Eco No Metrics Forecasting 2001 - ARCH Models

Forecasting Financial Markets
ARCH Models
Copyright © 1999-2006
Investment Analytics
Copyright © 1999-2006 Investment Analytics Forecasting Financial Markets – ARCH Models Slide: 1
Overview
¾ Asset Return Behavior
¾ Models
9ARCH
9GARCH
9GARCH-M
¾ Estimation procedures
¾ Applications
Defining Volatility
¾ Nonparametric Definition
9X is more volatile than y if:
9P(|X|>c) > P(|Y|>c) for all c
¾ Time Series Volatility
9Time series yt become more volatile when
• P(| yt+1 | >c) > P(| yt | > c) for all c
• Occurs if and only if σt=1 > σt
Estimating Historical Volatility
¾ Standard Deviation
X i = Ln ( Pi +1 /Pi )
σ = ∑ ( X i − X ) 2 /( N − 1)
¾ Parkinson (5x times more efficient)
1
σ=
2 N Ln(2)
∑ Ln( H i / Li )
¾ Garman & Klass (7 x efficiency)

1 1
A B S [ ∑ [ L n ( H i / L i )] 2 −
N 2
σ =
1
N
∑ ( 2 L n ( 2 ) − 1 )[ L n ( C i / C i − 1 )] 2 ]
Asset Return Characteristics
¾ Thick Tails
¾ Non-Normal Distribution
¾ Volatility Clustering
¾ Leverage
¾ Trading vs. Non Trading Periods
¾ Forecastable Events
¾ Volatility & Correlation
Thick Tails, Non-Normal Distribution
9Mandelbrot (1963), Fama (1963, 1965)
S&P500 Index Monthly Returns 1/1950-2/2001
X <= -0.061 X <= 0.075

5.0% 95.0%
12
10
Normal(0.007, 0.041)
8
Skewness = -0.6
4
Kurtosis = 5.7
2
0
-0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
Volatility is Stochastic
¾ Non-Constant Variance
9Changing conditional variance
9Highly volatile periods
9Interspersed with quiet periods
9Large returns tend to be followed by large
returns (positive or negative)
A Conditionally Heteroscedastic Series
10%
8%
6%
4%
2%
0%
-2%
-4%
-6%
Volatility Clustering
9Periods of high vol. followed by high vol.
9Followed by decay back to normal levels
SP500 Index Volatility

160%
140%
120%
100%
80%
60%
40%
20%
0%
Jan-50
Jan-52
Jan-54
Jan-56
Jan-58
Jan-60
Jan-62
Jan-64
Jan-66
Jan-68
Jan-70
Jan-72
Jan-74
Jan-76
Jan-78
Jan-80
Jan-82
Jan-84
Jan-86
Jan-88
Jan-90
Jan-92
Jan-94
Jan-96
Jan-98
Jan-00
Volatility Distributions
Enron: Volatility Distribution 9/83 - 2/01
X <= -2.8176 X <= -1.4954
5.0% 95.0%
1.2
0.8
LogLogistic(-5, 2.8, 12.4)

