Académique Documents
Professionnel Documents
Culture Documents
|
\ .
equals zero, i.e., where
* 0
0
SAA L
L
A
| |
| |
=
|
\ .
. This itself is an important result,
as it suggests that an underfunded plan must invest aggressively simply in order to keep
up with the liability, if controlling surplus volatility is part of the objective function. At
this point, the portfolio would be solely a hedging portfolio, with no companion risky
asset portfolio.
3. Surplus Utility
Now we can set up the total return/total risk form of surplus utility, again in these
same four parts:
Page 14; 3/3/2013
0
0
1
F S
A
R
L
Max U
| |
=
|
\ .
(18)
2
2 0 0
0 0
SAA L Q SAA SAA L Q
A A
L L
| | | | o
| | | |
+
| |
\ . \ .
Part 1
2 0 0 0 0
cov
0 0 0 0
2
SAA L Q PA Q PA SAA L Q
A A A A
f
L L L L
| | | | | | o
| | | |
+ A +
| |
\ . \ .
Part 2
2
2 2 0 0
2
0 0
PA Q PA PA Q
A A
f
L L
| | o + A Part 3
2
2 2 0 0 0
,
0 0 0
2
A L A A L L
A A A
L L L
e
o o e e e +
| |
| | | |
| +
| |
|
\ . \ .
\ .
Part 4
This is a valuable and general form of utility function (even though we have yet to
deal with residuals sourced in constraints on the underlying optimization problem). Lets
proceed to implement these utility functions in optimizable, vector-matrix form.
C. VECTOR-MATRI X VERSI ONS: MULTI PLE ASSET CLASSES
The development so far has been expressed in a single beta form. To be truly
useful in actual practice we need a multi-factor model, spanning the useful asset classes,
sub-asset classes, and styles that we are likely to be using.
The market-related component of any asset allocation policy, any manager, any
asset class, and any liability can be described in the form of some list, or more formally
some vector of factor weights. All of these vectors are in reality vectors of factor betas
across asset classes and styles, and can be thought of not only in that regression sense but
as a sort of vector of mini-CAPM betas in a very real, market-related risk sense of the
term, as each style or factor weight is in fact a representative of some fully-diversified
component of market related risk.
7
Moreover, these multiple factors can be readily
7
There are other factors that might be usable for asset allocation work besides asset classes and
styles, and that might be used one day for next generation asset allocation and strategy decisions. For
example, one can imagine a time in the future when we might use BARRAs factors, or some newer factor
set appropriate to the task of global strategic asset allocation and much more granular than todays asset
class factors.
Page 15; 3/3/2013
combined into a single factor representing the actual CAPM beta, ideally a world market
portfolio beta.
The notation we will develop will have as many consistencies as is possible with
the scalar algebra notation just used. In general, we have followed the standard
convention that scalars are italic, and not bolded, and now well follow the additional
convention that vectors are in lower case bold non-italic letters and that matrices are in
upper case non-italic, bold letters. Beta vectors will be row vectors, all others will be
column vectors. Transposes will be indicated with a bold superscripted T. There will be
various subscripts, hopefully intuitively connected to the scalar algebra subscripts, to
differentiate types of variables.
The beta vectors that we use for the assets,
q
(the weights or betas of the asset
classes in Portfolio Q),
A
(the current policy asset allocation weights, inclusive of both
SAA and PA effects),
SAA
(the strategic component of asset allocation weights),
PA
(the
portfolio active portion of the asset allocation weights), and for the liability,
L
, are
(1 x q) row vectors of asset class factor betas across the q Portfolio Q asset class and style
factors.
