w8 volatilityii l

Upload: khanh-toan

Post on 06-Jul-2018

223 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/17/2019 W8 VolatilityII L

    1/34

    ECON3350/7350Volatility Models-II

    Alicia N. Rambaldi

    Week 8

    1/34

  • 8/17/2019 W8 VolatilityII L

    2/34

    In this lecture

    Readings

    Introduction

    ARCH(p) and GARCH(p,q)Linear Models with ARCH Errors

    Extensions to the Basic GARCH ModelEGARCH

    TGARCHTesting for Leverage Eff ectsExample

    (G)ARCH-in Mean

    IGARCH ModelsStochastic Volatility Models

    Realised Volatility

    Coming Up

    2/34

  • 8/17/2019 W8 VolatilityII L

    3/34

    Reference Materials

    Author Title Chapter Call No

    Enders, W AppliedEconometricTime Series,

    3e

    3   HB139 .E552015

    Brooks,C Introductory

    Econometricsfor Finance,

    3e

    9   HG173 .B76

    2014

    Verbeek, M A Guide to

    ModernEconomet-rics,4e

    8.10   HB139

    .V465 2012

    3/34

  • 8/17/2019 W8 VolatilityII L

    4/34

    ARCH(q) and GARCH(p,q)

    I   ARCH (q )  process

    E {2t |=t −1} =  α0 + α12t −1 + α22t −2 + ... + αq 2t −q = α0 + α(L)

    2t −1

    where  α(L)   is a lag polynomial of order  q − 1,  α(1) <  1,α0 > 0.

    I   GARCH (p , q )  process

    ht  = α0

     +

    X j =1

    α j 2

    t −

     j  +

    Xi =1

    β i ht −i 

    = α0 + α(L)2t −1 + β (L)ht −1

    For a GARH(1,1), a non-negative  ht   requires  α0,α1  and  β 1   to

    be non-negative.4/34

  • 8/17/2019 W8 VolatilityII L

    5/34

    Regressions and Autoregressions

    I   Consider a regression for the mean of  y t   with  ARCH (q )  errors:

    y t  = β 1 + β 2x 2,t  + ... + β K x K ,t  + t 

    t  = ν t 

    q α0 + α12t −1 + ... + αq 

    2t −q 

    where  ν t  ∼ N (0, 1),  α(1) <  1,  α0 > 0 andE [t ν t −s ] = 0 ∀  t , s I  Consider an autoregressive model,  AR (1), of the mean of  y t 

    with  GARCH (1, 1)  errors:

    y t  = a0 + a1y t −1 + t    |a1| <  1

    t  = ν t 

    q α0 + α12t −1 + β 1ht −1

    α0,α

    1  and β 

    1  to be non-negative, α

    1 + β 

  • 8/17/2019 W8 VolatilityII L

    6/34

    Extensions to the Basic GARCH Model

    I

     Since the GARCH model was developed, a huge number of extensions and variants have been proposed.

    I Problems with GARCH(p,q) Models:- Non-negativity constraints may still be violated- GARCH models cannot account for leverage eff    ects

    I  Some of the most  important extensions are:

    I Asymmetric Models TGARCH and EGARCH (account forleverage eff    ects),

    I (G)ARCH-M models (particularly suited to study assetmarkets)

    I The IGARCH model (a constrained model that accounts forthe persistence of volatility)

    6/34

  • 8/17/2019 W8 VolatilityII L

    7/34

    The EGARCH Model

    I  EGARCH was suggested by Nelson (1991). The variance equation isgiven by

    log (ht ) = α0 + β 1loght −1 + λt −1

    h0.5t −1+ α1

    t −1h0.5t −1

    I  Advantage - Since we model the   log (ht ), then even if theparameters are negative,  ht  will be positive.

    I   s t −1  =  t −1

    h0.5t −1and  |s t −1| =

    t −1h0.5t −1 are iid standard normal.   s t −1  and

    |s t −1|− E (|s t −1|)  are zero mean iid.   E (|s t −1|) =p 

    2/π  undernormality.

