existing data, such as an
100-out-of-sample forecasting. Through that way, as mentioned before, RMSE is a
quite useful tool and has more exactness.

In this paper, we complete forecasting following this method:

First of all, there are 3624 observations in total in the dataset. Now we reserve the last
100 observations in order to compare to the forecasting value. Then we should regard
the front 3524 values as a “window” which has a certain length and 3524 observations.
Secondly, we estimate the model with this “window”, after that we can do a one-step

16
forecast of volatility. Take the univariate standard GARCH as an example of one-step
forecast,

hgold,t    0 gold

2
,t 1  hgold,t 1  1 silver ,t 1   2 platinum,t 1
2 2

hgold,t 1    0 gold

2
,t  hgold,t  1 silver ,t   2 platinum,t
2 2

Next, we repeat the first step except for creating a new window from 3524
observations to 3525 observations. The “window” gets wider and wider. We can keep
moving forward by repeating this two steps one-hundred times and then we can get
100 forecast values. Eventually, we can calculate RMSEs of all the values from all the
models and compare them.

First, we compare RMSEs of GARCH models.

Table 4: The results of RMSE of GARCH models

Standard +plat +silver +plat Standard +plat +silver +plat

+silver +silver

Normal Distribution Student-t Distribution

RMSE* 16.876 16.604 16.880 16.867 16.857 16.840 16.833 16.852

Note: RMSE is short for Root Mean Square Error.

From Table 4, adding platinum and silver error terms at the same time seems has a
much better result than every model else, meanwhile the group with t distribution
makes better performance than the other. The most important thing is that extra error
terms actually can help with forecasting gold price volatility since most of them turn
out a better result than using standard GARCH models without all the regressors no
matter what distribution we assume.

17
Next, we compare the results of EGARCH models. The results in Table 5 do not have
a spectacular better number among all. But in both normal distributions and t
distribution, adding only silver term shows its efficiency in helping forecasting gold
volatility. Similarly the group with t distributions performs better. Also same as
GARCH models, most models with extra terms have smaller RMSE.

Compared the results from the view of the kinds of model, by average, EGARCH
models provide a better final result in forecasting than GARCH models no matter
with or without regressors, especially excellent with t distribution.

Table 5: The results of RMSE of EGARCH models

Standard +plat +silver +plat Standard +plat +silver +plat

+silver +silver

Normal Distribution Student-t Distribution

RMSE* 16.773 16.792 16.764 16.826 16.715 16.706 16.700 16.728

Note: RMSE is short for Root Mean Square Error.

6. Conclusion

In summary, this paper aims to forecast gold prices under the help of other precious
metals such as silver and platinum. Among all the results we get, for models,
EGARCH models seem to be more efficient in forecasting volatility; for distributions,
unique properties of t distribution makes it more appropriate than normal distribution
in forcasting.

The result actually confirms a hypothesis we did not mention much, whether the
precious metals could help with forecasting volatility of gold price or not. Now we
can know it for sure that adding silver and platinum error terms makes GARCH

18
models better. The extra regressors increase the accuracy of the models in forecasting.

But there are still existing limitations.

First we assume all the estimation and forecasting are processed under the assumption
that the error term denotes a symmetric distribution such as normal distribution which
is not that enough for real situation.

Second, we consider all the data do not contain any breaks, however, after the
world-wide financial crisis, there may be some structural breaks during that time. This
needs many more steps of analysis before we applied the data to the models.

There have been a lot empirical applications of analyzing financial time series, they
seem to be the same kind of financial data and all applied to GARCH family models,
however the combination of different data and different point of views leads to totally
different results. In this paper, it is too complicated to explain that how and why the
silver and platinum could help with modeling and forecasting this way. We can dig it
deeper and deeper and find more details to improve the whole estimation and forecast
system of GARCH family models.

19
Reference

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

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

Bollerslev, T., (1987), A Conditionally Heteroskedastic Time Series Model for

Speculative Prices and Rates of Return, The Review of Economics and Statistics, Vol.
69, No. 3, pp. 542-547

Cryer J.D, Chan, (2008), Times Series Analysis with Applications in R, 2 nd ed.,

Springer Science Business Media, LLC

Casella, Berger, (2001) Statistical Inference text book 2nd Edition, Duxbury Press

Ding, Z., Engle, R.F., & Granger, C.W.J., (1993). A long memory property of stock
market returns and a new model. Journal of Empirical Finance, 1, pp.83-106.

Engle R.F. (1982). Autoregressive Conditional Heteroscedasticity with Estimates of

the Variance of United Kingdom Inflation. Econometrica.Vol. 50(4), pp. 987-1007

Gosset W.S., (1908) The Probable Error Of A Mean, Biometrika. 6 (1): 1–25.

Nelson, D. B. (1991). "Conditional heteroskedasticity in asset returns: A new

approach", Econometrica 59: 347-370.

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

Teräsvirta,(2006), An Introduction to Univariate GARCH Models. SSE/EFI Working

Papers, Economics and Finance, No. 646

20
Appendix

ARCH - Lagrange multiplier (LM) test

The ARCH - Lagrange multiplier (LM) test is a methodology to test for the lag length
of ARCH errors, in the other word, The ARCH-LM Test is for testing whether the
series has ARCH effects at all (Engle,1982).

In the ARCH-LM test, there is a regression ,

p
 t2   0  (  i t2 p )  et
i 1

In this regression, 1   2     q  0 is the null hypothesis, meanwhile, the

alternative hypothesis H1 : 1  0 or  2  0 or  2  0 means that not all  i

being not significant at the same time. Besides, the test statistics follows  2

distribution with p degrees of freedom.

21