Escolar Documentos
Profissional Documentos
Cultura Documentos
GARCH Models
Introduction ARMA models assume a constant volatility In nance, correct specication of volatility is essential ARMA models are used to model the conditional expectation They write Yt as a linear function of the past plus a white noise term
GARCH
10
|change| + 1/2 %
10
10
1970
1975
1980 year
1985
1990
1995
GARCH
30
20
10
20
30
40 1994
1996
1998
2000 year
2002
2004
2006
GARCH
35
30
20
15
10
1994
1996
1998
2000 year
2002
2004
2006
GARCH
GARCH models of nonconstant volatility ARCH = AutoRegressive Conditional Heteroscedasticity heteroscedasticity = non-constant variance homoscedasticity = constant variance
GARCH
ARMA unconditionally homoscedastic conditionally homoscedastic GARCH unconditionally homoscedastic, but conditionally heteroscedastic Unconditional or marginal distribution of Rt means the distribution when none of the other returns are known.
GARCH
Modeling conditional means and variances Idea: If is N (0, 1), and Y = a + b , then E (Y ) = a and Var(Y ) = b2 . general form for the regression of Yt on X1.t , . . . , Xp,t is Yt = f (X1,t , . . . , Xp,t ) + t (1) Frequently, f is linear so that f (X1,t , . . . , Xp,t ) = 0 + 1 X1,t + + p Xp,t . Principle: To model the conditional mean of Yt given X1.t , . . . , Xp,t , write Yt as the conditional mean plus white noise.
GARCH
Let 2 (X1,t , . . . , Xp,t ) be the conditional variance of Yt given X1,t , . . . , Xp,t . Then the model Yt = f (X1,t , . . . , Xp,t ) + (X1,t , . . . , Xp,t )
t
(2)
gives the correct conditional mean and variance. Principle: To allow a nonconstant conditional variance in the model, multiply the white noise term by the conditional standard deviation. This product is added to the conditional mean as in the previous principle. (X1,t , . . . , Xp,t ) must be non-negative since it is a standard deviation
GARCH
ARCH(1) processes Let 1 , 2 , . . . be Gaussian white noise with unit variance, that is, let this process be independent N(0,1). Then E ( t| and Var( t |
t1 , . . .) t 1 , . . . )
= 0, = 1. (3)
GARCH
10
at =
0 + 1 a2 t1 .
(4)
It is required that 0 0 and 1 0 It is also required that 1 < 1 in order for at to be stationary with a nite variance. If 1 = 1 then at is stationary, but its variance is Dene
2 t = Var(at |at1 , . . .)
GARCH
11
0 + 1 a2 t1 . + 1 a2 t1 }.
and
2 = 0 + 1 a2 t t 1 .
GARCH
12
If at1 has an unusually large deviation then the conditional variance of at is larger than usual at is also expected to have an unusually large deviation volatility will propagate since at having a large 2 deviation makes t +1 large so that at+1 will tend to be large.
GARCH
13
The conditional variance tends to revert to the unconditional variance provided that 1 < 1 so that the process is stationary with a nite variance. The unconditional, i.e., marginal, variance of at denoted by a (0)
GARCH
14
(7)
This gives us a (0) = 0 + 1 a (0). This equation has a positive solution if 1 < 1: a (0) = 0 /(1 1 ).
GARCH
15
GARCH
16
GARCH
17
1.4
ARCH(1) case 1
10 k
15
20
Var(at+k |at , . . .) for AR(1) and ARCH(1). (0) = 1 in both cases. For ARCH(1), 1 = .9. Case 1: 2 a2 = 1 . 5 . Case 2: a t t = .5.
