Dynamic Portfolio Choice Financial Econometrics

Review: DP and Econometrics

Leonid Kogan

MIT, Sloan

15.450, Fall 2010

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 1 / 22
Dynamic Portfolio Choice Financial Econometrics

Outline

1 Dynamic Portfolio Choice

2 Financial Econometrics

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 2 / 22
Dynamic Portfolio Choice Financial Econometrics

Outline

1 Dynamic Portfolio Choice

2 Financial Econometrics

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 3 / 22
Dynamic Portfolio Choice Financial Econometrics

Portfolio Choice: Static Approach

In models in which all options are redundant (e.g., Black-Scholes), dynamic

portfolio choice is relatively easy.
Solve a static problem:
Find the best state contingent payoff (under given utility) which is
budget-feasible.
Replicate the chosen payoff using dynamic trading in available assets.
Merton’s solution: CRRA utility with risk aversion γ, Black-Scholes model:
µ−r
φ�t =
γσ2
Myopic portfolio is optimal.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 4 / 22
Dynamic Portfolio Choice Financial Econometrics

Problem

Consider the Black-Scholes framework with parameters r , µ, and σ. Your

objective is to find an optimal investment strategy maximizing the expected
utility of terminal portfolio value
 
1
E0 (WT )1−γ
1−γ

subject to a lower bound on terminal wealth:

WT > W

1 Using the static approach, express the optimal terminal wealth as a function of
the SPD.
2 Show that one can implement the optimal strategy using European options on
the stock.
3 (*) Implement the optimal strategy using dynamic trading.

c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 5 / 22

Dynamic Portfolio Choice Financial Econometrics

DP

Dynamic programming principle.

Bellman equation.
Controlled Markov processes. Problem formulation.
Key examples: portfolio choice with time-varying moments of returns;
American option pricing.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 6 / 22
Dynamic Portfolio Choice Financial Econometrics

Outline

1 Dynamic Portfolio Choice

2 Financial Econometrics

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 7 / 22
Dynamic Portfolio Choice Financial Econometrics

Parameter Estimation
GMM

Estimate parameters using moment restrictions.

If the true distribution satisfies

E[f (xt , θ0 )] = 0, � 0 if θ �= θ0
E[f (xt , θ)] =

estimate θ0 using a sample analog of the population moments

1�
T
E �)] ≡
� [ f ( xt , θ f (xt , θ
�) = 0
T
t =1

Which moments to choose for estimation?

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 8 / 22
Dynamic Portfolio Choice Financial Econometrics

Parameter Estimation
MLE

MLE tells us that a particular choice of moments would work and would
produce the most precise estimates.
For IID observations, MLE prescribes estimating parameters as

θ � [ln p(xt , θ)]

� = arg max E
θ

In moment form, this implies

�
T
∂ ln p(x , θ)
=0
t =1
∂θ

MLE is a special case of GMM with a particular choice of moments, based on

the pdf.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 9 / 22
Dynamic Portfolio Choice Financial Econometrics

MLE for Dependent Observations

MLE approach works even if observations are dependent.
Consider a time series xt , xt +1 , ... and assume that the distribution of xt +1
depends only on L lags: xt , ..., xt +1−L .
Log likelihood conditional on the first L observations:
�T −1 �
�
� = arg max L(θ) = arg max
θ ln p(xt +1 |xt , ..., xt +1−L ; θ)
θ θ
t =L

AR(p) (AutoRegressive) time series model with IID Gaussian errors:

IID
xt +1 = a0 + a1 xt + ...ap xt +1−p + εt +1 , εt +1 ∼ N(0, σ2 )

Construct log likelihood:

�
T −1 √ (xt +1 − a0 − a1 xt − ...ap xt +1−p )2
L(θ) = − ln 2πσ2 −
2σ2
t =p

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 10 / 22
Dynamic Portfolio Choice Financial Econometrics

Parameter Estimation
Iterated expectations

Another approach to forming moment conditions is to use iterated

expectations.
For example, consider a linear model

yt = b0 + b1 xt + εt

Assume that
E[εt |xt ] = 0
Using iterated expectations, we can form two moments

E[(yt − b0 − b1 xt ) × 1] = 0
E[(yt − b0 − b1 xt ) × xt ] = 0

Recover standard OLS formulas.