1. Portfolio Active Beta in Multi-Asset Class Form
A bit of refreshing on the concept of portfolio active beta may be in order in
preparation for developing the concept in vector-matrix form. It is driven from a zero
position, if at all, by a special type of active return forecast. In the first section we
referred to single factor beta timing inputs, or exceptional market forecasts with respect
to the equilibrium return of Portfolio Q itself, but now were going to address exceptional
market forecasts on each component beta, the element of Portfolio Q. After parsing the
scalar version of the return function equation (2), above, we separated out from the total
return forecast
Q
f , the consensus expected return
Q
, and the exceptional market
forecast,
Q
f A , defining the relationship between them as
Q Q Q
f f = + A . In vector form to
capture the components of Q, we can write this return forecast
q
f for the (q x 1) asset
class and style return column vector as the sum of the consensus expected return vector
q
r and a vector of exceptional market forecasts
q
f :
Page 16; 3/3/2013
q q q
f = r +f . (19)
Well need additional notation,
q
V being the (q x q) variance-covariance matrix
for the asset classes or factors, and
n
V being the (n x n) variance-covariance matrix of
the n managers alphas.
8
We summarize the managers residual returns as alphas in a
single (n x 1) vector a across those n managers.
D. VECTOR-MATRI X FORM: ASSET-ONLY
1. Return model
With this notation, we can write directly in vector-matrix form, and thus make
usable, the four-part parsed return of equation (4). Well take advantage of the fact that
we can add the managers beta vectors together, getting a weighted average combined
beta vector, by multiplying the managers holding percentage vector (also the
optimization change variable)
n
h by the (n x q) matrix of the managers individual beta
vectors
n,q
B (the (n x q) matrix of the n managers q beta exposures; i.e., their normal
portfolios) , getting =
T
A n n,q
h B . (In this case, the forward-looking estimates of the beta
vectors of the n individual managers are taken as givens, estimated before the
optimization process.) Thus the holdings of the managers can act as the output vector for
the optimization by controlling manager weights one indirectly also controls the asset
allocation policy, within the confines of the betas available to the portfolio through the
candidate managers.
Beyond that point, Ill skip the algebraic progression of the parsing process, as
this represents a very straightforward mimicry of the parsing process used to obtain
equation (4), written in a form that assembles the returns from the individual manager
building blocks of this portfolio, an improved level of construction detail relative to
more typical asset class building block approaches. Here is the complete version of the
8
The latter is often thought of as a diagonal matrix, but need not be and strictly speaking, could
not be. Think of the extreme case: if there were only two managers in the world, and between them they
held all the securities, their alphas would have to be perfectly negatively correlated. In the real world,
diversification tends to drive managers from this bias to negative correlations towards a zero average, and
in practice zero is a good null hypothesis for manager correlations where there is no special contradictory
information.
Page 17; 3/3/2013
vector-matrix return model the labels below the vector-matrix terms translate each into
its scalar algebra equivalents:
Part 1
Part 2
Q Q Q
SAA A SAA SAA
PA
A F
f
R R
| | | |
|
A
| |
|
|
|
= + + +
|
|
|
|
\ .
T T T T T
SAA q q n n,q q q SAA q q SAA q q
q q q q q q T T T T
q q q q q q q q q q q q
V h B V V V
r - r f
V V V V
Part 4
Part 3
Q
A SAA
PA
f
o
| |
|
A
| |
|
|
|
+ +
|
|
|
|
\ .
T T T
n n,q q q SAA q q T
q q n n T T
q q q q q q
h B V V
- f h a
V V
(20)
a. Beta-related residuals
Well, I used the word complete. That was a bit of an overstatement. Equation
(20) is almost complete, but not fully.
Because we are going to use this in an actual optimization in a later step, we want
to make sure that we are properly representing beta-related residual returns, the alphas
from asset class misfit as opposed to the alphas from the active managers. This is a real
world issue that we didnt concern ourselves about when developing the scalar algebra
version earlier; we glossed over it. These alphas are uninformed by insights, simply being
a result of some deviations from market perfection. If we are in a happy world where
are beta expected returns lie neatly on the security market line, these alphas are all zeros,
but they will then be non-zero in practice for many consultants and advisors who ignore
such niceties. And the risk associated with these alphas will be positive in either event.