GARCH
18
independence implies zero correlation but not vice versa GARCH processes are good examples dependence of the conditional variance on the past is the reason the process is not independent independence of the conditional mean on the past is the reason that the process is uncorrelated
GARCH
19
20
40
60
ARCH(1) 4 2 0 5 2 4 6 10 5
AR(1)/ARCH(1))
20
40
60
15
20
40
60
GARCH
20
ARCH(1)
40
50
0 0
500
50 0
500
Probability
0 20 Data
40
GARCH
21
GARCH
22
Comparison of AR(1) and ARCH(1) AR(1) E (Yt ) = . Et (Yt ) = + (Yt1 ). ARCH(1) E (at ) = 0. Et (at ) = 0.
GARCH
23
ARCH(1) 0 = . 1 1
2 2 t = 0 + 1 a2 t1 . 2 Recall: t = Var(at |at1 , . . .).
GARCH
24
The AR(1)/ARCH(1) model Let at be an ARCH(1) process Suppose that ut = (ut1 ) + at . ut looks like an AR(1) process, except that the noise term is not independent white noise but rather an ARCH(1) process.
GARCH
25
at is not independent white noise but is uncorrelated Therefore, ut has the same ACF as an AR(1) process: u (h) = |h| h. a2 t has the ARCH(1) ACF: a2 (h) = 1
|h|
h.
need to assume that both || < 1 and 1 < 1 in order for u to be stationary with a nite variance
GARCH
26
t =
0 +
i=1
i a2 t i
GARCH
27
t =
0 +
i=1
i a2 ti +
i=1
2 i t i .
very general time series model: at is GARCH(pG , qG ) and at is the noise term in an ARIMA(pA , d, qA ) model
GARCH
28
Heavy-tailed distributions stock returns have heavy-tailed or outlier-prone distributions reason for the outliers may be that the conditional variance is not constant GARCH processes exhibit heavy-tails Example 90% N (0, 1) and 10% N (0, 25) variance of this distribution is (.9)(1) + (.1)(25) = 3.4 standard deviation is 1.844 distribution is MUCH dierent that a N (0, 3.4) distribution
GARCH
29
10
12
Normal plot normal 0.997 0.99 0.98 0.95 0.90 0.75 0.50 0.25 0.10 0.05 0.02 0.01 0.003 2 1 0
Data
Normal plot normal mix 0.997 0.99 0.98 0.95 0.90 0.75 0.50 0.25 0.10 0.05 0.02 0.01 0.003 10 0
Data
Probability
Probability
10
GARCH
30
For a N (0, 2 ) random variable X , P {|X | > x} = 2(1 (x/ )). Therefore, for the normal distribution with variance 3.4, P {|X | > 6} = 2(1 (6/ 3.4)) = .0011.
GARCH
31
For the normal mixture population which has variance 1 with probability .9 and variance 25 with probability .1 we have that P {|X | > 6} = 2{.9(1 (6)) + .1(1 (6/5))} = (.9)(0) + (.1)(.23) = .023. Since .023/.001 21, the normal mixture distribution is 21 times more likely to be in this outlier range than the normal distribution.
GARCH
32
Property Cond. mean Cond. var Cond. distn Marg. mean & var. Marg. distn
GARCH
33
All of the processes are stationary marginal means and variances are constant Gaussian white noise is the baseline process. conditional distribution = marginal distribution conditional means and variances are constant conditional and marginal distributions are normal Gaussian white noise is the source of randomess for the other processes therefore, they all have normal conditional distributions
GARCH
34
Fitting GARCH models Fit to 300 observation from a simulated AR(1)/ARCH(1) Listing of the SAS program for the simulated data
options linesize = 65 ; data arch ; infile c:\courses\or473\sas\garch02.dat ; input y ; run ; title Simulated ARCH(1)/AR(1) data ; proc autoreg ; model y =/nlag = 1 archtest garch=(q=1); run ;
GARCH
35
SAS output
Q and LM Tests for ARCH Disturbances Order 1 2 3 4 5 6 7 8 9 10 11 12 Q 119.7578 137.9967 140.5454 140.6837 140.6925 140.7476 141.0173 141.5401 142.1243 142.6266 142.7506 142.7508 Pr > Q <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 LM 118.6797 129.8491 131.4911 132.1098 132.3810 132.7534 132.7543 132.8874 132.8879 132.9226 133.0153 133.0155 Pr > LM <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 <.0001
GARCH
36
Approx Error t Value 0.3910 0.0266 0.1729 0.1167 1.23 -30.92 6.50 5.98
= .8226 AR parameter: this is +.8226 in our notation close to the true value of 0.8 estimates of the ARCH parameters: 0 = 1.12 (true value = 1) 1 = .70 (true value = .95)
GARCH
37
standard errors of the ARCH parameters are rather large approximate 95% condence interval for 1 is .70 (2)(0.117) = (.446, .934)
GARCH
38
65
70
75
80 year
85
90
95
Residuals when the S&P 500 returns are regressed against the change in the 3-month T-bill rates and the rate of ination.