0.6
0.4
0.2
0
-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0
Volatility(%) The Volatility Cone
Maximum
Average
Minimum
0 30 60
Days
Leverage Effect
¾ Black (1976)
9Stock price changes negatively correlated with
volatility
9Fixed costs provide a partial explanation
• Firm with equity and debt becomes more leveraged as
stock falls
• Raises equity returns volatility
9Correlation too large to be explained by leverage
alone
• Christe (1982), Schwert (1989)
Correlation in Returns & Volatility
Correlation: Returns & Volatility in
SP500 Index 1950-2001
1.0
0.8
0.6
0.4
0.2
0.0
Sep-51
Sep-53
Sep-55
Sep-57
Sep-59
Sep-61
Sep-63
Sep-65
Sep-67
Sep-69
Sep-71
Sep-73
Sep-75
Sep-77
Sep-79
Sep-81
Sep-83
Sep-85
Sep-87
Sep-89
Sep-91
Sep-93
Sep-95
Sep-97
Sep-99
-0.2
-0.4
-0.6
-0.8
-1.0
Non Trading Periods
¾ Trading Causes Volatility
9Volatility after weekends is higher
9Not 3 x higher (closer to 1.2 x)
9French & Roll (1986), NYSE & AMEX
9Baillie & Bollerslev (1989) FX
Forecastable Events
¾ News Announcements
9Associated with high ex-ante volatility
9Stock volatility is high around earnings dates
• Cornell (1978), Patell & Wolfson (1979, 1981)
9Bond volatility rises on macroeconomic news
• Harvey & Huang (1991, 1992)
¾ Seasonal Effects
9Volatility higher at the open and close
• Harris (1986), Gerity & Mulherin (1992)
Intra-Day Seasonality in the Bonds
115
110
105
100
95
90
85
80
8:20- 9.00- 10.00- 11.00- 12.00- 13.00- 14.00-
9:00 10.00 11.00 12.00 13.00 14.00 15.00
Volatility & Correlation
¾ Volatility & Serial Correlation in Returns
9 Strong inverse relationship
• Stock indices (LeBaron, 1992)
• FX markets (Kim 1989)
¾ Co-Movements in Volatility
9 Volatilities tend to change together
• Stocks: Black (1976)
• FX: Diebold & Nerlove (1989)
9 Also across markets
• Stock & bond volatilities move together (Schwert, 1989)
Volatility Correlation
Correlation: IBM vs JNJ Volatility
1.2
1.0
0.8
0.6
0.4
0.2
0.0
May-85 May-88 May-91 May-94 May-97 May-00
-0.2
-0.4
Asset Characteristics - Conclusions
¾ The Bad News
9iid Gaussian model inappropriate
¾ The Good News
9Correlation suggests few common factors may
explain variation
9Stochastic volatility models (GARCH, etc.)
GARCH Models
¾ Generalized Autoregressive Conditional
Heteroscedasticity
9Time-varying conditional volatility
9Handles volatility clustering
9Produces leptokurtic returns distributions
• Fatter tails than Normal
• Because changing variance allows for more extreme
values
9Many model variants (>20 !)
9Very widely applied in finance
Conditional Forecast
¾ AR(1) process: yt+1 = a0 + a1yt + εt+1
9 εt is white noise process
• Constant mean (0), variance σ2
• Uncorrelated
¾ Conditional Forecast
9Et(yt+1) = a0 + a1yt
¾ Forecast Error Variance
9Et[yt+1 - Et(yt+1)]2 = Et[yt+1 - (a0 + a1yt)]2 = Et(εt+1 )2 = σ2
Unconditional Forecast
¾ Unconditional Expectation is Constant
9E(yt+1) = a0 /(1 - a1)
• i.e. the long run mean
¾ Unconditional Variance is Constant
9E[yt+1 - E(yt+1)]2 = E[yt+1 - a0 / (1 - a1)]2 = σ2 / (1 - a1)2
9Unconditional forecast has greater variance
• Since 1 / (1 - a1) > 1
¾ Conditional Variance is Constant
9Et(εt+12) = Et(yt+1 - a0 + a1yt )2 = σ2
Auto-regressive Conditional
Heteroskedastic (ARCH) Model
¾ Suppose conditional variance is not constant
¾ Model conditional variance as an AR(p) process
9εt2 = α0 + α1 (εt-1)2 + α2 (εt-2)2 + . . . + αq (εt-q)2 + vt
• vt is white noise
¾ Multiplicative ARCH model (Engle):
9εt2 = [α0 + α1 (εt-1)2] vt2
• is white noise with σ2v = 1
• εt are independent of each other
• α0 > 0 and 0 < α1 < 1
Properties of ARCH
¾ Unconditional Mean
9 Eεt = E[vt(α0 + α1 (ε′t-1)2 )1/2]
= Evt E[(α0 + α1 (ε′t-1)2 )1/2] = 0
¾ Unconditional Variance
9Eεt2 = E[vt2(α0 + α1 (ε′t-1)2 )]
= Evt2 E[α0 + α1 (ε′t-1)2 ]
= α0 / (1 - α1)
¾ Covariance
9E[εtεt-i ] = 0; i ≠ 0
Properties of ARCH
¾ Conditional Mean of Errors
9E(εt / εt-1 , εt-2 . . . ) = Evt = E(α0 + α1 εt-12 )1/2] = 0
¾ Conditional Variance of Errors
9So far, heteroskedasticity has no effect on {εt }
• Mean, covariances are zero,variance constant
9Impact is on conditional variance
• E(εt2 / εt-1 , εt-2 . . . ) = σt2 = α0 + α1 εt-12
Properties of ARCH Series {yt}
¾ Conditional Mean: Et-1(yt) = a0 + a1yt-1
¾ Conditional Variance
• Var(yt / yt-1, yt-2, . . .) = Et-1(yt - a0 - a1yt-1 )2 = Et-1εt2 = α0 + α1 εt-12
9So minimum conditional variance is α0
• Since α1 and εt-12 > 0
¾ Unconditional Mean
9Eyt = E[a0 / (1 + a1) + Σa1i εt-i ] = a0 / (1 + a1)
¾ Unconditional Variance
9Var(yt ) = Σa12i var(εt-I) = [α0 / (1 - α1)] x [1 / (1 - a12)]
• Increasing in both α0 and | a1 |
Key Points about ARCH
¾ Errors Moments
9Zero mean, covariance, unconditional variance
¾ Error variance fluctuates
9For large εt , variance of εt will be large
9Periods of tranquility & volatility in {y}
¾ Errors are not independent
9Related through second moment
¾ Parameter values
9Restricted to ensure variance > 0 and series is stable
• α0 > 0 and 0 < α1 < 1
ARCH Example
ARCH Error Process ε t
¾ Parameters 2.0
1.5
9 α0 = 0.3, α1 = 0.9
1.0
0.5
0.0
9 a1 = 0.25 & 0.9 -0.5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
-1.0
¾ Effects & Interactions -1.5

