Arch Models

Most investors dislike risk taking and require a premium for holding assets with risky payos. The variance of an asset has been used to measure risk, and split the risk into a company specic component, which is diversiable, and a market component which cannot be diversied. This measure of the unconditional volatility does not recognize that there may be predictable patterns in stock market volatility. We will analyze models of conditional (on information at time t-1) volatility. These type of models have the implication for nance that investors can predict the risk. This type of models successfully characterize the fact that stock prices seem to go through long periods of high and long periods of low volatility. The fact that market participants may predict volatility has important implications The most important is that for periods where the investor has forecasted prices to be very volatile, she should either exit the market or require a large premium as a compensation for bearing an unusual high risk. Empirical Regularities of Asset Returns. i Thick Tails Asset returns tend to be leptokurtotic . The documentation of this empirical regularity is presented in Mandelbrot (1965). ii Volatility Clustering " ... large changes tend to be followed by large changes, of either sign and small changes tend to be followed by small changes " iii Leverage Eects The so-called "leverage eect" rst noted by Black(1976) refers to the tendency for stock prices to be negatively correlated with changes in stock volatility. A rm with debt and equity outstanding typically becomes more highly leveraged when the value of the rm falls. This raises the equity return volatility. iv) Non-Trading Periods Information that accumulates when nancial markets are closed is reected in prices after the markets reopen. If for example, information accumulates at a constant rate over calendar time, then the variance of the returns over the period from Friday close to the Monday close should be three times the variance from the Monday close to the Tuesday close. v) Forecastable Events

Patell and Wolfson (1979,1981) show that individual rms stock returns volatility is high around earning announcements.

Introduction: Conditional and Unconditional moments Before presenting the alternative Arch type models, we will briey review the dierence between conditional and unconditional moments. Let us assume that yt follows a random walk, i.e. yt = yt1 + t Then yt = y0 +
t X i=1

Unconditional Moments The unconditional mean and variance are; E(yt ) = y0 V (yt ) = t 2 A RW has a constant unconditional mean but a time varying unconditional variance. Conditional Moments The conditional mean and variance are; E(yt |yt1 ) = yt1 V (yt |yt1 ) = E(yt E(yt |yt1 ))2 = E(yt1 + t E(yt |yt1 ))2 = 2 A RW has a constant unconditional mean but a time varying unconditional variance. So while the unconditional variance of a random walk model tends to innite as t increase, the conditional variance is constant.

Univariate Parametric Models Arch Models In the linear Arch(q) model originally introduced by Engle(1982), the time varying conditional variance is postulated to be a linear function of the past q squared innovations. 2 = + t
q X i=1

i 2 = w + (L)2 ti t1

A sucient condition for the conditional variance to be positive is that the parameters of the model satisfy the following constraints; > 0 and 1 > 0, 2 > 0, , ....q > 0 Dening t 2 2 , the ARCH(q) model can be re-written as t t 2 = + (L)2 + t t t1 ( Notice that 2 = E(2 |2 , 2 , .....)) Since Et1 ( t ) = 0, the model corret t t1 t2 sponds to an AR(q) model for the squared innovations, 2 . Then, the process t is covariance stationary if and only if the sum of the positive autoregressive parameters is less than one, in which case the unconditional variance equals V ar(2 ) = /(1 1 2 ... q ). t Even though t s are serially uncorrelated they are clearly not independent through time. In accordance with the stylized facts for assets returns discussed above, there is a tendency for large (small) absolute values of the process to be followed by other large (small) values of unpredictable sign. The ARCH(1) Model Constant unconditional Variance but non-constant conditional Variance. Some useful statistical results are given below for the simplest ARCH(1) model, which is identical to the one used by Engle (1982). The main result is that this simple model exhibits constant unconditional variance but nonconstant conditional variance. Consider the following model yt = + t t = ut ( + 2 )1/2 , ut IIN (0, 1), > 0, > 0 t1 (NOTICE that ( + 2 )1/2 is the conditional standard deviation, t dened t1 as (E(2 |2 , 2 , .....))1/2 . t t1 t2 i) The conditional expectation of t is equal to zero E(t |t1 ) = E(ut |t1 )( + 2 )1/2 = 0 t1 3

