Time Varying Volatility

Asad Dossani

Fin 625: Quantitative Methods in Finance

1 / 53
Time Varying Volatility

A stylized feature of financial markets is that while returns are dif-

ficult to predict, volatility is to some extent predictable. Volatility
varies over time. It is persistent, meaning that periods of high and
low volatility are clustered together.

2 / 53
Daily S&P 500 Returns

3 / 53
Time Varying Volatility

Unlike returns, volatility is not observable. It has to be estimated

from the data. Volatility estimates and forecasts are important for
a variety of applications. These include portfolio optimization, risk
management, and derivatives pricing.

4 / 53
Historical Volatility

For simplicity, we assume that the mean return is zero. This is a

reasonable assumption when working with daily data. The historical
volatility is the average squared return over some historical period.
T
1 X 2
σ2 = rt
T
t=1

5 / 53
Exponentially Weighted Moving Average

The Exponentially Weighted Moving Average (EWMA) is an im-

provement over historical volatility. It allows more recent observa-
tions to have a stronger impact on the current level of volatility,
relative to older data points. It is a weighted average of the previ-
ous period’s squared return and previous period’s variance.

σt2 = (1 − λ)rt−1
2 2
+ λσt−1

6 / 53
Exponentially Weighted Moving Average

λ represents the decay factor. It is commonly set to 0.94 for daily

data. The EWMA is computed recursively each period. The initial
volatility is set to the unconditional mean or average squared return
over all periods.

σt2 = (1 − λ)rt−1
2 2
+ λσt−1

7 / 53
Autoregressive Conditional Heteroskedasticity

Autoregressive conditional heteroskedasticity (ARCH) models are a

more sophisticated approach to modeling and forecasting volatility.
ARCH models assume that the squared return follows an autore-
gressive process.

Both EWMA and ARCH models feature volatility persistence. Unlike

EWMA models, ARCH models also feature mean reversion. This
means that volatility reverts to a long run mean over time.

8 / 53
Martingale Difference Sequence

A martingale difference sequence (MDS) has an expected value of

zero, conditional on past information. If yt is a MDS, then:

Et−1 (yt ) = 0

9 / 53
ARCH(1) Model: yt ∼ ARCH(1)

We assume the daily log return yt has mean zero. (This assumption
is easily relaxed by first demeaning returns) The volatility is σt > 0,
and is determined by information available before time t. ut is
independent and identically distributed with mean zero and variance
one.

yt = σt ut
σt2 = a0 + a1 yt−1
2

ut ∼ IID(0, 1)

10 / 53
ARCH(1) Model: yt ∼ ARCH(1)

yt is a MDS and is unpredictable. The variance of yt depends on t.

Et−1 (yt ) = Et−1 (σt ut )

= σt Et−1 (ut )
=0
Vart−1 (yt ) = Vart−1 (σt ut )
= σt2 Vart−1 (ut )
= σt2

11 / 53
ARCH(1) Model: yt ∼ ARCH(1)

We define ηt as follows. ηt is a MDS and a white noise series.

ηt = (ut2 − 1)σt2
Et−1 (ηt ) = σt2 [Et−1 (ut2 ) − 1]
= σt2 [E(ut2 ) − 1]
=0
ηt = σt2 ut2 − σt2
σt2 ut2 = ηt + σt2

12 / 53
ARCH(1) Model: yt ∼ ARCH(1)

We show that yt2 follows an AR(1) model.

yt2 = σt2 ut2

yt2 = σt2 + ηt
yt2 = a0 + a1 yt−1
2
+ ηt
a0
E(yt2 ) =
1 − a1
a0 > 0, a1 ∈ (0, 1)

13 / 53
ARCH(1) Model: yt ∼ ARCH(1)

yt2 is positively autocorrelated, implying predictable volatility.

|k|
Corr(yt2 , yt+k
2
) = a1
k = 0, ±1, ±2, . . .

14 / 53
ARCH(1) Model: yt ∼ ARCH(1)

Suppose yt follows an ARCH(1) process:

yt = σt ut
σt2 = 0.5 + 0.5yt−1
2

ut ∼ IID(0, 1)

Express yt2 as an AR(1) process and compute E(yt2 ).

15 / 53
ARCH(1) Model: yt ∼ ARCH(1)

Let ηt = σt2 (ut2 − 1) = yt2 − σt2 . Then ηt is a martingale difference

sequence and hence a white noise process.