εt could be heteroscedastic, our estimator is still valid since our moment
restrictions are valid.
�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 11 / 22
Dynamic Portfolio Choice Financial Econometrics

Parameter Estimation
QMLE

QMLE helps formulate moment conditions when the exact form of the pdf is
not known.
Pretend that errors are Gaussian and use MLE to form moment restrictions.
Make sure that the moment restrictions we have derived are valid, based on
what we know about the model.
Intuition: we may only need limited information, e.g., a couple of moments, to
estimate the parameters. No need to know the entire distribution.
QMLE is a valid (consistent) approach, less precise than MLE but more
robust.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 12 / 22
Dynamic Portfolio Choice Financial Econometrics

Example: Interest Rate Model

Iterated expectations

Interest rate model:

rt +1 = a0 + a1 rt + εt +1 , E(εt +1 |rt ) = 0, E(ε2t +1 |rt ) = b0 + b1 rt

GMM with moment conditions derived using iterated expectations

E [(rt +1 − a0 − a1 rt ) × 1] = 0
E [(rt +1 − a0 − a1 rt ) × rt ] = 0
�� �
(rt +1 − a0 − a1 rt )2 − b0 − b1 rt × 1 = 0
�
E
�� �
E (rt +1 − a0 − a1 rt )2 − b0 − b1 rt × rt = 0
�

(a0 , a1 ) can be estimated from the first pair of moment conditions. Equivalent
to OLS, ignore information about second moment.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 13 / 22
Dynamic Portfolio Choice Financial Econometrics

Example: Interest Rate Model

QMLE

Treat εt as Gaussian N(0, b0 + b1 rt −1 ).

Construct log likelihood:

�
T −1
(rt +1 − a0 − a1 rt )2
L(θ) =
�
− ln 2π(b0 + b1 rt ) −
2(b0 + b1 rt )
t =1

(a0 , a1 ) can no longer be estimated separately from (b0 , b1 ).

Optimality conditions for (a0 , a1 ):

�
T −1
(rt +1 − a0 − a1 rt )
(1, rt ) � =0
b0 + b1 rt
t =1

This is no longer OLS, but GLS. More precise estimates of (a0 , a1 ).

Down-weight residuals with high variance.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 14 / 22
Dynamic Portfolio Choice Financial Econometrics

GMM Standard Errors

IID Observations

Under mild regularity conditions, GMM estimates are consistent:

� → θ0 (in
asymptotically, as the sample size T approaches infinity, θ
probability).
Define �
∂E� (f (xt , θ)) ��
�=
d � , � [f (xt , θ
�=E
S �)� ]
�)f (xt , θ
∂θ � ��
θ
GMM estimates are asymptotically normal:
√
� � �−1 �
� − θ0 ) ⇒ N 0, d� � S
T (θ � −1 d�

Standard errors are based on the asymptotic var-cov matrix of the estimates,
� �−1
�] = d� � S
T Var[θ � −1 d�

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 15 / 22
Dynamic Portfolio Choice Financial Econometrics

Problem

Suppose we observe a sequence of IID random variables Xt > 0,

t = 1, ..., T , with probability density

pdf (X ) = λe−λX , X >0

1 Write down the log-likelihood function L(λ).

2 Compute the maximum likelihood estimate b λ.
3 λ.
Derive the standard error for b

c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 16 / 22

Dynamic Portfolio Choice Financial Econometrics

Problem

Suppose you observe a series of observations Xt , t = 1, ..., T . You need to fit

a model
Xt +1 = f (Xt , Xt −1 ; θ) + εt +1
where E[εt +1 |Xt , Xt −1 , ..., X1 ] = 0. Innovations εt +1 have zero mean
conditionally on Xt , Xt −1 ,...,X1 . You also know that innovations εt +1 have
constant conditional variance:

E[ε2t +1 |Xt , Xt −1 , ..., X1 ] = σ2

The parameter σ is not known. θ is the scalar parameter affecting the shape
of the function f (Xt , Xt −1 ; θ).
1 Describe how to estimate the parameter θ using the quasi maximum likelihood
approach. Derive the relevant equations.
2 Derive the standard error for θ
b using GMM standard error formulas.

c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 17 / 22

Dynamic Portfolio Choice Financial Econometrics

GMM Standard Errors

Dependent observations

The relation
� � �−1 �
�] = 1 −1 � � � 1 � � � −1 � � �−1
Var[θ d S d
� = dS d
T T

is still valid. But need to modify the estimate S

�.
In an infinite sample,
∞
�
S= E [f (xt , θ0 )f (xt −j , θ0 ) � ]
j =−∞

Newey-West procedure for computing standard errors prescribes

�k
k − |j | 1 �
T
�=
S f (xt , θ �) �
�)f (xt −j , θ (Drop out-of-range terms)
k T
j =−k t =1

Of special importance: OLS with Newey-West errors.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 18 / 22
Dynamic Portfolio Choice Financial Econometrics

Additional Results

Delta method: distribution of h(θ

�) is approximately
� �� � �
∂h(θ
�) ∂h(θ
�)
N (h(θ), V (h)) , V (h ) = V (θ
�)
∂θ
� ∂θ
�

Hypothesis testing: construct a χ2 test of the hypothesis h(θ) = 0

Derive the var-cov of h(θ), V (h).
Construct the test statistic

ξ = h(θ �) ∼ χ2 (dim h(θ

�) � V (h)−1 h(θ �))

Model selection: pick an order of the AR(p) model using an AIC or BIC
criterion.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 19 / 22
Dynamic Portfolio Choice Financial Econometrics

Bootstrap: General Principle

Bootstrap is a re-sampling method which can be used to evaluate properties

of statistical estimators.
Bootstrap is effectively a Monte Carlo study which uses the empirical
distribution as if it were the true distribution.
Key applications of bootstrap methodology:
Evaluate distributional properties of complicated estimators, perform bias
adjustment;
Improve the precision of asymptotic approximations in small samples
(confidence intervals, test rejection regions, etc.)
Bootstrap bias correction (e.g., predictive regressions):
� � � �
E θ �� − θ
� − θ0 ≈ ER θ �

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 20 / 22
Dynamic Portfolio Choice Financial Econometrics

Boostrap Confidence Intervals

Basic bootstrap confidence interval. Nonparametric approach in IID samples.

For non-IID samples, use parametric bootstrap.

�
c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 21 / 22
Dynamic Portfolio Choice Financial Econometrics

Problem

Suppose you observe a series of observations Xt , t = 1, ..., T . You need to fit

a model
Xt +1 = f (Xt , Xt −1 ; θ) + εt +1
where E[εt +1 |Xt , Xt −1 , ..., X1 ] = 0. Innovations εt +1 have zero mean
conditionally on Xt , Xt −1 ,...,X1 . You also know that innovations εt +1 have
constant conditional variance:

E[ε2t +1 |Xt , Xt −1 , ..., X1 ] = σ2

The parameter σ is not known. θ is the scalar parameter affecting the shape
of the function f (Xt , Xt −1 ; θ).
1 Describe in detail how to use parametric bootstrap to estimate a 95% confidence
interval for θ.
2 Describe how to estimate the bias in your estimate of θ using parametric
bootstrap.

c Leonid Kogan ( MIT, Sloan ) Review: Part II 15.450, Fall 2010 22 / 22

 MIT OpenCourseWare
http://ocw.mit.edu

15.450 Analytics of Finance

Fall 2010

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