GARCH
39
This analysis uses RETURNSP = the return on the S&P 500 DR3 = change in the 3-month T-bill rate GPW = the rate of wholesale price ination RETURNSP is regressed on DR3 and GPW (factor model)
GARCH
40
Model RETURNSP = 0 + 1 DR3 + 2 GPW + ut ut is an AR(1)/GARCH(1,1) process Therefore, ut = 1 ut1 + at , at is a GARCH(1,1) process: at = t t where t =
2 0 + 1 a2 + 1 t1 t1 .
(8)
GARCH
41
returnsp = DR3 gpw species the regression model nlag = 1 species the AR(1) structure. garch=(p=1,q=1) species the GARCH(1,1) structure. archtest species that tests of conditional heteroscedasticity be performed
GARCH
42
SAS output The p-values of the Q and LM tests are all very small, less than .0001. Therefore, the errors in the regression model exhibit conditional heteroscedasticity. Ordinary least squares estimates of the regression parameters are:
Variable Intercept DR3 GPW DF 1 1 1 Estimate 0.0120 -0.8293 -0.8550 Standard Error 0.001755 0.3061 0.2349 t Value 6.86 -2.71 -3.64 Approx Pr > |t| <.0001 0.0070 0.0003
GARCH
43
Using residuals from the OLS estimates, the estimated residual autocorrelations are:
Estimates of Autocorrelations Lag 0 1 Covariance 0.00108 0.000253 Correlation 1.000000 0.234934
GARCH
44
Lag 1
Coefficient -0.234934
t Value -5.01
GARCH
45
Notice that these dier slightly from OLS estimates. Since all p-values are small, both independent variables are signicant. However, the Total R-square value is only 0.0551, so the regression has little predictive value.
GARCH
46
Since all p-values are small, all GARCH parameters are signicant. GARCH1 (0.7254) >> ARCH1 (0.1337) reasonably long persistence of volatility.
GARCH
47
I-GARCH models I-GARCH or integrated GARCH processes designed to model persistent changes in volatility A GARCH(p, q ) process is stationary with a nite variance if
q p
i +
i=1 i=1
i < 1.
GARCH
48
i +
i=1 i=1
i = 1 .