σt2 = 0.5 + 0.5yt−1

2

yt2 − ηt = 0.5 + 0.5yt−1

2

yt2 = 0.5 + 0.5yt−1

2
+ ηt
E(yt2 ) = 0.5 + 0.5E(yt−1
2
)
0.5
E(yt2 ) =
1 − 0.5
E(yt2 ) = 1

16 / 53
ARCH(p) Model: yt ∼ ARCH(p)

yt = σt ut
σt2 = a0 + a1 yt−1
2 2
+ · · · + ap yt−p
a0 > 0, aj ≥ 0
Et−1 (yt ) = σt Et−1 (ut )
=0
Vart−1 (yt ) = σt2 Vart−1 (ut )
= σt2

17 / 53
ARCH(p) Model: yt ∼ ARCH(p)

ηt is a MDS and a white noise series. yt2 follows an AR(p) model.

ηt = (ut2 − 1)σt2
Et−1 (ηt ) = σt2 [Et−1 (ut2 ) − 1]
= σt2 [E(ut2 ) − 1]
=0
yt2 = σt2 ut2
yt2 = σt2 + ηt
yt2 = a0 + a1 yt−1
2 2
+ · · · + ap yt−p + ηt

18 / 53
Forecasting Volatility: yt ∼ ARCH(1)

2
σt+1 = a0 + a1 yt2
2 2
σt+2 = a0 + a1 yt+1
2 2
Et (σt+2 ) = a0 + a1 Et (yt+1 )
2 2
Et (σt+2 ) = a0 + a1 σt+1
2
Et (σt+2 ) = a0 + a1 (a0 + a1 yt2 )
2
Et (σt+2 ) = a0 + a0 a1 + a12 yt2

19 / 53
Forecasting Volatility: yt ∼ ARCH(1)

2 2
σt+k = a0 + a1 yt+k−1
2 2
Et (σt+k ) = a0 + a1 Et (yt+k−1 )
2 2
Et (σt+k ) = a0 + a1 Et (σt+k−1 )
(k−1)
2 a0 (1 − a1 ) (k−1) 2
Et (σt+k )= + a1 σt+1
1 − a1
2 a0 (1 − a1k )
Et (σt+k )= + a1k yt2
1 − a1

20 / 53
Forecasting Volatility: yt ∼ ARCH(1)

Suppose yt follows an ARCH(1) process:

yt = σt ut
σt2 = 0.5 + 0.5yt−1
2

ut ∼ IID(0, 1)
2 ) and E (σ 2 ). The final answer should be a
Compute Et (σt+2 t t+k
function of yt and/or σt2 .
2

21 / 53
Forecasting Volatility: yt ∼ ARCH(1)

2 2
σt+2 = 0.5 + 0.5yt+1
2 2
Et (σt+2 ) = 0.5 + 0.5Et (yt+1 )
2 2
Et (σt+2 ) = 0.5 + 0.5σt+1
2
Et (σt+2 ) = 0.5 + 0.5(0.5 + 0.5yt2 )
2
Et (σt+2 ) = 0.75 + 0.25yt2
2 2
σt+k = 0.5 + 0.5yt+k−1
2 2
Et (σt+k ) = 0.5 + 0.5Et (yt+k−1 )
2 2
Et (σt+k ) = 0.5 + 0.5Et (σt+k−1 )

2 0.5(1 − (0.5)k )
Et (σt+k )= + 0.5k yt2
1 − 0.5
2
Et (σt+k ) = 1 − (0.5)k + (0.5)k yt2
22 / 53
Generalized Autoregressive Conditional Heteroskedasticity

The GARCH(1, 1) model is defined as follows:

yt = σt ut
σt2 = a0 + a1 yt−1
2 2
+ b1 σt−1
ut ∼ IID(0, 1)
a0 > 0, a1 ≥ 0, b1 ≥ 0
1 > a1 + b1

23 / 53
GARCH(1, 1) Model

Volatility is a weighted average of the squared return yt−1 2 with

2
weight a1 , and realized volatility σt−1 with weight b1 . The remaining
weight 1 − (a1 + b1 ) is on a0 /(1 − (a1 + b1 )), the long run variance.
yt is a MDS and is uncorrelated.

24 / 53
ARMA(1, 1) Representation
We can represent a GARCH(1, 1) model as an ARMA(1, 1) model.
If yt ∼ GARCH(1, 1), then yt2 ∼ ARMA(1, 1).