Notice that E(ut |t1 ) = E(ut ) = 0, since ut ~IIN(0,1) ii) The conditional variance is given by the following formula V ar(t |t1 ) = E(u2 |t1 )( + 2 ) = ( + 2 ) t t1 t1 Notice that E(u2 |t1 ) = E(u2 ) = 1, since ut ~IIN(0,1) t t Then the conditional mean and variance of yt are given by the following formulae; E(yt |yt1 ) = V ar(yt |yt1 ) = ( + 2 ) t1 Then, the conditional variance of yt is time varying. On the other hand it can be easily seen that the unconditional variance is time invariant whenever 2 t is stationary, i.e. V (yt ) = V (t ) = /(1 ) whenever the process is stationary. ( since V (t ) = E(2 ) = E( + 2 ) = + E(2 ) ) t t1 t1 First Order Autoregressive Process with ARCH eects. For more complicated models such as AR(1)-ARCH(1), we obtain similar results provided that the process for y is stationary, i.e. that the autoregressive parameter is smaller than one in absolute value. Assume the following rst order autoregressive process yt = yt1 + t where t = ut ( + 2 ) and ut ~IIN(0,1) , >0 , > 0 t1 then i) The conditional expectation of t is equal to zero E(t |t1 ) = E(u2 |t1 )( + 2 ) = 0 since E(ut |t1 ) = E(ut ) = 0 t t1 ii) The conditional variance is given by the following formula

V ar(t |t1 ) = E(u2 |t1 )( + 2 ) = ( + 2 ) t t1 t1 since E(u2 |t1 ) = E(ut ) = 1 t Then the conditional mean and variance of yt are given by the following formulae; E(yt |yt1 ) = yt1 V ar(yt |yt1 ) = ( + 2 ) t1 To nd the unconditional variance of yt we recall the following property for the variance; V ar(yt ) = E(V ar(yt |yt1 )) + V ar(E(yt |yt1 )) then i) E(V ar(yt |yt1 )) = E( + 2 ) = + E(2 ) = + V ar(t1 ) t1 t1 ii) V ar(E(yt |yt1 )) = 2 V ar(yt1 ) Then if the process is covariance stationary we have + V ar(t1 ) (1 2 ) (1 )(1 2 )

V ar(yt ) = = (Since V ar(t1 ) = /((1 ))

GARCH Models In empirical applications it is often dicult to estimate models with large number of parameters, say ARCH(q). To circumvent this problem Bollerslev (1986) proposed the Generalized ARCH or GARCH(p, q) model,
q X i=1 p X i=1

2 t

= +

i 2 + ti

i 2 ti

= + (L)2 + (L) 2 t1 t1 A sucient condition for the conditional variance in the GARCH(p, q) model to be well dened is that all the coecients in the innite order linear ARCH 5

model must be positive. Provided that (L) and (L) have no common roots and that the roots of the polynomial in L, (1 (L)) = 0 lie outside the unit circle, this positive constraint is satised, if and only if, the coecients of the innite power series expansion for (L)/(1 (L)) are non-negative. Rearranging the GARCH(p, q) model by dening t 2 2 , it follows t t that 2 = + ((L) + (L))2 (L) t1 + t t t1 which denes an ARMA( Max(p, q),p) model for 2 t By standard arguments, the model is covariance stationary if and only if all the roots of (1 (L) (L)) lie outside the unit circle. If all the coecients are positive, this is equivalent to the sum or the autoregressive coecients being smaller that 1. The analogy to ARMA class of models also allows for the use of standard time series techniques in the identication of the orders of p and q. In most empirical applications with nitely sampled data, the simple GARCH(1, 1) is found to provide a fair description of the data. Persistence and Stationarity Using the GARCH(1,1) model it is easy to construct multi period forecasts of volatility. When + < 1, the unconditional variance of t+1 is /(1). If we re-write the following GARCH(1,1) as 2 t = + 2 + 2 t1 t1 = + (2 2 ) + ( + ) 2 t1 t1 t1