Heres the deal: Weve shown the pure alpha from the managers, but we havent
shown any residuals from the possibility (indeed the likelihood) that the beta vectors of
our asset mix and of our liability mix might be interior to, or off, the CML. This isnt
an unusual or exceptional issue, but a usual occurrence under common asset allocation
Page 18; 3/3/2013
situations, most obviously related to the practical constraints on beta imposed by the
limited beta factors usually available in the manager mix.
There might be other reasons for this on the asset side of the problem, perhaps for
as obvious a reason as a lack of available borrowing or lending capabilities at the risk-
free rate. This creates the usual curved efficient frontier just tangent to, but in whole or in
part below, the CML; this difference from the CML associated with residuals. Commonly
used constraints all cause a variation on this same problem in the presence of
constraints, the frontier is going to be interior to the CML at most or all points.
The liability beta vector is almost certainly not efficient, and thus it is also interior
to the CML. If the liabilities are exactly hedged, we eliminate it from the possibility of
contributing to a residual, but that isnt always going to be the case. The bottom lines is
that we need to differentiate and account for the difference between these actual factor
beta vectors and their CML counterparts, and that requires us to deal with the somewhat
messy math of the residuals.
A direct way to start thinking about these beta-related residuals (inspired by
Appendix to Chapter 4 of G&K) is to solve for them, for example by manipulation of
equation (2) for asset-related residuals, getting:
( )
A F A Q
A A Q
R R f
r f
o |
o |
=
=
(21)
This simple manipulation confirms my assertion that there is a beta element to the
residuals, they arent solely the manager alphas as we are accustomed to expressing them.
To the extent that the beta vectors are interior to the CML, there will be a beta-related
residual component of returns, and also of risks.
b. Capital market line beta vectors and factor beta vectors
To write this in matrix form, well need to have the tools to differentiate between
the factor beta vector, loosely equivalent to an asset class holdings vector and which may
include residuals, and the pure or CML beta vector, which has no residuals as it is
composed of beta values that place it somewhere on the CML. This latter vector is the
classic two-fund portfolio, part Portfolio Q,
q
, the world market portfolio, and part the
Page 19; 3/3/2013
riskless asset. In this section well set up these tools, including the CAPM single factor
beta which controls the leverage of
q
in this two-fund format.
The CAPM beta is the ratio of a covariance of a portfolio with the market, to the
variance of the market. For example, if the factor beta vector for the SAA policy
9
is
SAA
,
and if for the asset classes as invested through managers it is the sum of those managers
asset class exposures:
=
T
A n n,q
h B (22)
then the scalar asset and SAA betas would be formed by dividing the covariance (with q)
of our factor beta vectors, by the variance of q:
A
| = =
T T T
A q q n n,q q q
T T
q q q q q q
V h B V
V V
(23)
SAA
| =
T
SAA q q
T
q q q
V
V
(24)
Lets use these scalar betas to find the residual terms in vector-matrix form.
Expressing equation (21) in this form:
( )
o = +
T T
n n,q q q T T
n n,q q n n q q T
q q q
h B V
h B f h a f
V
o
| |
= +
|
|
\ .
T T
n n,q q q T T
n n n n,q q q T
q q q
h B V
h a h B f
V
(25)
But we can simplify this with the scalar beta abbreviations so that we have an
easier equation to read, and it makes it easier to note that were just comparing the actual
factor beta vector of the assets to the CML beta vector that represents those same assets:
( )
A
o | = +
T
n n A q q
h a f , (26)
or, if the beta-related residual vector (the portion in the parenthesis) is indicated as a beta
vector with subscript R, then we can abbreviate this even further, to:
o = +
T
n n R q
h a f . (27)
9
If there are any constraints, the SAA policy will not be on the CML but will be on some interior
efficient frontier.
