Quantitative Finance I

Modeling Volatility (Lecture 4)

Winter Semester 2011/2012 by Lukas Vacha and Jozef Barunk
* If viewed in .pdf format - for full functionality use Mathematica 7 notebook (.nb) version of this .pdf

Characteristics of Volatility
Volatility is one of the most important concepts in finance, as it measures the risk of financial assets. Altough volatility is not directly observable, and can be just estimated, it has some important characteristics: (i) volatility clusters - there are periods of high/low volatility (ii) volatility evolves over time continuously (jumps in volatility are rare) (iii) volatility does not diverge to infinity (iv) negative and positive price shocks tend to have different impacts on volatility (leverage effect) (v) heavy-tailed, non-gausian distribution

Historical Volatility
Most straightforward way to measure volatility is to estimate time-series of variance on "rolling samples" For zero-mean variable (let us say return), this would be:
2 2 2 s2 = Irt-1 + rt-2 + ... + rt-q M q t

, where q is the latest observation used. This method may produce abrupt changes in estimates.

Exponential volatility
Alternatively, exponential moving average (EMA) estimator of volatility can be used. It uses all data points, and recent observations carry larger weights. (weights are exponentially decreasing).
2 s2 = H1 - lL rt-1 + ls2 , t t-1

where initial value could be unconditional variance in a historical sample. In most financial applications, l=0.94 is used.

Example on S&P 500 index data (change ticker in the code to get other data)
Note that characteristics of volatility can be observed from returns of S&P 500. Moreover, ACF plot suggests us, that returns are not serially correlated, but ACF plot of squared returns shows, that they are not independent. By modeling volatility, we will attempt to capture this kind of dependece.

QF_I_Lecture4.cdf

Prices
1500 1000 500 0

1980

1990

2000

2010

Returns
0.10 0.05 0.00 -0.05 -0.10 -0.15 -0.20 1980 1990 2000 2010

ACF of r
0.2 0.1 0.0 -0.1 -0.2

10

15

20

25

30

ACF of r2
0.2 0.1 0.0 -0.1 -0.2

10

15

20

25

30

Conditional Standard Deviation estimated by EMA (or EWMA)

S&P standard deviation estimate with EMA Hl=0.9L
0.08

0.06

0.04

0.02

0.00

1980

1990

2000

QF_I_Lecture4.cdf

Conditional Heteroscedastic Models

Log returns of an asset at time is either serially uncorrelated, or with minor lower order serial correlations, but it is dependent. Conditional heteroscedastic models are used for modeling s2 as an volatility estimate t by "allowing" heteroscedasticity (time-variation) and capturing that dependence.

ARCH(1)
The first model that provides a systematic framework for volatility modeling is Autoregressive conditional heteroscedasticity (ARCH) model. ARCH(1) model is most common from ARCH(m) models in financial theory. Basic idea is that the mean-corrected return is serially uncorrelated, but dependent, and that the dependence can be described by a simple function of lagged values: at = st et s2 = a0 + a1 a2 t t-1 Adding the assumption of normality, the model can be expressed in terms of information set available in time t: Ft (Engle 1982), (i.e. et is conditionally normally distributed) yt Ft-1 ~ NH0, ht L 2 ht = a0 + a1 yt-1 Example of other notation: yt = c + et st et ~ NH0, 1L s2 = a0 + a1 s2 , t t-1 where a0 > 0, a1 0. Unconditional mean of at is zero, unconditional variance is a0 H1 - a1 L, process is stationary, if a1 < 1, the fourth moment is finite if 3 a2 < 1. 1 Thus excess kurtosis is positive and the tail distribution of at is heavier than normal distribution.

