550.447 quantitative portfolio theory & performance ...daudley/447/jhuonly/550.447 sp13...
TRANSCRIPT
1
1.1
550.447Quantitative Portfolio Theory
& Performance Analysis
Week of March 4 & 11 (snow), 2013
Fast Algorithms, the Efficient Frontier & the Single-Index Model
1.2
Where we are
Chapters 1 - 3 of AL: Performance, Risk and MPT
Chapters 4, 5, and 11 of E&G The Opportunity Set, Efficient PFs and PF Selection
Chapters 5+ & 9 in Hull Closer Look & review of mean-variance calculations
for Risk, VaR, and Extreme Value Theory (EVT)
1.3
Assignment
For Week of March 4th
Read: A&L, Chapter 3 (Basic Elements of MPT) Read: E&G, Chapters 9 & 7 (Efficient Frontier and
Correlation Structure for Single-Index Model ) See Supplemental on Website
Problems: H5: 2, 5, 9, 11; 19 (Due Mar 4th) Problems: H6: 6 (Due Mar 4th) Problems: H9: 3, 6 (Due Mar 4th) Problems: EG7: 1, 2, 4 (Due Mar 13th) Problems: EG9: 1, 2 (Due Mar 13th)
1.4
Assignment
For Weeks of March 4th / March 11th
Read: A&L, Chapter 4 (Capital Asset Pricing Model and its Application to Performance Measurement)
Read: E&G, Chapters 9 & 7 (Efficient Frontier and Correlation Structure for Single-Index Model ) See Supplemental on Website
Problems: EG7: 1, 2, 4 (Due Mar 13th) Problems: EG9: 1, 2 (Due Mar 13th)
2
1.5
Assignment
Spring Break: March 18 – 23 Mid Term: NEW! April 3rd
Last Day of Class: Wednesday, May 1st
Final: Monday, May 13th ; 9:00am – Noon In the classroom: Whitehead 203
1.17
Fast Algorithms: Efficient Frontier
Markowitz and SharpeWe’ve looked at the formulations for the efficient
frontier, the complexity of the covariance structure, and the algorithmic solution
Sharpe provides an empirical simplification 1-factor model allows reduction of unknown
parameters in an appealing fashion Let’s view the algorithmic consequence of
Sharpe’s approach
1.18
Fast Algorithms: Efficient Frontier
Single-Index Model Suppose returns, Ri can be modeled as
Where Rm, is the return on the market index, and Where Rm , ei are r.v. with std dev σm , σei respectively ei is zero mean, and
ei is uncorrelated w/Rm : Estimates of from time series, regression
analysis – which provides an uncorrelated result (last) Finally, we assume, ei , ej are uncorrelated:
i i i m iR R e
cov( , ) 0i m i m me R E e R R 2, , and i i ei
0i jE e e 1.19
Fast Algorithms: Efficient Frontier
Single-Index Model Then we haveMean Return Variance of Return Covariance of returns
If the single index model holds, then for a PF The expected return is Variance
i i i mR R 2 2 2 2
i i m ei 2
ij i i m
1 1 1
N N N
P i i i i i i mi i i
R X R X X R
2 2 2
1 1 1
2 2 2 2 2 2
1 1 1 1
N N N
P i i i j iji i j
j i
N N N N
i i m i i j m i eii i j i
j i
X X X
X X X X
3
1.20
Fast Algorithms: Efficient Frontier
Single-Index Model So for the PF we can write
Giving a definition for PF alpha and beta
As for risk
Or
1 1
N N
P i i i i m P P mi i
R X X R R
1
N
P i ii
X
1
N
P i ii
X
2 2 2 2 2 2 2
1 1 1 1
2 2 2 2 2 2
1 1 1 1 1 1
N N N N
P i i m i i j m i eii i j i
j i
N N N N N N
i i j m i ei i i i i m i eii j i i i i
X X X X
X X X X X X
2 2 2 2 2
1
N
P P m i eii
X
1.21
Fast Algorithms: Efficient Frontier
Single-Index Model If we assume a large, equally weighted PF then
Which is 1/N times the average residual risk in the PF And as the PF becomes large, the importance of the
average residual risk diminishes to insignificant Thus So the measure of the contribution of a security to the
risk of a large PF is its beta – often used a measure of a securities risk
We will return to this shortly … but first …
2 2 2 2
1
1 1N
P P m eiiN N
1/22 2
1
N
P P m P m m i ii
X
1.22
Fast Algorithms: Efficient Frontier
Single-Index Model Key is the calculation of the Optimal Portfolio First we’ll go through the steps and then
rationalize/justify The steps are:1. Identify the universe of candidate securities2. Select those securities desirable to be included
in the optimal PF3. Calculate the weightings for those selected
1.23
Fast Algorithms: Efficient Frontier
Single-Index Model The main step is the selection process This is done using a measure of desirability – related
to the excess return to beta ratio This is a measure of additional return gained from a candidate
security per unit of non-diversifiable risk The additional return is relative to the risk-free rate
More formally, we use the ratio = the expected return on stock i = the return on the riskless asset = the expected change in the rate of return on stock i with
a 1% change in the market return
i F
i
R R
iFRiR
4
1.24
Fast Algorithms: Efficient Frontier