Page 20; 3/3/2013
The first term of equation (26), summing the managers alphas, is already familiar
weve been using it to represent the managers alphas already (Waring, et al [2000]). It
isnt a new component of the residual for our optimization.
The second term isnt really new, either. It is the misfit return of the same paper,
weve simply generalized it as a total portfolio measure. Perhaps it better renamed from
misfit return to beta-related residual return.
c. Total return equation, with beta-related residuals
The addition of this residual component allows us to state a generalized, usable
total portfolio return term, to truly complete equation (20). Expressed fully including
the new material in Part 4, it is:
A f
R R = (28)
+
T
SAA q q
q q T
q q q
V
r
V
Part 1
Q Q
SAA PA
f
| |
A
| |
+ +
|
|
\ .
T T T T
n n,q q q SAA q q SAA q q
q q q q T T T
q q q q q q q q q
h B V V V
- r f
V V V
Part 2
PA
|
| |
+
|
|
\ .
T T T
n n,q q q SAA q q
q q T T
q q q q q q
h B V V
- f
V V
Part 3
| |
+ +
|
|
\ .
R
T T
n n,q q q T T
n n n n,q q q T
q q q
h B V
h a h B f
V
Part 4
Or equivalently, we can state this quite compactly, using more abbreviated
notation:
( ) ( )
Part 1 Part 2 Part 3
Part 4
A f SAA A SAA SAA A SAA
R R | | | | | | = + + + +
+ +
q q q q q q q q
T
n n R q
r - r f - f
h a f
(29)
Page 21; 3/3/2013
2. Risk model:
The scalar algebra version of the risk model in equation (6), above, can also be
expressed in vector-matrix form, again across the same n managers and q asset classes or
styles (or other market factors).
Well start with equation (5) and follow a similar progression to that we used for
parsing the scalar algebra version. To keep it simple, well use the scalar values of beta
developed in equation (23), et seq:
2 2 2 2
A A Q
o | o e = + (5)
2
A
| =
T T
q q q n n n
V +h V h
Separating the beta of the assets into its strategic and portfolio active components:
( ) ( )
2
2
A SAA A SAA
o | | | = + +
T T
q q q n n n
V h V h
( ) ( )
2
2 2
Part 4
Part 1 Part 2 Part 3
2
A SAA SAA A SAA A SAA
o | | | | | | = + + +
T T T T
q q q q q q q q q n n n
V V V h V h (30)
Well, again, almost. In a world with no beta misfit residuals, we would be done.
But alas, it still isnt such a world. We need the beta misfit residuals to parallel the
additional term we added to the return term, in equation (28). Well do this by solving
equation (5) for residual risk, and rewriting it in vector-matrix form:
2 2 2 2
A A Q
o | o e = + (5)
2 2 2 2
A A Q
e o | o =
If we write
2
A
o directly, it will be in terms of the vector of asset factor betas, not
in terms of the scalar betas:
( )
2 2
2
2
A Q
A
A
| o
o
| = +
T T T
A q A n n n q q q
V h V h V
( )
2
A
| = +
T T T
n n n A q A q q q
h V h V V (31)
Were already familiar with the term on the left hand side of equation (31), the
residual risk introduced by the search for alpha engaged in by the managers. The new
term, in parentheses on the right, is the beta-related residual risk term. With this more
Page 22; 3/3/2013
complete residual risk term in hand, we can restate total portfolio risk, equation (30), in a
more complete and general form that will fit the real world:
2 2
A SAA
o | =
T
q q q
V Part 1 (32)
( ) 2
SAA A SAA
| | | +
T
q q q
V Part 2
( )
2
A SAA
| | +
T
q q q
V Part 3
( )
2
A
| + +
T T T
n n n A q A q q q
h V h V V Part 4
These return and risk equations admittedly look a bit hairy, particularly when
remembering that we have used abbreviated scalar forms for many beta terms, but this
parsing is valuable. And look at the bright side: this is what computers are for! Once
programmed, a long equation is as easy as a short equation but you better understand it
at least once!