The coecient measures the extent to which a volatility shock today feeds through into next periods volatility, while ( + ) measures the rate at which this eect dies over time. Recursively substituting and using the law of iterated expectations, the conditional expectation of volatility j periods ahead is, Et [ 2 ] = ( + )j ( 2 /(1 )) + w/(1 ) t+j t The multi period volatility forecast reverts to its unconditional mean at rate ( + ). IGARCH Models Integrated GARCH models are processes were the autorregresive part of the square residuals has a unit root, i.e., ( + ) = 1. For this case the conditional expectation of the volatility j periods ahead is Et [ 2 ] = 2 + j. t+j t 6

This process looks very much as a random walk with drift . Then, if t follows an IGARCH process the unconditional variance does not exist and therefore it is not covariance stationary. Nelson(1990) shows that the analogy with the random walk process should be treated with caution since the IGARCH process is not covariance stationary but it may be proved to be strictly stationary. For example when = 0, Et [2 ] = 2 , so volatility is a martingale. But t t+j the volatility remains bounded , since it cannot be negative, and then using the fact that a bounded martingale must converge, we can show that it converges to zero, a degenerate distribution. Despite the fact that it seems to be an empirical regularity that volatility is IGARCH (many estimated models have coecients that sum near 1) we regard this type of process as unlikely. (see section on structural breaks and GARCH models) EGARCH Models Even if GARCH models successfully capture thick tailed returns, and volatility clustering, are not well suited to capture the "leverage eect" since the conditional variance is a function only of the magnitudes of the lagged residuals and not their signs In the exponential GARCH (EGARCH) model of Nelson (1991) 2 depends t on both the size and the sign of lagged residuals. EGARCH(1,1) Models ln 2 = 0 + 1 ln 2 + 0 ([|t1 /t1 | (2/)1/2 ] + (t1 / t1 )) t t1 Obviously the EGARCH model always produces a positive conditional variance 2 for any choice of 0 , 1 , 0 and so that no restrictions need to be placed t on these coecients (except | 1 | < 1). Because of the use of both |t / t | and (t / t ), 2 will also be non-symmetric in t and, for negative , will exhibit t higher volatility for large negative t . Other ARCH Specications Glosten, Jagannathan and Runkle (1989) proposed the following specication: t = t t , where vt is iid. 2 = 0 + 1 2 + 1 2 + 2 2 It1 , t t1 t1 t1 where, It1 = 1 if t1 0 and It1 = 0 if t1 < 0. The non-negativity condition is satised provided that all the parameters are positive. If leverage eects do exist, 2 < 0.

Additional Explanatory Variables. It is straightforward to add other explanatory variables to a GARCH specication. Glosten, Jagannathan and Runkle(1993) add a short-term nominal interest rate to various GARCH models and show that it has a signicant positive eect on stock market volatility. 2 = + 2 + 2 + Xt1 , t t1 t1 where X is any positive variable. GARCH in Mean Models Many theories in nance assume some kind of relationship between the mean of a return and its variance . A way to take this into account is to explicitly write the returns as a function of the conditional variance or, in other words, to include the conditional variance as another regressor. GARCH in Mean Models allow for the conditional variance to have mean eects. Most of the time this conditional variance term will have the interpretation of a time varying risk premium. Consider the following model. yt = xt + 2 + t t and 2 = + (L)2 + (L)2 t t1 t1 Consistent estimation of and is dependent on the correct specication of the entire model. The estimation of GARCH in Mean type of models is numerically unstable and many empirical applications have used ARCH-M type of models which are easier to estimate. An ARCH in Mean model simply models the conditional variance as an ARCH model instead of modeling as GARCH, i.e. 2 = + (L)2 t t1 Example of an ARCH(1)-M Consider a simple version of the above model. yt = 2 + t t where t = vt t vt ~N(0,1) 2 = w + + 2 t t1 Then yt may be expressed as yt = ( + 2 ) + t t1 8