Single-Index Model The selection processWhen the measure is better than some
threshold, the security is included For convenience, rank the universe by excess return
over beta – from highest to lowest The optimum portfolio consists of all stocks for which
the excess return is greater than some cut-off, C* This far, the procedure is very simple were we to know
the cut-off Lets look at an example …
i F
i
R R
1.25
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process Consider the 10 stocks and RF = 5%
If C* = 5, then only the 1st 5 are included in optimal PF
1.26
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process So how do we find C* The cut-off is determined successively by assuming a
monotonically increasing-sized set including successively “lesser” securities
To do this we designate Ci as a candidate for C* Ci is calculated with i securities assumed in the optimal PF We know that we have found the optimum Ci , called C* , when
all securities used in the calculation of Ci have excess return to beta above Ci and all securities not used in the calculation have excess returns to beta below Ci
1.27
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process Finding C* (result now; verification later) The stocks are ranked by excess return to risk from high to low For a PF of i stocks, Ci is given by
But we can show that this is just where:= exp change in ROR on stock w/ 1% change in opt PF= exp return on opt PF
22
1
22
21
1
ij F j
mj ej
i ij
mj ej
R R
C
iP P Fi
i
R RC
iP
PR
5
1.28
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process Finding C* But we can show that this is just where:
= exp change in ROR on stock w/ 1% change in opt PF= exp return on opt PF
Of course we don’t know these until the opt PF is found, but we can still make use of the expression as …
We add securities to the candidate PF as long as Or from the above, as long as
iP P Fi
i
R RC
iP
PR
i Fi
i
R R C
iP P Fi Fi F iP P F
i i
R RR R R R R R
1.29
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process Finding C* The rhs is just the expected excess return on a particular stock,
based solely on the exp performance of the opt PF The lhs is the analysts estimated excess return on the stock Thus, if analysis of a particular stock leads the PF manager to
believe it will perform better than would be expected, based on the opt PF, it should be in the PF
Now we look at how =>can be used in our framework to find C*
i F iP P FR R R R
22
1
22
21
1
ij F j
mj ej
i ij
mj ej
R R
C
1.30
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process Finding C* – return to the Table (assume = 10%)
Which follows from …
22
1
22
21
1
ij F j
mj ej
i ij
mj ej
R R
C
2
m
1.31
Fast Algorithms: Efficient Frontier
Single-Index Model The 10 stocks and RF = 5%
6
1.32
Fast Algorithms: Efficient Frontier
Single-Index Model The selection process Finding C* The key point is Columns 5 & 6
are cumulative sums of columns 3 & 4, resp Column 7 is just
As soon as Col 7 reaches a maximum, or Col 7 is greater than Col 2 – because of our ordering scheme – we have found C*
2
2
column 51 column 6
mi
m
C
1.33
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings First we recall the case where we allow short sales w/
riskless lending and borrowing (similar in other cases) The criteria is to maximize the objective function
Subject to Thus with We can write
P F
P
R R
11
N
ii
X
1 11
N N
F F i F i Fi i
R R X R X R
1
2 2
1 1 1
N
i i Fi
N N N
i i i j iji i j
j i
X R R
X X X
1.34
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
And finding the maximum involves solving the N equations
Which can be shown (Chapter 6) as equivalent to solving
for i = 1, … , N Where the optimum proportion to invest in each stock is
1
kNk
ii
ZXZ
1 1 2 2i F i i N iNR R Z Z Z
0 where 1, ,i
d i NdX
1.35
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
Lets look at: But for our single index Sharpe’s model we have
Or combining terms
And solving
Where we define
21 1 2 2
1
N
i F i i N iN i i j ijjj i
R R Z Z Z Z Z
2 2 2 2 2
1 1
N N
i F i i j ij i i m ei j i j mj jj i j i
R R Z Z Z Z
2 2
1
N
i F i ei j i j mj
R R Z Z
2
2 2 21
*N
i F i m i i Fi j j
jei ei ei i
R R R RZ Z C
2
1*
N
m j jj
C Z
7
1.36
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
As an aside, to get the earlier “daunting” expression for C* we must express
andIn terms that do not invoke
To do so, we multiply the left hand equation by beta and sum
We can solve for
2
2 21
Ni F i m
i j jjei ei
R RZ Z
2
1*
N
m j jj
C Z
1
N
j jj
Z
22
2 21 1 1 1