ηt = σt2 (ut2 − 1)
yt2 = σt2 ut2
yt2 = σt2 + ηt
yt2 = a0 + a1 yt−1
2 2
+ b1 σt−1 + ηt
yt2 = a0 + a1 yt−1
2 2
+ b1 (yt−1 − ηt−1 ) + ηt
yt2 = a0 + (a1 + b1 )yt−1
2
+ ηt − b1 ηt−1
a0
E (yt2 ) =
1 − (a1 + b1 )

25 / 53
GARCH(1, 1) Model

Suppose yt follows an GARCH(1, 1) process:

yt = σt ut
σt2 = 0.5 + 0.25yt−1
2 2
+ 0.25σt−1
ut ∼ IID(0, 1)

Express yt2 as an ARMA(1, 1) process and compute E(yt2 ).

26 / 53
GARCH(1, 1) Model

Let ηt = σt2 (2t − 1) = yt2 − σt2 . Then ηt is a martingale difference

sequence and hence a white noise process.

σt2 = 0.5 + 0.25yt−1

2 2
+ 0.25σt−1
yt2 − ηt = 0.5 + 0.25yt−1
2 2
+ 0.25(yt−1 − ηt−1 )
yt2 = 0.5 + 0.5yt−1
2
+ ηt − 0.25ηt−1
E(yt2 ) = 0.5 + 0.5E(yt−1
2
)
0.5
E(yt2 ) =
1 − 0.5
E(yt2 ) = 1

27 / 53
Generalized Autoregressive Conditional Heteroskedasticity

The GARCH(p, q) model is defined as follows:

yt = σt ut
p
X q
X
σt2 = a0 + 2
ai yt−i + 2
bj σt−j
i=1 j=1

ut ∼ IID(0, 1)
a0 > 0, ai ≥ 0, bk ≥ 0
p
X q
X
1> ai + bj
i=1 j=1

28 / 53
GARCH(p, q) Model

Volatility is a weighted average of squared returns yt−i 2 with weights

2
ai , and realized volatility σt−j with weights bj . The remaining weight
1 − ( pi=1 ai + qj=1 bj ) is on a0 /(1 − ( pi=1 ai + qj=1 bj )), the
P P P P
long run variance. yt is a MDS and is uncorrelated.

29 / 53
Stationarity of GARCH Models

We can represent a GARCH(p, q) model as an ARMA([max p, q], q)

model. If yt ∼ GARCH(p, q), then yt2 ∼ ARMA([max p, q], q).

ηt = σt2 (ut2 − 1)
max
X p,q q
X
yt2 = a0 + 2
(ai + bi )yt−i + ηt − bj ηt−j
i=1 j=1

30 / 53
Stationarity of GARCH Models

This necessary and sufficient condition for stationarity is:

p
X q
X
ai + bj < 1
i=1 j=1

E(yt ) = 0
a0
Var(yt ) =
1 − ( pi=1 ai + qj=1 bj )
P P

This is the unconditional, or long run variance.

31 / 53
Forecasting Volatility: yt ∼ GARCH(1, 1)

2
σt+1 = a0 + a1 yt2 + b1 σt2
2 2 2
σt+2 = a0 + a1 yt+1 + b1 σt+1
2 2 2
Et (σt+2 ) = a0 + a1 Et (yt+1 ) + b1 σt+1
2 2
Et (σt+2 ) = a0 + (a1 + b1 )σt+1
2
Et (σt+2 ) = a0 + (a1 + b1 )(a0 + a1 yt2 + b1 σt2 )
2
Et (σt+2 ) = a0 + a0 a1 + a0 b1 + (a12 + a1 b1 )yt2 + (a1 b1 + b12 )σt2

32 / 53
Forecasting Volatility: yt ∼ GARCH(1, 1)

2 2 2
σt+k = a0 + a1 yt+k−1 + b1 σt+k−1
2 2 2
Et (σt+k ) = a0 + a1 Et (yt+k−1 ) + b1 Et (σt+k−1 )
2 2
Et (σt+k ) = a0 + (a1 + b1 )Et (σt+k−1 )

2 a0 [1 − (a1 + b1 )(k−1) ]
Et (σt+k )= + (a1 + b1 )(k−1) σt+1
2
1 − (a1 + b1 )

2 a0 [1 − (a1 + b1 )k ]
Et (σt+k )= + (a1 + b1 )(k−1) (a1 yt2 + b1 σt2 )
1 − (a1 + b1 )

33 / 53
Forecasting Volatility: yt ∼ GARCH(1, 1)