-2.0
9Larger α1, more

-2.5
ARCH Process yt = a 1yt-1 + ε t

persistent are shocks in {εt} 2.0
9Larger a1, more 1.0
0.0
persistent is change in {yt}
1
11
13
15
17
19
-1.0
-2.0
-3.0
-4.0 yt
y't
-5.0
GARCH Models
¾ GARCH(p, q)
9Error Process εt = vt√ ht
9 σv2 = 1
9 q p
ht = α 0 + ∑ α ε 2
i t −i + ∑ β i ht −i
i =1 i =1
¾ Error Process {εt}
9Conditional mean and variance are zero
9Conditional variance is ht
Properties of GARCH
¾ Disturbances of series {yt} follow ARMA process
9ARMA(p, q) process in series {εt2}
q p
Et −1ε t2 = ht = α 0 + ∑ α iε t2−i + ∑ β i ht −i
i =1 i =1
¾ Estimating a GARCH Model

9Fit ARMA model to series {yt}
9Evaluate sample autocorrelations of squared residuals
• Should suggest an ARMA(p, q) process in series {εt2}
GARCH(1,1) Model
¾ Model form:
ht = α0 + α1ε2t-1 + βht-1
¾ Estimate average, unconditional variance:
9Set ht = ht-1 = h
¾ Solution: h = α0 / (1 - α1 - β)
• For stationarity, α1 + β must be < 1
• Sum α1 + β is known as persistence
¾ Estimation
• Assume scaled residual rt = εt /ht1/2 is normally distributed
• Use numerical optimization to estimate parameters
GARCH Models
Parameter $/BP US Stocks T-Bond
σ (%pa) 11.33 12.02 9.72