N N N Ni F i i
i i m j ji i i jei ei
R RZ Z
22
2 21 1 1
1N N N
i F i ji i m
i i jei ej
R RZ
1.37
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
Which, when substituted into the expression for C* gives
as before
22
122
1 22
1
*1
Ni F i
mNi ei
m j j Nj j
mj ej
R R
C Z
1.38
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
On slide 1.27 we said we could show that this optimal cutoff could be expressed as
where:= exp change in ROR on stock w/ 1% change in opt PF= exp return on opt PF
We have defined Also, Zi is proportional to the optimal fraction of the PF the
investor should hold in each stock, Xi
The proportionality constant multiplying Xi is the ratio of the excess return of the optimal PF to the variance of its return
iP P Fi
i
R RC
iP
PR2
1*
N
m j jj
C Z
1.39
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
The proportionality constant is the ratio of the excess return of the optimal PF to the variance of its return
Thus
From the single index model, we recognize the PF beta, so
22
1*
NP F
m j jj P
R RC X
2
2 22 2 2
1 1
*N N
mP F P Fm j j m j j P F P
j jP P P
R R R RC X X R R
2P F
P
R R
8
1.40
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio Weightings
Multiplying by “1” ( = ) so
Where is the regression coefficient of the return on security i to the return on the PF
Which gives the other form for C* which we used in defining the selection process on 1.27
2 2
2 2
2
1*
1 cov( , ) 1
i m i P mP F P P F
i P i P
iP P FP F P F iP
i P i i
C R R R R
R Ri PR R R R
iP
i i
1.41
Fast Algorithms: Efficient Frontier
Single-Index Model The Optimal Portfolio WeightingsWe end the aside to formalize the approach and now
determine the optimal PF weightings The key equations are
and The second equation determines the relative investment in
each security, and The first scales the weights so they sum to 1: fully invested Note the significance that the residual variance, , in each
security plays in determining how much to invest in each
1
kNk
ii
ZXZ
2 *i i F
iei i
R RZ C
2ei
1.42
Fast Algorithms: Efficient Frontier
Single-Index Model Optimal Weightings For the example Our weightings are
Where the come from Col 4 of the first table, and The Zi : The Xi :
2
Col 4* Col 2 *i i Fi
ei i i
R RZ C C
i
1
kk N
ii
ZXZ
1
2
3
4
5
0.091 23.46%0.38750.0956 24.65%0.38750.0775 19.98%0.38750.110 28.36%0.38750.01375 3.54%0.3875
X
X
X
X
X
5.45 not 4.5 => 1.43
Fast Algorithms: Efficient Frontier
Single-Index Model Important to note that the previous result is exactly the
same as would have been found had the problem been solved using the established quadratic programming result Takes a fraction of the time Simple Calculations Easy to “debug”
The characteristics of a stock that make it attractive and relatively desirable are determined in advance or off-line Desirability of any stock is solely a function of its excess return to
beta ratio
9
1.45
Fast Algorithms: Efficient Frontier
Single-Index Model – With Short Sales Roughly the same technique as before Rank order via excess expected return over beta ratio Find C* and bifurcate set of stocks relative to C* Now, N = 10 , so go to last row (see next slide) for C10 = 4.52
Ratio higher than C* are longs, those lower are shortsWeightings use again, pos or neg Then % are found in two ways from either
Depending on definition of short sales (reinvest proceeds into longs (1st) or reinvest into risk-free rate, aka Lintner, (2nd ))
2 *i i Fi
ei i
R RZ C
1 1 or
N N
k k i k k ii i
X Z Z X Z Z
1.46
Fast Algorithms: Efficient Frontier
Single-Index Model – With Short Sales Returning to the example
1.47
Fast Algorithms: Efficient Frontier
Single-Index Model – With Short Sales Returning to the example
10
11.379i
iZ
1.48
Fast Algorithms: Efficient Frontier
Single-Index Model – With Short Sales Returning to the example
10
1.49
Fast Algorithms: Efficient Frontier
Single-Index Model Constant Correlation assumption and an
approach to selecting the optimum PF Utilizing the result that the efficient frontier can be
determined by solving the N simultaneous equations
But now we assume So Solving for Zi
2
1
N
i F i i j ijjj i
R R Z Z
2 2
1 1
N N
i F i i j i j i i i i j j i jj jj i
R R Z Z Z Z Z
ij i j
2
11
N
i i i F i j jj
Z R R Z
1.50
Fast Algorithms: Efficient Frontier
Single-Index Model Constant Correlation Solving for Zi
As before, to express C* in known terms, multiply by each and the and add up the N equations
Developing and solving for C* yields
Or
1
1 * where *1
Ni F
i j jji i
R RZ C C Z
i
1 1
**1 1
N Ni F
j jj j i
R R N CC Z
