18.

S096 Problem Set 5 Fall 2013

Volatility Modeling
Due Date: 10/29/2013

1. Sample Estimators of Diffusion Process Volatility and Drift

Let {Xt } be the price of a financial security that follows a geometric
Brownian motion process:
dX(t)
X(t) = µ∗ dt + σdW (t),
where

• σ > 0, is the volatility parameter

• µ∗ ∈ (−∞, ∞), is the drift parameter
• dX(t) is the infinitesimal increment in price.
• dW (t) is the increment of a standard Wiener Process, i.e, in-
finitesimal increments W (t + dt) − W (t) are i.i.d. Normal random
variables with zero mean and variance equal to ‘dt’.

Consider sampling values of the price process over a fixed time period
t ∈ [0, T ], at equal time increments h = T /n. Define
ti = i × h, i = 0, 1, . . . , n
Xi = X(ti ), i = 0, 1, . . . , n
Yi = log(Xi /Xi−1 ), i = 1, 2, . . . , n
Accept as given that:
Yi are i.i.d. N (µ · h, σ 2 · h) random variables,
(this is proven with the theory of diffusion processes/stochastic differ-
ential equations, with µ = µ∗ − 12 σ 2 ).

1(a) Prove that the Maximum-Likelihood Estimates: µ̂ and σ̂ for a

sample: y1 , y2 , . . . , yn , are given by
µ̂ = n1 ni=1 Yi
P

σ̂ 2 = n1 ni=1 (Yi − µ̂)2

P

1(b) Derive the distribution of µ̂ ; give specific formulas for the expec-
tation and variance of µ̂.

1
 1(c) Derive the distribution of σ̂ 2 ; give specific formulas for the ex-
pectation and variance of σ̂ 2 .
1(d) Consider increasing the number of increments n on the fixed time
period [0, T ], and let µ̂n and σ̂n2 be the corresponding MLEs of
the parameters. Determine the limiting distributions of µ̂n and
σ̂n2 .
1(e) A sequence of estimators θˆn for a parameter θ, is weakly consis-
tent if

lim P r(|θ̂n − θ|) = 0.

n→∞

For each of µ̂n and σ̂n2 , determine whether the sequence of esti-
mators is weakly consistent.

2. Consider the same process as in problem 1, but now, for fixed values
of µ and σ, consider sampling n values of the price process over a fixed
time perioPd t ∈ [0, T ], at variable increments hi > 0, i = 1, 2, . . . , n,
such that ni=1 hi = T. Define
ti = ij=1 hj , i = 0, 1, . . . , n
P

Xi = X(ti ), i = 0, 1, . . . , n
Yi = log(Xi /Xi−1 ), i = 1, 2, . . . , n
Accept as given that:
Yi are i.i.d. N (µ · hi , σ 2 · hi ) random variables,
(this is proven with the theory of diffusion processes/stochastic differ-
ential equations).

2(a) Derive the MLE for µ andPn its distribution for a fixed set of sam-
pling increments {hi } : i=1 hi = T .
2(b) Derive the MLE for σ 2 and P its distribution for a fixed set of
sampling increments {hi } : ni=1 hi = T .
2(c) If limited to sampling n + 1 price points of {Xt }, (including X0
and XT ) prove that
• For estimating σ 2 , sampling, the ML estimators vary with
the increment spacing, but the variance of these estimators
are all equal, regardless of the increment spacing.
• For estimating µ, all ML estimators are the same and have
the same variance, regardless of the increment spacing.

2
3. ARCH(1) Model Properties
Let yt = log(St /St−1 ) be the discrete returns of the price of a secu-
rity/portfolio {St , t = 1, 2, . . .}, and supppose that yt ∼ ARCH(1),
i.e.
y t = µ t + t ,
where µt is the mean return, conditional on Ft−1 , the information
available up to time (t − 1) and
t = Z t σ t ,
where Zt iid with E[Zt ] = 0, and var[Zt ] = 1, and
σt2 = α0 + α1 2t−1 .
Additionally, suppose that E[Zt3 ] = 0, and E[Zt4 ] = κ. (The parameter
κ is the Kurtosis of the Zt distribution with unit variance; if Zt is
Gaussian/normal, then κ = 3.
Prove that:

3(a) E[2t ] = α0 /(1 − α1 )

3(b) E[3t ] = 0
κα2 (1+α )
1
3(c) E[t4 ] = (1−α0 )(1−κα 2)
1 1
3(d) What constraints on α0 , α1 must be made in (c), to maintain
4-th order stationarity (bounded).
3(e) The kurtosis of t is
κ = E[4t ]/(E[2t ])2 .
(The fourth moment is normalized to be scale-free). If the distri-
bution Zt is Gaussian/normal (i.e., the scaled, conditional error
distribution of t ), does the unconditional distribution of t , have
a higher than that of the Gaussian distribution, ( i.e., heavier
tails)?