    The eff ect of a shock in   t   on   log (ht )If   t −1

    h0.5t −1is positive, eff    ect is  (α1 + λ)

    If   t −1

    h0.5t −1is negative, eff    ect is  (α1 − λ)

    Leverage eff    ect as long as  λ

    6= 0 and  λ <  0.

    7/34

  • 8/17/2019 W8 VolatilityII L

    8/34

    EGARCH notation in Enders

    I  Note that Enders uses diff erent notation. However, thenotation above is consistent with that used in EViews.

    I  Enders’ notation of the EGARCH is as follows

    log (ht ) = α0 + β 1loght −1 + α1t −1

    h0.5t −1

    + λ1 t −1h

    0.5t −1

    I  In this case the assumption is that  λ1 > 0 and  α1 )I If the random shock is  t  > 0 then the eff    ect is  λ1 + α1   (

  • 8/17/2019 W8 VolatilityII L

    9/34

    Representing The Leverage Eff ect

    -When  λ <  0 positive shocksgenerate less volatility than negative

    shocks (“bad news”).

    -The standardised   t −1/h0.5t −1   is unitfree and permits a more natural

    interpretation of the size and

    persitence of the shocks.

    Source: Enders (2015)

    9/34

  • 8/17/2019 W8 VolatilityII L

    10/34

    The EGARCH model (cont.)

    I  It is possible to extend the EGARCH model by includingadditional lags.

    I  Note that we can rewrite as:

    log (ht ) = α0 + β 1loght −1 + (α1 + λ)s t −1 if  t −1 > 0

    log (ht ) = α0 + β 1loght −1 + (α1 − λ)s t −1 if  t −1  0 while  λ 

  • 8/17/2019 W8 VolatilityII L

    11/34

    Volatility Prediction - EGARCHI  Assuming the parameters are known and the innovations are

    Gaussian, we proceed with a natural predictor

    ht  = exp [log (ht )] = exp 

    α0 + β 1loght −1 + λ

    s t −1 + α1

    s t −1

    E t ht + j  = E t  {exp [log (ht + j )]}   j  > 0

    I  The Gaussian case can be re-written to find a more explicitexpression (see for example Tsay (2002), pp 105)

    log (ht ) = (1− β 1)α0 + β 1loght −1 + g (s t −1)

    g (s t −1) = λs t −1 + α1(

    s t −1−p 2/π)I  The one-step ahead forecast is given by

    ht +h  = h

    2β1

      exp [(1−β 1)α0] exp [g (

    )]

    11/34

  • 8/17/2019 W8 VolatilityII L

    12/34

    The TGARCH or GJR Model

    I  Due to Glosten, Jaganathan and Runkle (1994), the

    threshold-GARCH:

    ht  = α0 + α12t −1 + λd t −1

    2t −1 + β 1ht −1

    where,

    d t −1 = 1 if  t −1  0.

    12/34

  • 8/17/2019 W8 VolatilityII L

    13/34

    Testing for Leverage Eff ects

    I  Estimate an ARCH or GARCH (Use a "diagnostic" approach

    to the test)

    I Form the standardised residuals (from ARCH or GARCH)

    s t  = e t /ĥ0.5

    t    for t  = 1,...,T 

    I An  F - statistic,  H 0   : a1  = a2  = .. = 0 using the regression

    s 2t    = a0 + a1s t −1 + a2s t −2 + ...

    If there are no leverage eff ects the squared errors should beuncorrelated with the level of the error term and the

    computed  F  statistic will fail to reject the null hypothesis.I A   t −test  H 0   : a1  = 0 using the regression

    s 2t    = a0 + a1d t −1 + εt 

    13/34

  • 8/17/2019 W8 VolatilityII L

    14/34

    Testing for Leverage Eff ects (cont.)