Kurtosis of ARCH(1) processes

The fourth moment of ARCH(1) process: EAet 4 E =
3 a0 2 H1-a1 L
2

1-a1 2 1-3 a1 2

, with 3 a1 2 < 1 r3

kurtHet L =

EAet 4 E EAet E
2 2

=3

1-a1 2 1-3 a1 2

QF_I_Lecture4.cdf

Example : a1 = 0.1

Example : a1 = 0.4

Example : a1 = 0.5

QF_I_Lecture4.cdf

Example : a1 = 0.9

Examples of ARCH(1) artificial processes

Sample ARCHH1L

ARCHH1L volatility

ACF function of a2 t

PACF function of a2 t

a0 a1

0.82 0.823

New Random Case

10

-5

-10

100

200

300

400

500

Example - Simulation of ARCH process ARCH(m)

at = st et , s2 = a0 + a1 a2 + ... + am a2 , t t-1 t-m where at is mean corrected return at = rt - mt , {et < is i.i.d random variables with zero mean and variance 1, somteimes denoted as h - innovations, a > 0, a 0 for i > 0.

QF_I_Lecture4.cdf

at = st et , s2 = a0 + a1 a2 + ... + am a2 , t t-1 t-m where at is mean corrected return at = rt - mt , {et < is i.i.d random variables with zero mean and variance 1, somteimes denoted as ht - innovations, a0 > 0, ai 0 for i > 0. One can easily observe, that large past squared shocks 9a2 =i=1 imply large conditional variance s2 , or t-i t volatility. Thus under ARCH, large shocks tend to be followed by large shocks - ARCH effect
m

Examples of ARCH(m) artificial processes

Note that you have to input a0 > 0, ai 0 for i > 0. in form 8a0 , a1 ..., am <
Input Parameters of ARCHHmL:
am 80.5, 0.1<

Sample ARCHHmL

ARCHHmL volatility

ACF function of a2 t

PACF function of a2 t

New Random Case

0.4

0.2

0.0

-0.2

-0.4 0 5 10 15 20 25 30

Test for ARCH effects in time series

We use Lagrange Multiplier test (LM test), with the null hypothesis of no ARCH effects:
2 2 et = a0 + Is=1 as et-s M + nt q

H0 : a1 =. .. = as = 0 LM = N.R2 ~ c2 HpL

QF_I_Lecture4.cdf

Forecasting with ARCH

Forecasts of the ARCH can be obtained recursively at origin h, the 1-step ahead forecast of s2 is: h+1 s2 H1L = a0 + a1 e2 + ... + am e2 h h h+1-m l-step ahead forecast will be: s2 HlL = a0 + m ai s2 Hl - iL, h i=1 h where s2 Hl - iL = e2 h h+l-i if l - i 0.

Weaknesses of ARCH Models

ARCH model has also following weaknesses: (i) assumes that positive and negative shocks have the same effect on volatility, but empirical testing shows, that assets respond differently to positive and negative shocks. (ii) model gives us description of the conditional variance, but does not explain the real behavior and source of variance (iii) ARCH models tend to overpredict the volatility as they respond slowly to large isolated shocks.

GARCH models GARCH(m,s)

As ARCH model requires many parameters to adequately describe volatility process of returns, extension is used - Generalized autoregressive conditional heteroscedasticity (GARCH) model. It is assumed, that mean equation is adequately described by ARMA, letting at = rt - mt be the mean-corrected log return, GARCH(m,s) model is described by following equation: at = st et , s2 = a0 + m ai a2 + s b j s2 j , t i=1 t-i j=1 twhere {et < is i.i.d random variables with zero mean and variance 1, a0 > 0, ai 0, b j 0, and Hai + bi L < 1, so unconditional variance of at is finite, whereas its conditional variance s2 evolves i=1 t over time. Note, that GARCH model reduces to a pure ARCH(m) model for s = 0.
maxHm,sL

QF_I_Lecture4.cdf

Example GARCH(m,s) artificial processes

Sample GARCHH1,1L

GARCHH1,1L volatility

ACF function of a2 t

PACF function of a2 t

am bs

80.1, 0.2, 0.3< 80.05, 0.2<

New Random Case

Export Simulated Series

1.5

1.0

0.5

0.0 0 100 200 300 400 500

GARCH(1,1)
GARCH(1,1) is most common for financial applications, as the dependencies are mostly very weak. It takes form of: at = st et , s2 = a0 + ai a2 + b1 s2 , t t-1 t-1 where Ha1 + b2 L < 1. Large past shocks of return and volatility in t-1 gives large shocks to volatility in time t, thus volatility clustering is quite well captured.