3
4. Using Daily Open/High/Low/Close Data on the S&P500 Index from
2006-20012 , annual sample variances were computed of changes in the
log index value of the daily Close.
The following table gives the annual sample variances, day counts, and
annualized volatilities
Annual Sample Variances of Logarithmic Returns:

daily.variance days volatility

2006 3.981351e-05 251 0.09996595
2007 1.018599e-04 251 0.15989632
2008 6.677100e-04 253 0.41101171
2009 2.950132e-04 252 0.27265972
2010 1.294545e-04 252 0.18061706
2011 2.164385e-04 252 0.23354335
2012 6.459046e-05 250 0.12707327

The differences in the sample variances and volatilities appears quite

large for some years. Are the year-by-year differences significant?
To address this question, consider modeling the returns for any given
year as a simple random sample from a Gaussian distribution:
{y1 , y2 , . . . , yn } : yi i.i.d. N (µ, σ 2 ).
√
The table gives values of σ̂ 2 , n, and nσ̂, where
1 Pn
σ̂ 2 = n−1 2
i=1 (yi − µ̂) ,

µ̂ = n1 ni=1 yi .
P
with

4(a) Under the Gaussian model, given n, µ, σ 2 , prove that the distri-
bution of σ̂ 2 is
σ2
σ̂ 2 ∼ n−1 × χ2n−1 .
that is, a scaled Chi-square distribution with degrees of freedom
σ2
equal to (n − 1), and scale factor equal to n−1
4(b) Statistical methodology defines confidence intervals for unknown
parameters by computing a likely interval for the parameter es-
timate given the unknown parameter, and then inverting the in-
terval to correspond to the parameter instead of the estimate.
For a 95% confidence interval (two-sided), the development is as
follows:

4
 • With 95% probability, the χ2n−1 random variable will fall
within the interval from the 0.025 percentile, to the 0.975
percentile, i.e.,
P r(q0.025 < χ2n−1 ≤ q0.975 ) = 0.975 − 0.025 = 0.95
where
P r(χ2n−1 ≤ q0.025 ) = 0.025
P r(χ2n−1 ≤ q0.975 ) = 0.975
• Replacing the random variable χ2n−1 with ( n−1 σ2
)σ̂ 2 gives:
n−1 2
P r(q0.025 < ( σ2 )σ̂ ≤ q0.975 ) = 0.95
which can be inverted to
−1) 2 (n−1)
P r(σ̂ 2 (n 2
q0.975 ≤ σ ≤ σ̂ q0.025 ) = 0.95
The following table gives the percentiles of the Chi-square distri-
butions for degrees of freedom ranging from 249 to 252 (one less
than the annual day counts).
df q0.025 q0.975 ll.factor ul.factor
249 207.1856 294.6008 0.8452115 1.201821
250 208.0978 295.6886 0.8454840 1.201358
251 209.0102 296.7763 0.8457550 1.200898
252 209.9227 297.8637 0.8460245 1.200442
The last two columns are
ll.f actor = qn0.−975
1
and ul.f actor = n−1
q0.025
which when multiplied by the unbiased sample estimate σ̂ 2 , define
the confidence interval for σ 2 .
• Using data for 2008, compute the two-sided 95% confidence
interval for σ 2 , based on daily log returns.
• Express
√ the interval in terms of the annualized volatility
( 253σ). Does the sample annual volatility for any other
year fall in the confidence interval for 2008?
4(c) The return variance / volatility varies considerably from year to
year. To evaluate the statistical significance of the difference in
values for any two years, we can use the F -Distribution. Consider
2007 and 2008.
Under the assumption (i.e., a null hypothesis H0 ) of Gaussian/normal
daily returns and that the variances of the returns are constant/
the same for all days in the two years it follows from 4(a) that:
−1)
X = ( n2007
σ2
2
)σ̂2007 ∼ χ2df1 , where dfX = (n2007 − 1)
−1)
Y = ( n2008
σ2
2
)σ̂2008 ∼ χ2df2 , where dfY = (n2008 − 1)

5
and X and Y are independent random variables. The statistic
Y /dfY σ̂ 2
S= X/dfX =( σ̂2008
2 )
2007
has the F -Distribution with degrees of freedom dfY for the nu-
merator and dfX for the denominator. (Verify by looking up the
definition of the F -distribution.)
Under the null hypothesis, the numerator and denominator of
S are estimates of the same return variance. Their ratio varies
about 1 due to the independent variation in the numerator and
denominator of scaled Chi-squared random variables.
The methodology of hypothesis testing in statistics uses the fact
that the test statistic has a known distribution under the null
hypothesis. The null hypothesis is accepted / rejected so long
as the test statistic is not extreme. We choose a test α-level,
the probability of (falsely) rejecting the null hypothesis if true,
say α = 0.05. From this, extreme ranges of the test statistic are
defined that occur with probability α when the null hypothesis is
true. For α = 0.05, a two − sided alternative is considered using
q0.025 and q0.975 , the percentiles of the F distribution given by:
P r(FdfY ,dfX < q0.025 ) = 0.025
P r(FdfY ,dfX < q0.975 ) = 0.975
The null hypothesis is accepted if
q0.025 < S < q0.975
From the package R, we provide the percentiles of the F -distribution
when df1 = n2008 − 1, and df2 = n2007 − 1 :

> qf(0.025, df1=252, df2=250)

[1] 0.7804173
> qf(0.975, df1=252, df2=250)
[1] 1.281525
so the null hypothesis is accepted if
q0.025 = 0.7804 < S < 1.2815 = q0.975
• Compute the test statistic S = S0 for testing the daily return
variance for 2008 is equal to the daily return variance for
2007.
• Given the value of the test statistic S0 , determine the α−level
at which the null hypothesis is on the boundary of being just
accepted/rejected.

6
 (This level is called the P -value of the test statistic. Report-
ing a test statistic’s P -value provides evidence concerning
for/against the test null hypothesis which can be provided
without having to specify an α-level.)
• Repeat the previous two questions, for testing the equality
of the return variance for 2008 to that for 2006. (Note: the
degrees of freedom for 2006 are the same as those for 2007 so
the same F distribution is applicable)

7
MIT OpenCourseWare
http://ocw.mit.edu

18.S096 Mathematical Applications in Financial Industry

Fall 2013

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.