Lecture 3

Risk and Return

Dr HK Pradhan
Professor of Finance & Economics
XLRI Jamshedpur
 Measuring Risk

• Risk is measured in the volatility of assets prices

• How much an asset is expected to change in value?
• Market risk can be discerned from the past behavior of the
market prices
• Past behavior would reflect future (or expected return),
given all other factors
 Know your data
Data frequency
• Monthly, weekly, daily, intra-day
• Plot and understand levels and return
• Compute measures of volatility (risk)
• Understand their underlying distribution of the data
– Normal, Log normal, Weibull, Exponential,
Uniform
Euro/Per USD rate: Sept 2007-April 2008 Euro/Per USD rate: June 2008-August 2008

Euro/Per USD rate: 23rd August 2008

Intra-Day Data
 Return
• Risk is measured from return
• Let us define daily return on an asset as follows:
• The change in the logarithm of the daily closing price of
the asset  S 
rt = Ln ( S t ) − Ln ( S t −1 ) = Ln  t 
 S t −1 
• The arithmetic return is instead defined as
rt +1 = (St +1 − St ) St or rt +1 = St +1 St − 1

• Log return & percentage return are close on a daily basis

 Advantage of Log Return
• From the log return we can easily calculate the
compounded return over K-days as the sum of daily
return
K K
Rt +1:t + K = ln(St + K ) − ln(St ) =  ln(St + k ) − ln(St + k −1 ) =  Rt + k
k =1 k =1
K
ln(St + K ) / ln(St ) =  Rt + k
k =1

• Called the geometric average of single period return

• Log return does not consider negative prices

 Distribution Assumptions
• Log returns are assumed to follow somewhat normal distribution
(i.e. the underlying asset prices follow a lognormal distribution)
• A random variable X is said to have a lognormal distribution if its
natural logarithm, Y = ln(X), has a normal distribution

 
ln
St   N (  * t ,  t )  = mean &  = sd
i.e.  
 S0 
 Distribution
properties by its
four moments (all
unconditional)

First moment: mean, ()

Second moment: standard deviation, ()
Third moment: Skewness, (skew)
Fourth moment: Kurtosis (kurt)

Distribution properties have significant

implications on risk
 Mean
• Define Pi as the price of an asset at end of day i
• Define Return as ri= ln(Pi/Pi-1) or continuously
compounded return over multiple-day
• The standard estimates of mean from n observations is:

n
1
r =  ri is unconditional Mean of
n i =1 the return

• Provides an expected return by holding an asset over the

time horizon
 Standard Deviation
• Xi as the price of an asset at end of day i
• Return as ui= ln(Xi/Xi-1) or continuously compounded
return over multiple-day
• The standard estimate of volatility from n observations is:
n
1
n =
2

n − 1 i =1
(ri − r ) 2 is unconditional sd of the
return

1 n is unconditional Mean of
r =  ri the return
n i =1
• It is unconditional as no weight attached to past
observations
 Skewness
• Skewness: Skewness is a
measure of the asymmetry
of the distribution
1 n

n i =1
(ri − r ) 3

sk =
3
1 n
and the sample variance defined as  =  (ri − r ) 2
2
n
n i =1

• Skewness is a measure of symmetry, a –ve skew implies

return data in tail
• We are interested in measuring tail risk of the distribution
 Kutosis
• Kurtosis relates to the
balance between the tails of
the distribution versus its
center n
1

n i =1
(ri − r ) 4

kurt =
 4

• Leptokurtic distributions have heavy concentrations around the

mean
• What is the implication of excess Kurtosis??
• If a high kurtosis when combined with –ve skewness can have
significant risk implications
 Are asset return (log) normal?
• If asset returns were normally
distributed, then the mean and
volatility would fully characterize
risk.
• Portfolio returns exhibit non-
normality, no single measure
completely describes portfolio
risk.
• Often characterized by extreme 0.0004000 45

values, resulting in a heavy tail 0.0003500 40

35
losses 0.0003000
30
• Fatter tails mean a higher 0.0002500
25
0.0002000
probability of large losses than 0.0001500
20
15
the normal distribution would 0.0001000 10
suggest 0.0000500 5
0.0000000 0
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
1
6