    I  Estimate the TARCH or EGARCH model,

    ht  = α0 + α12t −1 + λd t −1

    2t −1 + β 1ht −1

    or

    log (ht ) = α0 + β 1loght −1 + λt −1

    h0.5t −1+ α1

    t −1h0.5t −1 I   Test:

    H 0 : λ =  0

    I using a  t − test I PLEASE NOTE TYPO IN ENDERS (It should read

    H 0  : α1  = 0, page 157)

    14/34

  • 8/17/2019 W8 VolatilityII L

    15/34

    Example

    An Example of the use of a GJR ModelData: monthly S&P 500 returns, December 1979- June 1998.Source: Brooks (2002)

    y t  = 0.172

    (3.198)

    ht    =   1.243 + 0.0152t −1 + 0.498ht −1 + 0.6042t −1d t −1

    (t − stat ) (16.37) (0.44) (14.99) (5.77)

    Testing the leverage eff ect

    H 0 : λ = 0

    Computed  t  = 5.77. The computed value is larger than 1.645 (fora one-tail test) and thus we conclude that there are significant

    leverage eff 

    ects.15/34

    E l

  • 8/17/2019 W8 VolatilityII L

    16/34

    Example

    GJR Model (cont.)

    I  Data: monthly S&P 500 returns, December 1979- June 1998.

    I

     The news impact curve plots the next period volatility (ĥt )that would arise from various positive and negative values of  2t 

    16/34

  • 8/17/2019 W8 VolatilityII L

    17/34

    (G)ARCH-in Mean

    I  We expect a risk to be compensated by a higher return. So

    why not let the return of a security be partly determined by itsrisk?

    I  Engle, Lilien and Robins (1987) suggested the ARCH-Mspecification.

    y t  = µt  + t 

    µt  = β  + δ ht 

    µt  = β  + δ (α

    0 +

    Xi =1

    αi 2

    t −

    i )

    where,β ,  δ ,  α0  and  αi  are constants, and  δ  > 0.

    17/34

    (G) ( )

  • 8/17/2019 W8 VolatilityII L

    18/34

    (G)ARCH-in Mean (cont.)

    I  In Engle, Lilien and Robins (ELR)

    I y t  =excess return from holding a long-term asset relative to aon-period treasury bill

    I µt  =risk premium necessary to induce the risk-averse agent tohold the long-term asset rather than the one-period bond.

    I t  =  unforecastable shock to the excess return on thelong-term asset.

    18/34

    Th ELR ARCH M M d l ( )

  • 8/17/2019 W8 VolatilityII L

    19/34

    The ELR ARCH-M Model (cont.)

    I   The expected excess return from holding the long-term assetmust be just equal to the  risk premium:

    E t −1y t  = µt 

    I  The assumption is that the risk premium is an increasingfunction of the conditional variance of  t 

    I The greater the conditional variance of returns, the greater thecompensation necessary to induce the agent to hold thelong-term asset .

    19/34

    Th (G)ARCH M M d l ( )

  • 8/17/2019 W8 VolatilityII L

    20/34

    The (G)ARCH-M Model (cont.)

    I   A general (G)ARCH-M model is given by:

    y t  = x 0

    t β  + γ g (ht ) + t  +r 

    X j =1

    θ j t − j 

    ht  = α0 +

    q X j =1

    α j 2t − j  +

    p Xi =1

    β i ht −i 

    I   Usual  g (ht )  are:

    I log (ht ),√ 

    ht ,  (ht )m m = 1, 2, ..

    20/34

    E l

  • 8/17/2019 W8 VolatilityII L

    21/34

    Example

    Example. ELR - Excess yields on six-month treasury bills

    I   r t  quarterly yield on a three-month treasury bill held from  t   to

    t  + 1.I Rolling over all proceeds, at the end of two quarters and

    individual investing $1 at the beginning of the  t  will have(1 + r t )(1 + r t +1)  dollars.

    I   R t  quarterly yield on a six-month treasury bill, buying andholding the six-month bill for the full two quarters will result in(1 + R t )

    2 dollars.

