5) Asymemtric - GRACH

Financial Risk Management:
Asymmetric Volatility Modeling: TARCH and EGARCH
Giacomo Morelli
March 9, 2022
1 / 34
Leverage Effect
In the ARCH/GARCH world, a large absolute return today predicts

higher future volatility in tomorrow.
The sign of the return doesn’t play any role in volatility dynamics.
Is this plausible? Let’s see what the data has to say!
2 / 34
Volatility Clustering
Figure: S&P 500 Lagged Returns vs. Absolute Returns.
3 / 34
Leverage Effect
Positive and negative returns can have different effects on volatility.

A negative return on a stock decreases the equity value of the firm. If
the debt is constant this implies that the leverage ratio of the firm is
increasing, thus making the firm more risky.
The leverage effect implies that a negative return increases volatility
more than a positive one of the same size.
Thus, leverage effect implies that volatility dynamics are asymmetric
in the sense that volatility reacts differently to positive and negative
shocks.
4 / 34
The Negative Sign/Negative Size Bias Tests
Engle and Ng (1993) introduce simple residual tests that allow us to

asses the evidence of asymmetric behavior:
The Negative Sign Test
The Negative Size Bias Test
The intuition behind the tests:
Estimate your favourite GARCH model
r2
Consider the series of squared standardized residuals ẑt2 = σ̂t2 .
t
If the model is correctly specified, then ẑt2 should not exhibit any
dynamics. In particular, the fact that on period t − 1 returns were
positive or negative should not have any forecasting power.
5 / 34
The Negative Sign/Negative Size Bias Tests
The Negative Sign/Negative Size Bias Tests are based on the

following auxiliary regression.
The Negative Sign Test:
ẑ 2 = c0 + c1 I(rt−1 <0) + ut
The Negative Size Bias Test:
ẑ 2 = c0 + c1 rt−1 I(rt−1 <0) + ut
The test statistics are the t-stats of the c1 coefficients in the two
regressions.
6 / 34
S&P 500 returns: Asymmetry after GARCH?
Figure: S&P 500 Lagged Returns vs. Std Residuals. Volatility asymmetry test:
Negative Sign, Stat = 12.7839, p-value = 0.0004
7 / 34
Asymmetric Volatility Models
In general, the Negative Sign/Negative Size Bias Tests signal the

presence of volatility asymmetry in the vast majority of stocks.
(Exchange rates returns do not usually exhibit asymmetry. Why?)
In what follows we are going to see some of the most commonly used
asymmetric volatility specifications:
TARCH (aka GJR-GARCH) (Zackonian (1993))
EGARCH (Nelson (1991))
In practical applications, the TARCH(1,1) is often found to be one of
the most successfully forecasting specifications.
8 / 34
TARCH
The TARCH model is defined as

√
rt = σ 2 zt zt ∼ D(0, 1)
where D is such that P(z < 0) = P(z > 0) = 0.5.

The TARCH(1,1) variance equation is
σt2 = ω + αrt−1
2 2
+ γrt−1 I− 2
t−1 + βσt−1
where I−
t is 1 if rt < 0 and 0 otherwise, ω > 0, α > 0, β > 0,
α + γ/2 + β < 1.
The persistence of the TARCH is α + γ/2 + β.
9 / 34
TARCH: Remarks
The TARCH(1,1) variance equation can alternatively be written as:

! 2 + βσ 2 t − 1 r
2 ω + αrt−1 t−1 ≥ 0
σt = 2 + βσ 2 t − 1 r
ω + (α + γ)rt−1 t−1 < 0
Thus, the TARCH allows for volatility asymmetry through the γ

coefficient. If the γ coefficient is positive, volatility increases more
after those days in which rt is negative.
For the vast majority of asset exhibiting asymmetric dynamics, γ is
positive.
(Some assets like gold have negative γ. Can you think of a reason
why gold has negative asymmetry?)
10 / 34
TARCH: Properties
Unconditional variance:
ω
σ2 =
(1 − α − γ/2 − β)
k-step ahead forecast of the conditional variance is:
σT2 +k|T = σ 2 + (α + γ/2 + β)k−1 (σT2 +1|T − σ 2 )
11 / 34
EGARCH
The EGARCH model is defined as