3. Utility
Collecting return and risk terms developed above, we get the following fully
generalized version of the utility function (a reminder: were still in asset-only space).
This uses the abbreviated versions of the beta terms, but of course would have to
implemented with the full versions:
Max
p f
U R =
n
h
(33)
2
SAA SAA SAA
| | +
T
q q q q q
r V (Part 1)
( ) ( )
cov
2
A SAA SAA A SAA SAA
| | | | | | ( + +
T
q q q q q q q
- r f - V (Part 2)
( ) ( )
2
A SAA PA A SAA
| | | | +
T
q q q q q
- f V (Part 3)
( ) ( )
2
A A e
| |
(
+ + +
T T T T
A q q n n A q A q q q n n n
f h a V V h V h (Part 4)
The optimization change variable is simply
n
h , the manager weights, the basic
unit of trade for a sponsor portfolio, and the key to summing up both the asset beta and
the alpha terms. Because this term is hidden through our abbreviations, the operation of
this utility function is a bit obscured. But this is a key to using this as a total return
optimization function. Through the manager weights, which tie in the managers betas,
the optimizer allows for both active beta and alpha to be optimized relative to the SAA
Page 23; 3/3/2013
policy, and simultaneously for the SAA policy to be established. This is important, and
handy. What this means to the investor is that the output of the optimizer tells her what
managers to hold and in what weights, and this is controlling all the important underlying
variables, including strategic asset allocation policy in Part 1, tactical asset allocation
policy in Part 3, and manager optimization policy in Part 4. Thats a lot!
E. VECTOR-MATRI X FORM: SURPLUS UTI LI TY
1. The surplus residual terms
This time, well develop the residual terms before building the return and risk
models. Well solve for the total surplus residual return (the unsystematic residuals plus
the beta-related residuals) by manipulating equation (12), repeated here solely for
convenience:
0 0 0
0 0 0
1
S F A L Q A L
A A A
R R f
L L L
| | o o
| | | | | |
= + +
| | |
\ . \ . \ .
(12)
Rearranging to solve for the residual return with respect to beta, and with respect
to both the assets and the liability:
0 0 0
0 0 0
1
A L S F A L Q
A A A
R R f
L L L
o o | |
( | | | | | |
=
( | | |
\ . \ . \ .
Because the term
0
0
1
S F
A
R R
L
| |
|
\ .
is simply the excess return of the surplus,
S
r ,
this simplifies further. We subtract the portfolio surplus beta, and rearrange, to get a
complete expression of the entire residual:
0 0
0 0
total residual
A L S A L Q
A A
r f
L L
o o | |
| | | |
=
| |
\ . \ .
0 0 0
0 0 0
unsystematic residuals
S
L A L
r
A A A
L L L
o | |
(
(
| | | | | |
( = +
| | |
(
\ . \ . \ .
(
T
A L q n n q q
f h a f
Page 24; 3/3/2013
0 0 0
0 0 0
L A L
A A A
L L L
o | |
( | | | | | |
= +
( | | |
\ . \ . \ .
T
n n A L q q
h a f (34)
Notice that we have used the same type of compact notation that we developed
above to describe the CML beta of the liability,
L
| =
T
L q q
T
q q q
V
V
.
This can be further abbreviated, and although well lose the details we may see
the intuition more clearly:
0 0
0 0
A L L S
A A
L L
o o o |
| | | |
( = +
| |
\ . \ .
T
n n S q q
h a f (35)
This reads that the total surplus residual return, on the left hand side, equals the
sum of the residuals from manager alphas uncorrelated with the asset class factors, and
the beta-related residuals from the combined liability model and asset class vectors where
they are off of the CML.