    I   The excess yield,  y t , due to holding the six-month bill is

    y t  = (1 + R t )2

    − (1 + r t +1)(1 + r t )which is approximately equal to:

    y t  = 2R t  − r t +1 − r t 21/34

  • 8/17/2019 W8 VolatilityII L

    22/34

    Example

    (cont.)

    y t  = 0.142 + t 

    (4.04)

    The excess yield of 0.142% per quarter is more than four standard

    deviations from zero.

    LM ARCH  = 10.1 which is larger than the critical value from the  χ2(1)

    at 1% (6.635).

    However, the post-1979 period showed higher volatility than theearlier period.

    22/34

  • 8/17/2019 W8 VolatilityII L

    23/34

    Example

    The ARCH-M estimates are:

    ŷ t  = −0.0241 + 0.687ĥt (−1.29) (5.15)

    ĥt  = 0.0023 + 1.64(0.42t −1 + 0.3

    2t −2 + 0.2

    2t −3 + 0.1

    2t −4)

    (1.08) (6.30)

    Results:

    I   Risk premium are time-varying

    I  Estimate of 1.64 implies the unconditional variance is infinite

    (the conditional variance if finite)I  During volatily periods, the risk premium rises as risk-averse

    agents seek assets that are conditionally less risky.

    23/34

    IGARCH Models

  • 8/17/2019 W8 VolatilityII L

    24/34

    IGARCH Models

    I  Sometimes when we estimate GARCH(1,1) models we findthat the sum of the estimated  α1  and  β 1  coefficients is closeto 1.

    I  Due to Nelson (1990), constraining  α1 + β 1 = 1 yields the

    Integrated-GARCH or  IGARCH (1, 1)  model

    ht  = α0 + (1 − β 1)2t −1 + β 1Lht I   It yields a more parsimonious representation of the error

    process (there is one less parameter to estimate)

    I forces  ht  to act like a process with a unit root.

    24/34

    IGARCH Models (cont )

  • 8/17/2019 W8 VolatilityII L

    25/34

    IGARCH Models (cont.)

    I   However, solving for  ht 

    ht  =  α0/(1−β1) + (1−β 

    1)∞

    Xi =0

    β i 

    12

    t −

    1−

    I Unlike a true nonstationary process,  ht   is a geometricallydecaying function of current and past realisations of the  2t sequence.

    25/34

    IGARCH (cont )

  • 8/17/2019 W8 VolatilityII L

    26/34

    IGARCH (cont.)

    I

     As it is a decaying function of current and past values of theshocks it can be estimated like any other  GARCH (p , q )

    I   The  one −step-ahead forecast of the conditional variance is

    E (ht +1|

    =t ) = α0 + ht 

    I   The   j −step-ahead forecast of the conditional variance is

    E (ht + j |=t ) = j α0 + ht I  Thus, appart from the intercept the forecast of the conditional

    variance for the next period is the current value of theconditional variance. The unconditional variance is infinite.

    26/34

    Stochastic Volatility Models

  • 8/17/2019 W8 VolatilityII L

    27/34

    Stochastic Volatility Models

    I  An alternative class of models to those in the  GARCH (p , q )family are the stochastic volatility models (SV).

    I  They diff er principally because the conditional variance is not adeterministic function of past information.

    I  They are particularly designed to deal with volatility clustering.

    I  SV models contain a second error term which enters theconditional variance specification and are written asstate-space  models

    27/34

    Stochastic Volatility Models (cont )

  • 8/17/2019 W8 VolatilityII L

    28/34

    Stochastic Volatility Models (cont.)

    I  The simple forms can be estimated using maximum likelihoodor Monte Carlo methods via Kalman filtering

    I  A SV model is defined as

    at  = ht t 

    (1 − β 1L − . . .β mLm) ln ht  = α0 + ν t where  at   are returns,  ν t 

     ∼N (0,σ2ν ),  t 

     ∼N (0, 1),

    α0,β 1, . . . , β m  are parameters

    28/34

    Stochastic Volatility Models (cont )

  • 8/17/2019 W8 VolatilityII L

    29/34

    Stochastic Volatility Models (cont.)