√
rt = σ 2 zt zt ∼ D(0, 1)
The EGARCH(1,1) variance equation is

" #
2 |rt−1 | rt−1 2
log σt = ω + α −m +γ + β log σt−1
σt−1 σt−1
$
where m = E|zt | and β ∈ (0, 1). If D ∼ N , then m = π2 .
The persistence of the EGARCH is β.
12 / 34
EGARCH: Remarks
Notice that because of log specification, the nonnegativeness of the

conditional variance is automatically granted.
The EGARCH allows for volatility asymmetry through the γ
coefficient. If γ is negative, volatility increases more after those days
in which rt is negative.
Note that because of the log specification, ω does not have to be
greater than zero.
(We are not going to introduce formulas for moments, acf and
forecasts of the EGARCH.)
13 / 34
TARCH Simulation Illustration
Figure: Returns, rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

innovations.
14 / 34
Figure: Volatility, σt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

innovations.
15 / 34
Figure: ACF rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.
16 / 34
Figure: ACF rt2 . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.
17 / 34
Figure: QQPlot rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.
18 / 34
19 / 34
EGARCH Simulation Illustration
Figure: Returns, rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

innovations.
20 / 34
Figure: Volatility, σt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

innovations.
21 / 34
Figure: ACF rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.
22 / 34
Figure: ACF rt2 . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.
23 / 34
24 / 34
25 / 34
Modeling S&P 500 Volatility
We have already detected the evidence of asymmetry in the volatility

dynamics of the S&P 500. Let’s assess the if TARCH and EGARCH
models improve the fit over the GARCH.
We assume that returns are described by
$
rt = c + σt2 zt zt ∼ N (0, 1)
and that the variance equation is going to be one of the GARCH,

TARCH or EGARCH.
26 / 34
S&P 500 Estimates
Figure: GARCH, TARCH, EGARCH Estimates.
27 / 34
S&P 500 Volatility Comparison
Figure: GARCH, TARCH, EGARCH Estimates.
28 / 34
S&P 500 Volatility Comparison: Zoom
Figure: Zoom on the European Sovereign Debt Crisis Period. Cumulative Returns.
29 / 34
Figure: Zoom on the European Sovereign Debt Crisis Period. GARCH, TARCH,
EGARCH.
30 / 34
Remarks
As expected, the specifications show evidence of asymmetric

dynamics.
The standardized residuals of the asymmetric specifications don’t
exhibit evidence of asymmetry.
Also, usually the standardized residuals of the asymmetric models
improve residual diagnostics over the GARCH(1,1)
(however, they also have more parameters).
31 / 34
News Impact Curves
It is of interest to construct graphs which summarize the impact of a

shock on the system.
New Impact Curves (NIC) have been introduced in Engle & Ng
(1993) to accomplish exactly this task.
The News Impact Curves plots the impact of a return on the future
conditional variance (assuming that the process is currently at its
unconditional level).
32 / 34
News Impact Curves
GARCH: NICG (r ) = (ω + βσ 2 ) + αr 2
The impact of returns on the conditional variance is a quadratic
function.
TARCH: NICT (r ) = (ω + βσ 2 ) + (α + γI(r <0) )r 2
The impact of returns on the conditional variance is a quadratic
function with different curvatures depending on the sign of r .
% &
EGARCH: NICE (r ) = exp(ω + β log σ 2 + α |rσ| − m + γ σr )
where m = E[|z|].
33 / 34
Figure: News Impact Curves: Evidence from the S&P 500.
34 / 34

5) Asymemtric - GRACH

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

5) Asymemtric - GRACH

Uploaded by

Copyright:

Available Formats

Financial Risk Management:

Asymmetric Volatility Modeling: TARCH and EGARCH

In the ARCH/GARCH world, a large absolute return today predicts

Is this plausible? Let’s see what the data has to say!

Figure: S&P 500 Lagged Returns vs. Absolute Returns.

Positive and negative returns can have diﬀerent eﬀects on volatility.

Engle and Ng (1993) introduce simple residual tests that allow us to

The Negative Sign/Negative Size Bias Tests are based on the

The Negative Size Bias Test:

ẑ 2 = c0 + c1 rt−1 I(rt−1 <0) + ut

In general, the Negative Sign/Negative Size Bias Tests signal the

The TARCH model is defined as

where D is such that P(z < 0) = P(z > 0) = 0.5.

The TARCH(1,1) variance equation can alternatively be written as:

Thus, the TARCH allows for volatility asymmetry through the γ

k-step ahead forecast of the conditional variance is:

σT2 +k|T = σ 2 + (α + γ/2 + β)k−1 (σT2 +1|T − σ 2 )

The EGARCH model is defined as

The EGARCH(1,1) variance equation is

Notice that because of log specification, the nonnegativeness of the

Figure: Returns, rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

Figure: Volatility, σt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

Figure: ACF rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: ACF rt2 . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: QQPlot rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: QQPlot rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: Returns, rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

Figure: Volatility, σt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian

Figure: ACF rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: ACF rt2 . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: QQPlot rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

Figure: QQPlot rt . ω = 0.01, α = 0.001, γ = 0.19, β = 0.9. Gaussian innovations.

We have already detected the evidence of asymmetry in the volatility

and that the variance equation is going to be one of the GARCH,

Figure: GARCH, TARCH, EGARCH Estimates.

Figure: GARCH, TARCH, EGARCH Estimates.

As expected, the specifications show evidence of asymmetric

It is of interest to construct graphs which summarize the impact of a

Figure: News Impact Curves: Evidence from the S&P 500.

You might also like