While were dealing with the residuals, lets figure out a full definition of surplus
residual risk, expecting to find a risk term parallel to the surplus residual return, equation
(34):
2
2 2 2 0
0
S A L Q S
A
L
o | | o e =
| |
+
|
\ .
(15)
2
2 2 2 0
0
0
0
0
0
2
2
2 0 0
0 0
2
2 0 0
0 0
beta-related surplus residual r
S S A L Q
A L
A L
S
S
S
A
L
A
L
A
L
A A
L L
A A
L L
o
e o | | o
| |
| |
e
e
=
| |
|
\ .
| | | |
| |
= +
| |
|
\ .
\ . \ .
| | | |
| |
= +
| |
|
\ .
\ . \ .
T
T
A L q A L q q q
T
T
A L q A L q q q
V V
V V
isk
Page 25; 3/3/2013
0
0
2
2 2 0 0
,
0 0
Unsystematic surplus residual risk
2
0 0
0 0
beta-related surplus residual risk
2
A L
S A L L
A
L
A A
L L
A A
L L
| |
e e e
(
(
| |
( = +
|
(
\ .
(
| | | |
| |
+
| |
|
\ .
\ . \ .
T
n n n
T
T
A L q A L q q q
h V h
V V
(
(
(
(
(
(36)
We lose the detail, but this can be abbreviated and expressed as:
2
2 2 2 0 0
,
0 0
beta-related surplus residual risk
Unsystematic surplus residual risk
2
S A L L S
A A
L L
e e e |
| |
= + +
|
\ .
T T T
n n n S q S q q q
h V h V V (37)
Again, there is no point in using vector-matrix algebra to represent the liabilitys
alpha-related residual risks nor its covariance with the asset residual risks, so we will
continue to show these terms in scalar form.
While looking at the unsystematic portion of residual risk, there are some
observations that we can make: The third of these terms,
2
L
e is a constant, and could be
dropped. And one would expect the middle term, the correlation of the asset and liability
unsystematic residuals, to tend towards zero, so this risk term shouldnt be especially
important to the optimization process However, the total risk may more accurately be
represented if an estimate of these risks are made and included.
10
2. Surplus return
With this result, we can write our surplus return equation (14) in matrix form,
including the more complete surplus residual of equation (35):
0
0
1
S f
A
R R
L
| |
=
|
\ .
(38)
0
0
SAA L
A
L
| |
| |
+
|
\ .
q q
r Part 1
10
These liability residual terms can become very real and important, where as in some cash
balance plans the liability itself is credited with a rate that includes an explicit manager alpha with its
attendant residual risk, or where it includes company stock with all the residual risk of a single stock
portfolio.
Page 26; 3/3/2013
( )
0 0
0 0
SAA L A SAA
A A
L L
| | | |
| |
+ +
|
\ .
q q q q
f r Part 2
( )
0
0
A SAA
A
L
| | +
q q
f Part 3
0
0
L S
A
L
o |
| |
( + +
|
\ .
T
n n S q q
h a f Part 4
Note that the alpha, or residual return, for the liability,
L
o , i.e., the portion of the
liability return not related to market risk factors, is simply a scalar alpha, not one in
vector-matrix form. With todays data, its probably impossible to accurately estimate
this number but thats ok, as it is a constant and the largest error wont affect the
optimization result. While we wont be precise, we can probably back into a number that
is reasonable. One probably improves the clarity of the risk-return view by making a dull
axe estimate and including it simply so that it shows up in the output, but I would
understand that some analysts might exclude it and simply note that the resulting return
number is understated for lack of its inclusion, its true magnitude being unknown. The
same, of course, will be true on the risk side.
I should be clear this liability alpha is trivial to the optimization, but it may not
be trivial to the sponsor that actually experiences this very real element of return and risk.