I-GARCH processes are either non-stationary or have an innite variance. Here are some simulations of ARCH(1) processes:
GARCH
49
= .9
1
0.5
1.5
2 1 = 1
2.5
3.5 x 10
4
4
100
100 0 5 x 10
4
0.5
1.5
2 1 = 1.8
2.5
3.5 x 10
4
4
5 0
0.5
1.5
2.5
3.5 x 10
4
4
GARCH
50
= .9
1
0.999 0.997 0.99 0.98 0.95 0.90 0.75 0.50 0.25 0.10 0.05 0.02 0.01 0.003 0.001 100 80 60 40 20 0 20 40
Probability
1 = 1 0.999 0.997 0.99 0.98 0.95 0.90 0.75 0.50 0.25 0.10 0.05 0.02 0.01 0.003 0.001 50 0 50 1 = 1.8 0.999 0.997 0.99 0.98 0.95 0.90 0.75 0.50 0.25 0.10 0.05 0.02 0.01 0.003 0.001 4 3 2 1 0 1 2 3 4 x 10
4
Probability
100
150
Probability
GARCH
51
Comments on the gures all three processes do revert to their mean, 0 larger the value of 1 the more the volatility comes in sharp bursts processes with 1 = .9 and 1 = 1 looks similar none of the processes in the gure show much persistence of higher volatility to model persistence of higher volatility, one needs an I-GARCH(p, q ) process with q 1 Next gure shows simulations from I-GARCH(1,1) processes
GARCH
52
30 25 20 15 10 5 0 0 1000 2000
30 25
200
40 35
30 25 20 15 10 5 0 1000 2000
1 = 0.95, GARCH(1,1) 30 20 10
1 = 0.4, GARCH(1,1) 40
20 100 0 0 100
50
0 10 20 20 200 0 1000 40 2000 0 1000 300 2000 0 1000 100 2000 0 1000 2000 50
GARCH
53
To t I-GARCH in SAS:
proc autoreg ; model returnsp =/nlag = 1 garch=(p=1,q=1,type=integrated); run ;
The default value of type is nonneg which only constrains the GARCH coecients to be non-negative. type=integrated in addition imposes the sum-to-one constraint of the I-GARCH model
GARCH
54
What does innite variance mean? let X be a random variable with density fX the expectation of X is
xfX (x)dx
GARCH
55
If
xfX (x)dx =
(9)
and
0
xfX (x)dx =
(10)
then the expectation is, formally, + not dened if both integrals are nite, then the expectation is the sum of these two integrals
GARCH
56
Exercise: fX (x) = 1/4 if |x| < 1 and fX (x) = 1/(4x2 ) if |x| 1 fX is a density since
fX (x)dx = 1
xfX (x)dx =
and
0
xfX (x)dx =
GARCH
57
What are the implications of having no expectation? assume sample of iid sample from fX law of large numbers sample mean will converge to the expectation law of large numbers doesnt apply if expectation is not dened there is no point to which the sample mean can converge it will just wander without converging
GARCH
58
1 = .9 0.5
0.5
1 0 4 2 0 2 4 6 0 15 10 5 0 5 0
0.5
1.5
2 1 = 1
2.5
3.5 x 10
4
4
0.5
1.5
2 1 = 1.8
2.5
3.5 x 10
4
4
0.5
1.5
2.5
3.5 x 10
4
4
GARCH
59
What are the implications of having innite variance now suppose that the expectation of X exists and equals X the variance
(x X )2 fX (x)dx
if this integral is + then the variance is innite law of large numbers sample variance will converge to the variance variance of X is innity the sample variance will converge to innity
GARCH
60
= .9
1
8 6 4 2 0 0 150
0.5
1.5
2 1 = 1
2.5
3.5 x 10
4
4
100
50
0 0 4 3 2 1 0 0 x 10
5
0.5
1.5
2 1 = 1.8
2.5
3.5 x 10
4
4
0.5
1.5
2.5
3.5 x 10
4
4
GARCH
61
GARCH-M processes if we t a regression model with GARCH errors could use the conditional standard deviation (t ) as one of the regression variables when the dependent variable is a return the market demands a higher risk premium for higher risk so higher conditional variability could cause higher returns
GARCH
62
or for I-GARCH-M
proc autoreg ; model returnsp =/nlag = 1 garch=(p=1,q=1,mean,type=integrated); run ;
GARCH
63
GARCH-M example: S&P 500 GARCH(1,1)-M was t in SAS is the regression coecient for t = .5150 standard error = .3695 t-value = 1.39 p-value = .1633
GARCH
64
since p-value = .1633 could accept the null hypothesis that = 0 no strong evidence that there are higher returns during times of higher volatility. volatility of S&P 500 is market risk so this is somewhat surprising (think of CAPM) is may be that the eect is small but not 0 ( positive, after all) AIC criterion does select the GARCH-M model
GARCH
65
E-GARCH E-GARCH models are used to model the leverage eect prices become more volatile as prices decrease E-GARCH, model is
q p
log(t ) = 0 +
i=1
1 g (
ti )
+
i=1
i log(ti ),
GARCH
66
> 0,
< 0,
< 0 so that 0 < + < typically, 1 < = .7 in the S&P 500 example E (| t |) = 2/ = .7979 (good calculus exercise)
GARCH
67
= 0.7
6 5 4 6 5 4
=0
g( t )
3 2 1 0
g( t )
2 0 2 4
3 2 1 0
1 4
t = 0.7
1 4
t = 1
6 5 4
6 5 4
g( t )
3 2 1 0
g( t )
2 0
t
3 2 1 0
1 4
1 4
0
t
The g function for the S&P 500 data (top left panel) and several other values of .