    Example

    A SV model for the daily returns of an asset price,  y t   is:

    y t  = a + σexp (1

    2

    θt )t    t  ∼

    N (0, 1)

    θt +1 = φθt  + ηt    ηt  ∼ N (0, σ2η)   0

  • 8/17/2019 W8 VolatilityII L

    30/34

    Realised Volatility

    I   The conditional variance is latent (ie not observed)

    I  Models from the  GARCH (p , q )  family (and SV type models)are estimates of the latent conditional variance

    I   Criticism are that standard latent volatility models fail todescribe in an adequate manner the low, but slowly decreasing,

    autocorrelations in the squared returns that are associatedwith high excess kurtosis of returns.

    I  Search for an adequate framework has led to the analysis of high frequency intraday data.

    I  Merton (1980) noted that the variance over a fixed intervalcan be estimated arbitrarily, although accurately, as the sum of squared realisations, provided the data are available at asufficiently high sampling frequency

    30/34

    Realised Volatility (cont )

  • 8/17/2019 W8 VolatilityII L

    31/34

    Realised Volatility (cont.)

    I  Andersen and Bollerslev (1998) showed that ex-post daily

    foreign exchange volatility is best measured by aggregating 288squared five minute returns.

    I   The five-minute frequency is a trade-off  between accuracy,which is theoretically optimized using the highest possiblefrequency, and microstructure noise that can arise through thebid-ask bounce, asynchronous trading, infrequent trading, andprice discreteness, among other factors

    I   Ignoring the remaining measurement error, which can beproblematic, the ex-post volatility essentially becomes

    “observable.”I   As volatility becomes “observable,” it can be modeled directly,

    rather than being treated as a latent variable

    31/34

    Realised Volatility (cont.)

  • 8/17/2019 W8 VolatilityII L

    32/34

    Realised Volatility (cont.)I  Consider a simple discrete time model in which the  daily returns of 

    a given asset are typically characterized as

    r t  = h0.

    5t    ηt 

    where  ηt   is a sequence of independently and normally distributedrandom variables with zero mean and unit variance,  ηt   NID (0, 1)and  h0.5t    is the (time-varying) standard deviation of daily returns.

    I  Suppose that, in a given trading day  t , the logarithmic prices are

    observed tick-by-tick.I Consider a grid  Λt  = {τ 0,..., τ nt }  containing all observation

    points, andI p t ,i ,   i  = 1,...,nt , to be the   i 

    th (logarithmic) price observationduring day  t , where  nt   is the total number of  observations at

    day t .I Furthermore, suppose that  r t ,i  = p t ,i  − p t ,i −1   is the   i th

    intraperiod return of day  t , such that

    r t  =

    nt 

    Xi =0

    r i ,t 

    32/34

    Two unbiased estimators of RV

  • 8/17/2019 W8 VolatilityII L

    33/34

    uI   The realised variance is defined as the sum of all available

    intraday high frequency squared returns given by

    RV (all )t    =

    nXi =0

    r 2t ,i 

    I   The squared daily return can be written as

    r 2t   =

      nXi =0

    r t ,i 

    !2

    I  They are unbiased as it can be shown that

    E (r 2t  |=t ,0) = E (RV (all )t    |=t ,0) = ht   (=t ,0   is the information setavailable prior to the start of day  t ).

    I  Further reading, see for example:I McAleer, Michael and Medeiros, Marcelo C.(2008) ’Realized

    Volatility: A Review’,   Econometric Reviews , 27: 1, 10 — 45

    33/34

    Coming Up

  • 8/17/2019 W8 VolatilityII L

    34/34

    g p

    I   Vector Autoregressive Models

    34/34