101
 Measuring Volatility
– Volatility of asset returns
• How much an asset change in value change? Represented
by the standard deviation
n
1
 =
2
n 
n − 1 i =1
(ri − r ) 2

• There are several ways to measure volatility

– Realized volatility (Sd, EWMA, GARCH family)
– Implied Volatility (VIX)
• Measures such as downside volatility also used to assess
asymetry
– Different asset classes exhibit different volatility parameters
– Their volatility change over time
Volatility
 Variance Autocorrelations
0.15

0.1

Let us approximate the variance as: 0.05

= r
n
1 n
0

if r =  ri  0  2 2 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

-0.05

n i =1 t ,n i =1 n −i -0.1

-0.15

We can measure some kind of variance autocorrelations.

Corr(rt 2+1 , rt 2+1− )  0, for small 
• Equity and equity indices often display negative correlation
between variance and returns.
• This often termed the leveraged effect, a drop in stock price
will increase the leverage of the firm as long as debt stays
constant
 Autocorrelations of Squared Returns
(autocorrelation of variance)
• Squared returns show persistence, 0.4

implying the persistence in the 0.35

0.3

scale of fluctuations 0.25

0.2

– This can be known as volatility 0.15

persistence: large price 0.1

0.05

movements followed by another 0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
large price movements, not -0.05

necessarily in the same Volatility Clustering

directions
– Some people use absolute value (
Corr r 2 t , r 2 t −  0 )
of the increments or even runs for small 
test
 Time Varying Volatility (Risk)
0.08
• Asset prices display 0.06
time varying
variance 0.04

0.02

0
• This has important -0.02
1 25 49 73 97 121 145 169 193 217 241 265 289 313 337 361 385 409 433 457 481 505

implications on risk -0.04

management -0.06

-0.08

• Capturing this time varying parameters is key to volatility

computations
• Short periods vs long periods, create different amplitude
Unconditional vs Conditional Volatility
Unconditional Volatility(mean return is zero)

= r
1 n n
if r =  ri = 0
n i =1
 2
t ,n
2
i =1 n −i

Conditional Volatility (with assigning weights)

n

 
n
 t2,n = i rn2−i where i =1
i =1
i =1

0 r 2 t −1 + 1 r 2 t −2 + .... + n −2 r 2 t −n −1 + n −1 r 2 t −n
 t,n =
0 + 1 + .... + n −2 + n −1
When we measure unconditional variance, we assume
equal weights to past returns
Exponentially Weighted Moving Average
(EWMA)
RiskMetrics uses an exponentially weighted moving average
(EWMA) model, the weights assigned to the u2 decline
exponentially as we move backwards through time

0 r 2 t −1 + 1 r 2 t −2 + .... + n −2 r 2 t −n −1 + n −1 r 2 t −n
 t,n = 0 1 n−2 n −1
 +  + .... +  + 
……………….with some manipulations

 n2 =  n2−1 + (1 −  )rn2−1
Conditional volatility, measured by a combination of past
squared return and a recursive term measuring past variance
 Exponentially Declining Weights
• Recent volatility is being 1
emphasized 0.9
0.8
• Less emphasis on long 0.7

memory 0.6
0.5
• Weights quickly dimish 0.4
0.3
as we move past 0.2
0.1

0.94a  tolerance level

0
1 35 69 103 137 171 205 239 273 307 341 375 409 443 477

•
Can we find out the optimal decay factor?
 Volatility estimation: exponential
moving averages reflect better picture
4.00 Daily percentage changes DM/$
2.00