GARCH
68
SAS ts the E-GARCH model xed as 1 estimated E-GARCH model is specied by using type=exp as in:
proc autoreg ; model returnsp =/nlag = 1 garch=(p=1,q=1,mean,type=exp); run ;
GARCH
69
Back to the S&P 500 example SAS can t six dierent AR(1)/GARCH(1,1) models type = integrated, exp, or nonneg GARCH-in-mean eect can be included or not following table contains the AIC statistics models ordered by AIC (best tting to worse)
GARCH
70
AIC statistics for six AR(1)/GARCH(1,1) models t to the S&P 500 returns data. AIC is change in AIC between a given model and E-GARCH-M.
GARCH
71
AR(2) and E-GARCH(1,2)-M, E-GARCH(2,1)-M, and E-GARCH(2,2)-M models were tried none of these lowered AIC none had all parameters signicant at p = .1
GARCH
72
GARCH
73
Variable Intercept DR3 GPW AR1 EARCH0 EARCH1 EGARCH1 THETA DELTA
DF 1 1 1 1 1 1 1 1 1
Estimate -0.003791 -1.2062 -0.6456 -0.2376 -1.2400 0.2520 0.8220 -0.6940 0.5067
Standard Error 0.0102 0.3044 0.2153 0.0592 0.4251 0.0691 0.0606 0.2646 0.3511
t Value -0.37 -3.96 -3.00 -4.01 -2.92 3.65 13.55 -2.62 1.44
Approx Pr > |t| 0.7095 <.0001 0.0027 <.0001 0.0035 0.0003 <.0001 0.0087 0.1490
GARCH
74
The GARCH zoo Heres a sample of other GARCH models mentioned in Bollerslev, Engle, and Nelson (1994): QARCH = quadratic ARCH TARCH = threshold ARCH STARCH = structural ARCH SWARCH = switching ARCH QTARCH = quantitative threshold ARCH vector ARCH diagonal ARCH factor ARCH
GARCH
75
GARCH Models in Finance Remember the problem of implied volatility it depended on K and T ! Black-Scholes assumes a constant variance But GARCH eects are common so the Black-Scholes model is not adequate Having a volatility function (smile) is a quick x but not logical
GARCH
76
0.04 implied volatility 0.035 0.03 0.025 0.02 0.015 150 100 50 maturity 0 35 45 40 exercise price 55 50
GARCH
77
Options can be priced assuming the log-returns are a GARCH process (rather than a random walk) Multinomial (not binomial) tree to have dierent levels of volatility Need to keep track of price and conditional variance
GARCH
78
Ritchken and Trevor use an NGARCH (nonlinear asymmetric GARCH) model: log(St+1 /St ) = r + ht ht /2 + ht t+1
ht+1 = 0 + 1 ht + 2 ht (t+1 c)2 where t is WhiteNoise(0, 1). c = 0 is an ordinary GARCH model is a risk premium Under the risk-neutral (martingale) measure: log(St+1 /St ) = (r ht /2) + ht+1 = 0 + 1 ht + 2 ht ( where c = c + and
t
ht
t+1 ,
2 c ), t+1
is WhiteNoise(0, 1).