Example GARCH(1,1) artificial processes

Note that condtition i=1 Hai + bi L < 1 must be met. Initial values are set to simulate GARCH(1,1) process with a0 = 0.5, a1 = 0.2, b1 = 0.4
maxHm,sL

QF_I_Lecture4.cdf

Sample GARCHH1,1L

GARCHH1,1L volatility

ACF function of a2 t

PACF function of a2 t

a0 a1 b1

0.581 0.667 0.296

New Random Case

Export Simulated Series

10

-5

-10

100

200

300

400

500

GARCH-M model
The return of a security may sometimes depend directly on volatility. To model this, we use GARCH in mean (GARCH-M) model. GARCH(1,1) - M is formalized as: rt = m + cs2 + at t at = st et , s2 = a0 + ai a2 + b1 s2 , t t-1 t-1

where m and c are constant. c is also called risk premium

Example GARCH(1,1)-M artificial processes

Note, that with positive risk premium c, returns are positively skewed, as they are positively related to its past volatility

10

QF_I_Lecture4.cdf

Sample GARCHH1,1L-M

ACF function of a2 t

PACF function of a2 t

risk premium c a0 a1 b1

-0.375 0.5 0.38 0.318

New Random Case

Export Simulated Series

-2

-4

100

200

300

400

500

ARIMA-GARCH Models
Common way to build ARCH model is to remove any linear dependencies in the data (i.e.) by ARMA model - for most series, we remove also sample mean from the data.), and use residuals for testing the ARCH effects. This can be done either using Ljung-Box statistics, or ACF, PACF functions, or by Lagrange multiplier test (LM test). If the statistic is significant, then conditional heteroscedasticity of at is detected, and PACF of a2 is used to determine order of ARCH effect. For GARCH, it is not so easy to determine the t order, but empirical findings can serve again, that in most financial applications, we use lower order models such as GARCH(1,1), GARCH(2,1).

QF_I_Lecture4.cdf

11

Example ARIMA(p,d)-GARCH(1,1) artificial processes

Sample ARIMAHp,d,qL-GARCHH1,1L

ACF

ARIMA@p,q,dD parameters
process is unit-root stationary
10 difference ARHpL parameters MAHqL parameters 80.9, -0.4< 80.1, 0.5<

GARCH@1,1D parameters
a0

-5 a1 -10 0 100 200 300 400 500 New Random Case b1

Export Simulated Series

IGARCH - Integrated GARCH

IGARCH models are unit-root (integrated) GARCH models, their key feature is, that past squared shocks is persistent. IGARCH(1,1) is formalized as: at = st et , s2 = a0 + b1 s2 + H1 - b1 L a2 t t-1 t-1

12

QF_I_Lecture4.cdf

Example IGARCH(1,1) artificial processes

IGARCHH1,1L
Simulated series Simulated Volatility

a0 b1

0.611 0.6

New Random Case

Export Simulated Series

800

600

400

200

0 0 100 200 300 400 500

Stochastic Volatility Model

rt = mt + et , et = gIs2 M ht , t s2 = a0 + a1 s2 + nt , t t-1 where 8ht < ~ iidH0, 1L, 8nt < ~ iidI0, s2 M n

Empirical example

Homework #4
Deadline: Tue 15.10.2010, 3:00 pm
Homework may be returned in class, or sent via email to [email protected]
:] Exercise 1 [:

Estimate ARMA-(G)ARCH model on the real stock market data (use dataset which contains (G)ARCH effect).

QF_I_Lecture4.cdf

13

:] Exercise 1 [:

Estimate ARMA-(G)ARCH model on the real stock market data (use dataset which contains (G)ARCH effect). * Please include your computations program with your results.