0.00
-2.00

Volatility – Exponential moving average

2.00
Weights for daily observations
Volatility –
1.50 Simple moving 4.0%

average 3.0%
2.0%
1.00 1.0%
0.0%
250 150 50
0.50
days of historical data

1999 2000 2001

 Time Scaling of Volatility
• Risk increases with time, but not linearly
• Uncertainty increases with the square root of time
• Use the number of trading days as opposed to actual
days to scale volatility
• 1 Week volatility = daily volatility 5 = daily volatility * 2.24
• 1 month Vol = 1 day Vol 21 = daily volatility * 4.58
• 1-Year Vol = 1 day Vol 252 = daily volatility * 15.87
Volatility Forecasting using EWMA
– Riskmetrics λ= 0.94 for daily volatility forecasting
– Today’s Variance depends on yesterday’s estimate
of the variance and yesterday’s return^2
 n2 =  n2−1 + (1 −  )rn2−1
– Example
– Volatility in last period 0.13456%
– Change in the market variable(return) -0.07719%

 n = 0.08763% = 0.94 x 0.13456% + 0.06 x 0.07719^ 2

– Estimated value of volatility 0.08763%
 Questions
• If stock price is 50 and the annual volatility
is 30%, what is the one standard deviation
move in one week of the stock?
• Putting less weight to (say, = 0.85 ), what
will be the implications on volatility?
Generalized Autoregressive Conditional
Heteroskedastic(GARCH) Model

• Engle(1982) assumed that the conditional

variance of returns is not constant, and that
asset returns may show alternative periods
on higher or lower volatility (termed as
volatility clustering), with time varying
parameters
• GARCH models are used when the
variance of the error term is not constant.
That is, the error term is heteroskedastic.
• Many variations of GARCH have
emerged:
 Engle’s ARCH/GARCH
Let us recall our Riskmetrics EWMA volatility calculation:

0 r 2 t −1 + 1 r 2 t − 2 + .... + n − 2 r 2 t − n −1 + n −1 r 2 t − n
 t2, n =
0 + 1 + .... + n − 2 + n −1
n

 
n
 t2,n = i rn2−i where i =1
i =1
i =1
There is a long run variance component in the volatility process
n
=  VL + i =1  r where  +  i = 1
n
 2
t ,n
2
i n −i
i =1

 t2,n = L + i =1 i rn2−i
n
with  VL =  ,

GARCH  t2 =  +  ut2−1 +  t2−1 EWMA  n2 =  n2−1 + (1 −  )rn2−1

with  +  +  =1
 GARCH (1,1) Model
GARCH (1,1) Model therefore represents he following

 =  + u
t
2 2
t −1 +  2
t −1

 2
t −1
volatility of the previous day

 (=  VL ) called the persistence and not greater than one

 +  +  = 1 are weights which sum to one
In GARCH (1,1) Model the variance  t −1is calculated from
2

a long run average variance rate V as well as from

recent variance & return L
Estimate the long run variance V

 =  + u
t
2 2
t −1 +  2
t −1

 +  +  =1
or  =1 −  − 

and  = V or V =
1− − 
where λ is the mean reversion parameter
 EWMA vs GARCH

2 2
EWMA t =  n −1 + (1 −  )u n −1

GARCH(1,1)  t =  +  t2−1 +  ut2−1

 in GARCH is equivalent to decay factor ( ) in EWMA

 in GARCH is equivalent to (1- ) in EWMA (AR term)
EWMA is a special case of GARCH where  = 0
 Estimating GARCH
• Basically we are estimating the variance of our return ri
from t observations of return, when the underlying
distribution is normal.
• Assuming that the mean is zero and the variance  2 the
likelihood of ri being observed is the probability density
function of x, when x=

1  rt 2 
lt = exp
 − 2 2 

2 t2  t 
Which is the pdf under a normal distribution as
 Estimating GARCH
• The best estimate of is volatility  t2 = v the value
that maximizes the expression (likelihood function)
below, by changing the GARCH parameters:
𝑚
1 −𝑢𝑖 2
𝑀𝑎𝑥 ෑ exp( )
2𝜋𝑣 2𝑣
𝑖=1