Suppose yt follows an GARCH(1, 1) process:

yt = σt ut
σt2 = 0.5 + 0.25yt−1
2 2
+ 0.25σt−1
ut ∼ IID(0, 1)
2 ) and E (σ 2 ).The final answer should be a func-
Compute Et (σt+2 t t+k
tion of yt and/or σt2 .
2

34 / 53
Forecasting Volatility: yt ∼ GARCH(1, 1)

2 2 2
σt+2 = 0.5 + 0.25yt+1 + 0.25σt+1
2 2 2
Et (σt+2 ) = 0.5 + 0.25Et (yt+1 ) + 0.25σt+1
2 2
Et (σt+2 ) = 0.5 + (0.25 + 0.25)σt+1
2
Et (σt+2 ) = 0.5 + (0.5)(0.5 + 0.25yt2 + 0.25σt2 )
2
Et (σt+2 ) = 0.75 + 0.125yt2 + 0.125σt2

35 / 53
Forecasting Volatility: yt ∼ GARCH(1, 1)

2 2 2
σt+k = 0.5 + 0.25yt+k−1 + 0.25σt+k−1
2 2 2
Et (σt+k ) = 0.5 + 0.25Et (yt+k−1 ) + 0.25Et (σt+k−1 )
2 2
Et (σt+k ) = 0.5 + (0.25 + 0.25)Et (σt+k−1 )

2 0.5[1 − (0.25 + 0.25)k ]

Et (σt+k )=
1 − (0.25 + 0.25)
+ (0.25 + 0.25)(k−1) (0.25yt2 + 0.25σt2 )

2 0.5[1 − (0.5)k ]
Et (σt+k )= + (0.5)(k−1) (0.25yt2 + 0.25σt2 )
0.5
2
Et (σt+k ) = 1 − (0.5)k + (0.5)(k−1) (0.25yt2 + 0.25σt2 )

36 / 53
Integrated GARCH Model

We assume a1 + · · · + ap + b1 + · · · + bq = 1. Consider the

IGARCH(1, 1) model. The model is stationary, but is not mean
reverting. yt2 has no long run mean.

yt = σt ut
σt2 = a0 + a1 yt−1
2 2
+ b1 σt−1
b 1 = 1 − a1
a0
σt2 = 2
+ (1 − b1 )(yt−1 2
+ b1 yt−2 + b12 yt−3
2
+ ...)
1 − b1

37 / 53
ARMA(p, q) − GARCH(p1 , q1 ) Model

We assume that innovations in an ARMA model follow a GARCH

structure. The time varying conditional mean is µt = Et−1 (rt ). The
time varying conditional variance is σt2 = Vart−1 (rt ).

rt = µ + β1 rt−1 + · · · + βp rt−p + yt + α1 yt−1 + · · · + αq yt−q

yt = σt ut
p1
X q1
X
σt2 = a0 + 2
ai yt−i + 2
bj σt−j
i=1 j=1

38 / 53
Exponential GARCH(1, 1) Model

We use log volatility and capture the asymmetric effects of shocks

on volatility. α < 0 means that volatility rises more in response to
a negative shock, relative to a positive shock

ht = log(σt )
yt = exp(ht )ut
ht = ω + βht−1 + αut−1 + γ(|ut−1 | − E|ut−1 |)

39 / 53
Asymmetric Power GARCH Model

The power index δ exploits the stronger autocorrelation of power

functions of returns, e.g. absolute returns have stronger autocor-
relations than squared returns. The asymmetric effect is captured
by d. Positive values of d indicate a greater response to negative
shocks. a0 , a1 , b > 0, d ∈ (−1, 1), δ ∈ (0, 2], ut ∼ IID(0, 1).

yt = σt ut
σtδ = a0 + a1 (|yt−1 | − dyt−1 )δ + bσt−1
δ

40 / 53
GARCH in Mean (Risk Premium)

The first term is the conditional expected return, which is related to

the underlying risk, and the second term is the volatility around the
return. The risk premium is expressed in terms of either variance or
standard deviation.

rt = g (σt ) + yt
yt = σt ut
σt2 = a0 + a1 yt−1
2 2
+ b1 σt−1
g (x) = θ0 + θ1 x (linear in variance)
√
g (x) = θ0 + θ1 x (linear in standard deviation)

41 / 53
Stochastic Volatility

Volatility is modeled as a latent state space variable not explicitly

linked with observed returns. ut and et are independent and g (.) is
a known function.