α0 0.00685 0.00233 0.00410
α1 0.0678 0.0213 0.0132
β 0.9186 0.9740 0.9749
Persistence
(α1 + β) 0.9864 0.9953 0.9890
Persistence Parameter
1.00
Variance
0.90
0.80
0 20 25
Days ahead
GARCH Estimation
¾ Step 1
9Fit ARMA model to series {yt}
9Calculate sample variance
1 T 2
σˆ = ∑ εˆi
2
T t =1
¾ Step 2
9Plot sample autocorrelations of series {εt2}
T
∑ t
(εˆ 2
− σˆ 2
)(εˆ 2
t −i − σˆ 2
)
ρ (i ) = t =i +1 T
∑ t
(εˆ 2
t =i
− σˆ 2 2
)
GARCH Estimation
¾ Step 3
9Test significance of autocorrelations
• 95% of coefficients should lie in range ±1.96/√T
• Ljung-Box test
9Significant coefficients indicate GARCH errors
¾ Step 4
9Fit ARMA(p, q) model to series {εt2}
• Use standard Box-Jenkins methodology
Likelihood Function in Regression
¾ Simple Linear Regression: yt = βxt + εt
9 εt ~ IID No(0, σ2)
¾ Likelihood
9 L = (-n/2)[Ln(2π) +Ln(σ2)] - (1/2σ2)Σ(yt - βxt)2
¾ MLE Estimates
9 (σ′)2 = Σ(εt)2 / n
9 β′ = Σ(xtyt ) / Σ(xt )2
¾ Standard Error
9 σ′β = σ′ / {Σ(xt - xmean)2}1/2
Likelihood Estimation in GARCH
¾ Now assume εt2 = [α0 + α1 (εt-1)2] vt2
9Conditional variance ht = α0 + α1ε2t-1
• ht = α0 + α1(yt-1 –βxt-1)2
¾ Likelihood
9L = (-n/2)Ln(2π) - (1/2)Σ Ln(ht) - (1/2)Σ ht(yt - βxt)2
• Maximize L with respect to α0, α1 and β
• Nonlinear model
• Use search algorithm
Lagrange Multiplier Test
¾ For ARCH disturbances (Engle 1982)
¾ Use OLS to fit AR(n) to series {yt}
¾ Form series {εt2}
¾ Estimate regression relationship
9ε′t2 = α0 + α1 (ε′t-1)2 + α2 (ε′t-2)2 + . . . + αq (ε′t-q)2
¾ Testing
9If no ARCH effects α1 = α2 = . . . = αq = 0
9TR2 ~ χq2 under the null hypothesis of no ARCH
• If test statistic is large, reject null of no ARCH
Example: US Inflation
¾ Bollerslev (1986)
9 Compared standard ARMA to GARCH
¾ ARMA δ t = 0.240 + 0.552δ t −1 + 0.177δ t −2 + 0.232δ t −3 − 0.209δ t −4 + ε t
9 All coefficients significant
9 No significant ACF or PACF coefficients
9 However, ACF & PACF of squared residuals
show significant autocorrelations
¾ GARCH(1,1)
δ t = 0.141 + 0.433δ t −1 + 0.229δ t − 2 + 0.349δ t −3 − 0.162δ t − 4 + ε t
ht = 0.007 + .135ε t2−1 + 0.829ht −1
9 Similar forecasts
9 Changing confidence limits
• Expand during periods of volatility
ARCH-M Models
¾ Extension of ARCH
9ARCH in the mean
9Engle, Lilien & Robins (1987)
¾ Applications in Asset Markets
9Idea: risk premium should be increasing
function of conditional variance of returns
ARCH-M Models for Asset Returns
¾ Assume yt = µt + εt
9yt is the excess return over risk free rate (T-Bills)
9 µt is required risk premium
9εt is unforecastable shock to excess return
¾ Assume Risk Premium Follows ARCH Process
9µt = β + δht (δ > 0)
q
ht = α 0 + ∑ α iε t2−i
i =1
ARCH-M Process Example
ARCH Error Process ht
3.5
3.0
¾ Model: 2.5
yt = −1 + ht + εt
2.0
1.5
1.0
0.5
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
ARCH Process yt = β + δ ht + ε t
4.0
3.0
2.0
1.0
0.0
1
11
13
15
17
19
-1.0
-2.0
GARCH-M Model
¾ Process: yt = µt + εt
q p
E ε = ht = α 0 + ∑ α ε
t −1 t
2 2
i t −i + ∑ β i ht −i
i =1 i =1
¾ Unconditional Variance is Constant

α0
Var (ε t ) = q p
>0
1 − ∑αi − ∑ βi
i =1 i =1
Likelihood Estimation
Normal Errors
¾ Minimize Log(Likelihood) to find optimal
parameters
(n − q) 1 n 1 n ε t2
Log ( L) = − 2 log(2π ) − ∑ Log (ht ) − ∑
2 2 t = q +1 2 t = q +1 ht
Likelihood Estimation:Student-t Errors
¾ Appropriate for Modeling Asset Returns
9Distribution of standardized residuals εt / √ht
has fatter tales than Normal distribution
• υ is # degrees of freedom, estimated along with
parameters θ
n
Log ( L) = ∑ L (Θ,υ )
t = q +1
t
⎧ ⎛ υ 1 ⎞⎫ 1
Lt (Θ,υ ) = − log ⎨ B⎜ , ⎟⎬ − Log (υ − 2)
⎩ ⎝ 2 2 ⎠⎭ 2
1 ⎛υ +1⎞ ⎛ ε t2 ⎞
− Log (ht ) − ⎜ ⎟ Log ⎜⎜1 + ⎟⎟
2 ⎝ 2 ⎠ ⎝ ht (υ − 2) ⎠
EGARCH Models
¾ Exponential GARCH(p, q)
⎛ ε t −i ⎞ q * ⎛ ε t −i
q ⎞ p
Log (ht ) = α 0 + ∑ α i ⎜⎜ 1/ 2 ⎟⎟ + ∑ α i ⎜⎜ 1/ 2 − ε ⎟⎟ + ∑ β i ht −i
i =1 ⎝ ht −i ⎠ i =1 ⎝ ht −i ⎠ i =1
9Where
⎛ εt ⎞ 2
ε = E ⎜⎜ 1/ 2 ⎟=
⎟ if ε t ≈ N (0,1)
⎝ ht ⎠ π
Factor GARCH
¾ Volatility Equivalent of CAPM
σ t2 = β 2σ Mt
2
+ σ ε2t
IBM Volatility Factor Model
100 y = 0.496x + 26.05