Use solver to maximize the above expression given constraints

Look at the distribution?
 Parameters Relations
• The relations among the various parameters
of the GARCH model:
 = V
( +  +  ) = 1
 = (1 −  −  )
• Typically the values  +  = 1 so that
mean reversions takes place
• If  +  = 1 then we have a situation of
mean fleeing, as  becomes negative
 Question
• SENSEX closes today at 17,862, up from the previous day
closing of 17,777. The historical volatility of the Sensex index
was estimated at 1.4% per day.
• GARCH (1,1) parameters estimated as follws
a = 0.03, b = 0.95 & w = 0.000002
• What is the estimate of new annualized volatility for today?

• Return =ln(17,862/17,777)= 0.0100478

• Variance== 0.000002 +0.03*0.0100478^2+0.95*0.014^2 = 0.00019

 t2 =  +  ut2−1 +  t2−1

• Daily Volatility = 1.3829% ( 0.00019 )

• Annualized Volatility = 21.95% (1.3829% 252 )
 Volatility Forecasting in GARCH
 2 t =  +  ut2−1 +  t2−1
 2
t = V +  u 2
t −1 +  2
t −1

 2
t = (1 −  −  ) V +  u 2
t −1 +  2
t −1

See Hull Page 226 for the solutions..

 2
t =V + ( +  ) (t 2
t −1 −V )
You only need the volatility today to forecast
tomorrow’s volatility, given the long run volatility
 Multi-Period Risk Calculations
• If the returns at horizon K are not normally
distributed?
• We have the GARCH(1,1) model where

( )
K
= K +  ( +  )
k −1
 2
t +1:t + K
2 2
t +1 −   K
2 2
t +1
k =1

• As the variance does mean revert and it

therefore does not scale by the horizon K, and
again the returns over the next K days are not
normally distributed.
37
 What do you expect in
tomorrow’s variance?
 2
t = (1 −  −  )V +  u 2
t −1 +  2
t −1

−−−−−−−−
 2
t = V +  (u 2
t −1 − V ) +  ( 2
t −1 −V )
Tomorrow’s variance is the long-run average variance with
something added (subtracted) if today’s squared return is above
(below) its long-run average, and something added (subtracted)
if today’s variance is above (below) its long-run average.
 Implied Volatility
• Volatility computed from historical data:
realized volatility or statistical volatility.
• Implied volatility is derived from option
prices: as predictive volatility of option’s
future prices (derived from Black-Sholes
option pricing model)
 Implied Volatility
Derived from Black-Scholes(1973) model
C = [S N (d ) − e −rT X N (d − t )]
where, N(d) is the cumulative distribution function of a
standard normal calculated at d
1 2
ln( S / X ) + (r +  ) T
d= 2
 T
• Implied volatility depended on the maturity of the option
• Implied volatility differs with the strike price of option
(creating volatility smiles or skews) so one can take derive
a measure of weighted average volatility
 Implied Volatility Measures
• Historical volatility: Backward looking
• Implied volatility: Forward looking
• Implied volatility builds on future (expected)
values of variables, therefore can be used as
the indication of future volatility
• However traders prefer historical volatility
measyres
 Caution:
• Historical data can be noisy
• Historical data can be misleading
• if a market is maturing, if there is regime switch, policy
shifts occurring swiftly over the measurement period,
or are affected by specific events that had a large
impact on asset prices (9/11, 26/11, global financial
crisis)
• Illiquidity can have problems affect smooth or
continuous historical data
• say for example, prices bonds on certain days not
available
• We analysed base don one sample from the
underlying distribution and getting more
observations may change the shape of the
estimated densities
 You could try these calculations
• Get at least two asset prices (stock prices, exchange rate,
prices of gold, oil, etc)
• Identify the distribution properties of their series
• Interpret mean, sd, Kurtosis & Skewness
• Find out volatility clustering in data
• Compute autocorrelations of return^2
• Draw conclusions on risk pattern of asset prices
• What economic logic, risk implications
Thank You