Then the expected value of yt is E(yt ) = + E(2 ) t1 and using that E(2 ) = /(1 )then t1 E(yt ) = + /(1 ) Which can be viewed in nance models as the unconditional expected return for holding a risky asset. Testing for Arch Before attempting to estimate a GARCH model you should rst check if there are ARCH eects in the residuals of the model. Clearly we should not explicitly model (and estimate) the conditional volatility of series as GARCH when there are not signs of Arch eects. The original Lagrange Multiplier test for ARCH proposed by Engle (1982) is very simple to compute, and relatively easy to derive. Under the null hypothesis it is assumed that the model is say an AR(p) model yt = 1 yt1 + 2 yt2 + ... + p ytp + t where t is a Gaussian white noise process, t |It1 ~N(0, 2 )where It is the information set. The Alternative hypothesis is that the errors are ARCH(q). The test for ARCH(q) eect simply consists on regressing 2 = 1 2 + 2 2 + ... + q 2 + t t t1 t2 tq Under the null hypothesis that 1 = 2 = ... = q = 0, and T R2 is asymptotically distributed (q), where T is the number of observations. While this is the most widely used test we should be cautious in interpreting the results. If the model is misspecied it is quite likely to reject the null hypothesis simply because most of the time serial correlation in the residuals will induce serial correlation in the squared residuals. Structural Breaks and ARCH eects It has been shown ( Diebold (1986) ) that breaks in the variance, which are not taken into account by the econometrician, will look as ARCH eects when the whole sample is used. In other words, it might be that for a sub sample the unconditional variance changes from say to and then back to the previous level. In this case to model the conditional variance as an ARCH model will be the wrong thing to do. In this case, it is recommended to divide the sample and test for ARCH for the sub periods, if no ARCH eects are found for any of the 9

sub periods but are found for the whole sample that is a clear indication of a break in the unconditional variance and not of ARCH eects. Many researchers wrongly estimated GARCH Models in many situations where there was only a change in regime. For example many papers use GARCH models to t interest rate series for USA when the change in the Volatility was simply a result of the dierent operative procedures of the Federal Reserve (a dierent distribution). GARCH Eects and Sampling Frequency It can be proved that GARCH models do not temporarily aggregate, or in other words if a model is GARCH using daily data cannot be GARCH with weakly data and so on. Given that we dont observe the data generating process in practice is very dicult to determine at which sampling frequency the data presents GARCH eects (if it has at all). Nevertheless there are some well established empirical regularities that show that the higher is the sampling frequency (say daily) the higher the GARCH eects. Weakly and every forth night data seem to also present GARCH eect. Monthly data usually does not have GARCH eects and whenever these are detected, are usually due to a structural break of the unconditional variance. Estimating GARCH Models Maximum likelihood Estimation with Gaussian Errors The estimation of GARCH type models is easily done by conditional maximum likelihood. If the model to be estimated is yt = xt + t Where xt is a (row) vector of predetermined variables, which could include lagged variables, is a parameter vector and t ~N(0, 2 ), where the conditional t variance is assumed to be GARCH(1,1), i.e. ; 2 = + 2 + 2 t t1 t1 Then the conditional distribution of yt is f (yt |xt , It1 ) = (2 2 ).5 exp(.5(yt xt )2 /2 ) t t Then the conditional log likelihood is
T X log L(, , , |It1 ) = (.5 log(2) .5 log( 2 ) .5 2 (yt xt )2 ) t t t=1

Notice that at time 1 we need initial values for 0 and 2 . These values are 0 usually assumed to be the equilibrium values, that is 2 = /(1 ) and 0 0 = (/(1 )).5