1 1
1* 1 *1 1 1 1
N Ni F i F
j ji i
R R R RNC CN
1*
1
Ni F
j i
R RCN
1.51
Fast Algorithms: Efficient Frontier
Single-Index Model Constant Correlation Approach Securities can be ranked according to their excess
return to standard deviation quotient
The cutoff, as used before, is found using C*
Lets look at an example …
i F
i
R R
11
ij F
ij j
R RC
i
1.52
Fast Algorithms: Efficient Frontier
Single-Index Model Constant Correlation Approach
11
1.53
Fast Algorithms: Efficient Frontier
Single-Index Model(let ρ = 0.5)
1.54
Fast Algorithms: Efficient Frontier
Single-Index Model Constant Correlation Approach To determine weightings in each of the 3 qualifiers The optimum to invest in each chosen stock is Xi where
And
1 *
1i F
ii i
R RZ C
1
kNk
ii
ZXZ
1.55
Fast Algorithms: Efficient Frontier
Single-Index Model Constant Correlation Approach The Example where C* = 21/4
1.56
Fast Algorithms: Efficient Frontier
Single-Index Model A final example
A final example
12
1.57
Fast Algorithms: Efficient Frontier
Single-Index Model A final example Using the earlier rule; ranking is 1, 3, 2, 5, 4
and C* = 11.82 so 1 & 3 are in
1.58
Fast Algorithms: Efficient Frontier
Single-Index Model The last thing for the single index model is
parameterization for each security, i The model The alpha and beta are determined by regression on
the returns series Beta
Alpha
i i i m iR R e
1
2 2
1
N
it it mt mtim t
i Nm
mt mtt
R R R R
R R
i it i mtR R
1.59
Fast Algorithms: Efficient Frontier
Single-Index Model The last thing for the single index model is
parameterization for each security, i The model Alpha and beta from regression on the returns series Model Statistics Estimate Error Coefficient of Determination – the square of the correlation
a measure of how much of the variation in a single stock is due to a variation in the market
Standard Error in beta for security i :
i i i m iR R e
2
2
1
1 N
ei it i i mtt
R RN
2
im i m mim i
i m i m i
i ei m 1.60
Fast Algorithms: Efficient Frontier
Single-Index Model How explanatory is this parameterization? How much association in beta one period to the next? Beta’s on large PF contain a great deal of information about
future betas – not so good for individual securities
Historical betas better predictors on PF than on stocks
13
1.61
Fast Algorithms: Efficient Frontier
Single-Index Model – How explanatory? Can the predictive ability be improved? Blume’s Results – Betas tend to one
Adjust prediction in next period Adds undesirable modification of average for a population of
stocks – this bias on individual betas should be removed Forecast improved by subtracting bias after adjusting
& under model 1.02 => 1.033 for `62-`68 For individual beta: subtract 1.033, add 1.02 for `62-`68
( 4̀8 5̀4) 1 ( 5̀5 6̀1) 1.02
2 10.343 0.677i i
1.62
Fast Algorithms: Efficient Frontier
Single-Index Model – How explanatory? Can the predictive ability be improved? Vasicek’s Results – adjust prediction: take a weighted average
of stock’s historical beta, , and the previous period average beta, , over a sample of stocks ML just takes average (equally weighted) Vasicaek:
= variance in distribution of historical period betas over a set of stocks
= variance in distribution of beta for security imeasured in the historical period,
This is a Bayesian Technique
221 1
2 1 12 2 2 21 11 1
ii i
i i
2
1
21i
1 1 1i ei m
1i1
1.63
Fast Algorithms: Efficient Frontier
Single-Index Model – How explanatory? Can the predictive ability be improved? Both Blume & Vasicek lead to better betas than not adjusting Is beta the right criteria? In PF analysis the key inputs are expected returns, variances
and correlations, so: Better to consider ability to give better correlations! This is
what matters to PF optimization Study of Correlation matrix forecast Use the historical correlation matrix itself Matrix from historical beta forecast Matrix from Blume-adjusted betas using 2 prior periods Matrix from Vasicek-adjusted Bayesian technique
2ij i j m
iji j i j
1.64
Fast Algorithms: Efficient Frontier
Single-Index Model – How explanatory? Can the predictive ability be improved? Historical correlation matrix was worst by far All single index models did better – surprising when it might be
though that these loose information in achieving simplification! Bayesian Technique is best if results are forced to have
stationary average correlation to the period where the model was fitted
More can be said about other factors that influence beta and how fundamentals can provide improvements to the forecast - later