yt = exp(ht )ut
p
X
ht = c + bj ht−j + et
j=1

ut ∼ IID(0, 1)
et ∼ IID(0, σe2 )

42 / 53
Realized Volatility

Realized Volatility (RV) is a nonparametric estimator of the variance

that is computed using high frequency data. For example, intraday
returns can be used to construct an estimate of the daily volatility.
Alternatively, daily returns can be used to constructed an estimate
of the monthly volatility. Suppose the log price pt is continuously
available, and is driven by a standard Weiner process with constant
mean and variance.

dpt = µdt + σdWt

43 / 53
Realized Volatility

Suppose we wish to estimate the daily volatility using intraday re-

turns. RV is estimated by sampling pt throughout the trading day.
Suppose prices on day t are sampled on a regular grid of (m + 1)
points: 0, 1, . . . , m. Let pi,t denote the i th observation of the log
price. The m-sample RV on day t is defined as:

m
(m)
X
RVt = (pi,t − pi−1,t )2
i=1
m
(m)
X
2
RVt = ri,t
i=1

44 / 53
Realized Volatility

(m)
RVt is asymptotically unbiased and a consistent estimator of σ 2 .

(m) µ2
E[RVt ]= + σ2
m
(m)
lim E[RVt ] = σ2
m→∞

(m) µ2 σ 2 σ4
Var[RVt ]=4 2
+2
m m
(m)
lim Var[RVt ]=0
m→∞

45 / 53
Realized Volatility

Suppose the price process has a time varying drift and stochastic
(m)
volatility. Then RVt is a consistent estimate of the integrated, or
average variance over the measurement interval.

dpt = µt dt + σt dWt
Z t+1
(m) p
lim [RVt ] − → σs2 ds
m→∞ t

46 / 53
Implementing Realized Volatility

Observed returns are contaminated by noise. Traded prices are only

observed at the bid and the ask. This produces a bid ask bounce
where consecutive prices oscillate between the two. Suppose ob-
∗ , and an independent
served prices consist of true efficient price pi,t
mean zero shock νi,t .

∗
pi,t = pi,t + νi,t

47 / 53
Implementing Realized Volatility

The observed return ri,t can be decomposed into the actual unob-
∗ and an independent noise term η
served return ri,t i,t = νi,t − νi,t−1 .
ηi,t is a MA(1) and is serially correlated.

∗ ∗
pi,t − pi,t−1 = (pi,t + νi,t ) − (pi,t−1 + νi,t−1 )
∗ ∗
pi,t − pi,t−1 = (pi,t − pi,t−1 ) + (νi,t − νi,t−1 )
∗
ri,t = ri,t + ηi,t

48 / 53
Implementing Realized Volatility

The estimate of RV is biased upward. τ 2 is the variance of ηi,t and

RV
dt is the realized volatility if efficient returns could be observed.
The bias is increasing in m.
m
(m)
X
2
RVt = ri,t
i=1
m
(m)
X
∗
RVt = (ri,t + ηi,t )2
i=1
m
(m)
X
∗2 ∗ 2
RVt = (ri,t + 2ri,t ηi,t + ηi,t )
i=1
(m) dt + mτ 2
RVt = RV

49 / 53
Implementing Realized Volatility

One solution is to use sparse sampling (i.e. not using all observed
prices). Another solution is to filter the data using an MA(1) model.
In this case, RV is computed using the error term ˆi,t from the
following model:

ri,t = θi,t−1 + i,t

50 / 53
Implementing Realized Volatility

Most markets do not trading 24 hours a day or 7 days a week. In

practice, we augment the high frequency returns using the squared
close to open (CtO) return in order to construct an estimate of the
total close to close (CtC) variance.
(m) 2 (m)
RVCtC ,t = rCtO,t + RVt

51 / 53
Modeling Realized Volatility

If RV is observable, it can be modeled using standard ARMA mod-

els. The heterogeneous autoregression (HAR) model captures the
dynamics of RV in a parsimonious model. RV is a function of the
RV of the previous day, the average RV of the previous week, and
the average RV of the previous month. The HAR model is some-
times estimated in logs.

52 / 53
Modeling Realized Volatility

RVt = φ0 + φ1 RVt−1 + φ5 RV t−5 + φ21 RV t−21 + ut

log RVt = φ0 + φ1 log RVt−1 + φ5 log RV t−5 + φ21 log RV t−21 + ut
5
1X
RV t−5 = RVt−i
5
i=1
21
1 X
RV t−21 = RVt−i
21
i=1

53 / 53