10

Maximum likelihood Estimation with non Gaussian Errors The unconditional distribution of many nancial time series seems to have fatter tails than the normal. GARCH eects may not account for this and we need to use another distribution for t . A tractable distribution is the t-distribution. We proceed as before but replace the Normal density function by f (t ) = ([( + 1)/2]/(/2))(( 2) 2 ).5 [1 + (2 /( 2 ( 2))](+1)/2 t t t Where n is a parameter to be estimated which represents the degrees of freedom. We estimate as before numerically subject to the constraint that n is greater than 2. How to compare Between Dierent GARCH - Specications Most of the GARCH models are non-nested (they cannot be written as a restricted version of a more general process). Therefore the comparison between dierent GARCH Models is no straight forward. Misspecication Tests on the standardized residuals. We have seen above that the residuals may be written as the product of a WN and the conditional standard deviation. For example for an ARCH(1) this can be written as t = vt ( + 2 )1/2 t1 Therefore we can test for the existence ARCH eects in the standardized residuals vt = t /( + 2 )1/2 t1 The model that cleans the standardized residuals is a candidate to be the true model. Some Other Ways Of Discriminating between alternative ARCH models. We are going to present two alternative ways of discriminating between ARCH models; (i) based on the use of auxiliary regressions of the squared residuals, (ii) based in their forecasting ability. (i) Comparison between alternative models based on the use of auxiliary regressions

11

Pagan and Scwhert (1989) suggest to use the following auxiliary regression as a mean of choosing between dierent Arch models. 2 = + 2 + t t t This regress the squared residuals on the tted variance of the alternative GARCH models. If the chosen GARCH model is appropriate to explain the conditional volatility of the series under scrutiny, you should expect to be zero, to be one and the t (R2 ) to be good. Pagan and Scwhert (1989) propose to test the joint hypothesis H0 ) = 0, = 1 H1 ) 6= 0, 6= 1 As a second step, they propose to compare the models that were not rejected on the basis of goodness of t. The argument being, the one with better t the better that mimics the conditional variance. They also propose to express the previous regression in logarithms to account for scale eects and then compare the goodness of t of this alternative auxiliary regression. (ii) Measuring the Accuracy of Forecasts of Dierent Arch Models. Hamilton (1994) propose to use the forecasting ability of the dierent ARCH models as a way of comparing these models. As we said before, the ARCH type of models have the property that they allow to forecast the conditional variance of a series, therefore a criteria which may enable us to choose between dierent models is to choose that one that forecast better. Various measures (loss functions) have been proposed for assessing the predictive accuracy of the forecasting ARCH models. The Mean Squared Error M SE = (1/T )( The Mean Absolute Error. M AE = (1/T )(
T X t=1 T X 2 (t 2 )2 ) t t=1

|2 2 |) t t

The Mean Squared Error of the log of the squared residuals. [LE]2 = (1/T )(
T X (ln(2 ) ln( 2 ))2 ) t t t=1

The Mean Absolute Error of the log of the squared residuals. 12

[M AE]2 = (1/T )(

T X t=1

|ln(2 ) ln( 2 )|) t t

For all the models we calculate the proportional improvement over a model which assumes constant variance, i.e., 2 = 2 (to account for scale eects). t The model that provides the largest proportional improvement is the one to be preferred. Hamilton also propose to compare the forecasting performance at dierent horizons (4 and 8 periods). That will slightly modify the above formulae in the following way; The Mean Squared Error M SE = (1/T )( The Mean Absolute Error.
T X |2 2 |) t+ t M AE = (1/T )( t=1 T X (2 2 )2 ) t+ t t=1

The Mean Squared Error of the log of the squared residuals.

The Mean Absolute Error of the log of the squared residuals. [M AE]2 = (1/T )(
T X t=1

| ln(2 ) ln( 2 )|) t+ t

where 2 is the forecast of the variance periods ahead given information at t time t.

13