R2 = 0.22
80
IBM Vol
60
40
20
0
0 20 40 60 80 100
SP500 Vol
Lab: Estimating an ARCH Model
for Cable
USD/GBP Spot Rate
2.80
2.60
2.40
2.20
2.00
1.80
1.60
1.40
1.20
1.00
ARCH Model Procedure
¾ Fit AR(1) Model
9Form series {yt} = ∆Ln(rt)
9yt = a0 + a1 yt-1 + εt
¾ Examine ACF of Squared Residuals
¾ Fit ARCH Model to Error Process
9Estimate model ht = εt2 = α0 + α1εt-12
¾ Lagrange Multiplier Test for ARCH Effects
9nR2 ~ χ2(q)
Solution: ARCH Model for Cable
¾ Evidence of AR(1) ARCH process:
Squared Residuals ACF and PACF
0.25 ACF
PACF
0.20 Upper 95%
Lower 95%
0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
Solution: ARCH Model for Cable
MLE Max Likelihood
α0 AIC -2616.55
α1 0.3635 BIC -2612.98
DW 2.07
2
R 18.0% p
2
nR 47.452 0.00%
ANOVA DF SS MS F p
Model 1 1.04E-05 1.04E-05 57.46 0.00%
Error 261 4.74E-05 1.82E-07

Total 262 5.79E-05
GARCH in Research – Equities
¾ Noh, Engle, Kane (1993/4)
9 Test of Efficiency of the S&P 500 Index Option Market using
Variance Forecasts
• Straddle trading using GARCH vs IV forecasting
9 Significant profits generated by GARCH system
• Return 1.62% /day, after transaction costs
¾ Highly significant ARCH effects found
9 Stocks (Engle & Mastufa, 1992)
9 Indices (Akgiry, 1989)
9 Futures (Schwert, 1990)
GARCH in Research – Fixed Income
¾ Weiss (1984)
916 different macroeconomic variables
• Including AAA corporate bind yields
• Very significant ARCH effects
¾ Hong (1988) – excess return 3m vs. 1m Tbills
9GARCH(1,1) model, persistence > 1
9Engle, Lillien & Robins (1987)
• Similar results for 6m vs. 3m Tbills
GARCH in Research - Forex
¾ High Frequency FX Processes
9Well described by GARCH(1,1,) Model
• Hsieh (1989)
• Taylor (1986)
• McCurdy & Morgan (1988)
• Kugler & Lenz (1990)
9Less significant for monthly data
• Diebold (1988), Baillie & Bollerslev (1989)
Summary: GARCH Models
¾ Major development of last decade
¾ Now widely used in financial markets forecasting
9Option volatility
9Equity & index markets
9FX markets
¾ Many financial time series exhibit non-stationarity
9Due to heteroscedasticity
9Can be modeled successfully with GARCH
9Models how the process deals with shocks

Eco No Metrics Forecasting 2001 - ARCH Models

Caricato da

Informazioni sul documento

Descrizione originale:

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Eco No Metrics Forecasting 2001 - ARCH Models

Caricato da

Copyright:

Formati disponibili

Forecasting Financial Markets

¾ Garman & Klass (7 x efficiency)

S&P500 Index Monthly Returns 1/1950-2/2001

X <= -0.061 X <= 0.075

SP500 Index Volatility

LogLogistic(-5, 2.8, 12.4)

9 a1 = 0.25 & 0.9 -0.5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

¾ Effects & Interactions -1.5

9Larger α1, more

ARCH Process yt = a 1yt-1 + ε t

9Larger a1, more 1.0

¾ Estimating a GARCH Model

Parameter $/BP US Stocks T-Bond

σ (%pa) 11.33 12.02 9.72

¾ Unconditional Variance is Constant

IBM Volatility Factor Model

100 y = 0.496x + 26.05

Error 261 4.74E-05 1.82E-07

Potrebbero piacerti anche