GARCH
79
Now there are ve unknown parameters: h0 1 , 2 , and 3 c These parameters are estimated by nonlinear least-squares Implied GARCH parameters
GARCH
80
Volatility smile is explained as due to GARCH eects: nonconstant variance nonnormal marginal distribution
GARCH
81
From Chapter 7, Bank of Volatility, of When Genius Failed by Roger Lowenstein: Early in 1998, Long-Term began to short large amounts of equity volatility. This simple trade, second nature to Rosenfeld and David Modest, would be indecipherable to 999 out of 1,000 Americans. Equity vol comes straight from the Black-Scholes model.
GARCH
82
The stock market, for instance, typically varies by about 15 percent to 20 percent a year. And when the model told them that the markets were mispricing equity vol, they were willing to bet the rm on it. The options market was anticipating volatility in the stock market of roughly 20 percent. Long-term viewed this as incorrect ... Thus, it gured that options prices would sooner or later fall.
GARCH
83
Long-Term began to short options on the S&P 500 are similar European options. They were selling volatility. In fact, it sold insurance (options) both waysagainst a sharp downturn and against a sharp rise.
GARCH
84
1996
1998
2000
2002
GARCH
85
AR(1)/EGARCH(1,1) model 0.5 0.45 S&P500, conditional SD, annualized 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 1997
1998
1999
2000
2001
Annualized SD =
GARCH
86
data capm ; set Sasuser.capm ; logR_sp500 = dif(log(close_sp500)) ; logR_msft = dif(log(close_msft)) ; run ; proc gplot ; plot logR_sp500*date ; run ; proc autoreg ; model logR_sp500 = /nlag=1 garch=(p=1,q=1,type=exp) ; output out=outdata cev=cev ; run ; proc gplot ; plot cev*date ; run ;
GARCH
87
nlag 1 1 1 1 1 1 0 2 1 2 2 1
p 2 1 1 1 2 1 1 2 2 2 3 3
q 2 2 2 2 1 1 1 2 2 3 2 2
E-GARCH no no yes yes yes yes yes yes yes yes yes yes
M-GARCH no no no yes no no no no no no no no
AIC 14,366 15,002 15,105 15,103 15,105 15,107 14,366 15,112 15,104 15,100 15,100 15,101
GARCH
88
In fact, it sold insurance (options) both waysagainst a sharp downturn and against a sharp rise.
110 stock price 100
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
GARCH
89
The Greeks C (S, T, t, K, , r) = price of an option = C (S, T, t, K, , r) S = C (S, T, t, K, , r) t C (S, T, t, K, , r) R= r V= C (S, T, t, K, , r) Delta Theta Rho Vega
GARCH
90
Put-call parity Put and call prices with same K and T are related: P (S, T, t, K, , r) = C (S, T, t, K, , r) + er(T t) K S. Therefore, the call and put have the same vegas P (S, T, t, K, , r) = C (S, T, t, K, , r) but dierence deltas P (S, T, t, K, , r) = C (S, T, t, K, , r) 1 S S 0 < (call) = (d1 ) < 1 1 < (put) = (d1 ) 1 < 0
GARCH
91
From previous slide: 1 < (put) = (d1 ) 1 < 0 Suppose we buy N1 call options and N2 put options. Delta of the portfolio is N1 (d1 ) + N2 ((d1 ) 1) which is zero if N1 1 (d1 ) = N2 (d1 )
Hedging is possible with a put and call with dierent K and T each then has its own value of d1
GARCH
92
Selling equity vol was very clever, but as Lowenstein remarks: This wasso unlike the partners credorank speculation.