There are risks in DB plans that cannot be hedged away in the markets under any
circumstances. These can be moderated by being less ambitious in increasing benefits,
by pricing benefits correctly, by using mortality tables with more improvement built in,
etc. But not all risk can be eliminated through the asset portfolio. This is the alpha of the
liability, or the residual risk of the liability. It is risk not related to the markets.
3. Surplus risk
2
2 0
0
S SAA L
A
L
o | |
| |
=
|
\ .
T
q q q
V Part 1 (39)
( )
0 0
0 0
2
A SAA SAA L
A A
L L
| | | |
| |
+
|
\ .
T
q q q
V Part 2
Page 27; 3/3/2013
( )
2
2
0
0
A SAA
A
L
| |
| |
+
|
\ .
T
q q q
V Part 3
2
2 2 0 0
,
0 0
2
A L L S
A A
L L
e e |
(
| |
( ( + + +
|
(
\ .
T T T
n n n S q S q q q
h V h V V Part 4
4. Surplus utility
Now we assemble the risk and return terms into a utility function:
0
0
Max 1
S f
A
U R
L
| |
=
|
\ .
(40)
2
0 0
0 0
SAA L SAA SAA L
A A
L L
| | | |
| | | |
+
| |
\ . \ .
T
q q q q q
r V (Part 1)
( )
( )
0 0
0 0
0 0
cov
0 0
2
SAA L A SAA
A SAA SAA L
A A
L L
A A
L L
| | | |
| | | |
( ( | |
+ +
( ( |
\ . (
(
| |
(
|
(
\ .
q q q q
T
q q q
f r
V
(Part 2)
( ) ( )
2
2
0 0
0 0
A SAA PA A SAA
A A
L L
| | | |
| |
+
|
\ .
T
q q q q q
f V (Part 3)
2
2 0 0 0
| ,
0 0 0
2
|
2
L A L L
S S
A A A
L L L
e o
e |
o e e
| |
( (
| | | |
( ( + +
| |
( (
\ . \ .
(
( ( ( +
T T
n n n n n
T T
S q q S q S q q q
h a h V h
f V V
(Part 4)
Ive abbreviated this through my notation a great deal, so that some of the
operations cant be seen. The most important one that cant be seen with this abbreviation
is the optimization change vector, the manager holdings vector
n
h . It is found inside of
every
A
| through the required term found in its numerator, =
T
A n n,q
h B . So apologies
there is much accounting to do in properly handling total returns, and even more so in a
surplus context.
Note that part four is really two parts the informed residuals and the beta-related
residuals.
Page 28; 3/3/2013
This is a very complete total return utility function, capable of informing a
number of the common strategy decisions that every plan sponsor and other serious
investor routinely faces:
- Basic strategic asset allocation across asset classes, on either or a liability-relative
or on an asset-only basis.
- Making the manager selection decision and the manager structure decision.
- Making tactical decisions to re-weight asset classes and styles based on
exceptional market forecasts (views).
- Beta-alpha separation decisions.
F. TACTI CAL BETS, BETA TI MI NG, AND THE LI KE
1. A Bit More on Beta Timing
There are two reasons why an optimization might result in a beta different than
the strategic asset allocation beta,
SAA
| , i.e., why it might have a non-zero
PA
| . Weve
just discussed the beta-related residuals.
The other of these reasons is the active management reason discussed earlier,
the forecasting of non-zero exceptional market forecasts. Recall, this results in an active
beta, the beta that is different than the beta of the SAA policy in single beta space, and
different from the CML SAA policy beta vector in multiple beta space, i.e.,
PA p SAA
| | | = , and the more generally useful
PA p SAA
= - . We discussed beta timing
briefly above, in section A.1.d, but now that we have built the vector-matrix version we
can make some additional observations.
This is beta timing, or market timing, or active beta, which in different
applications is known as tactical asset allocation, style rotation, country rotation, etc. A
sponsor or other investor could hold intentional mis-weights of the beta factors in the
total portfolio benchmark (such as styles, asset classes, or other factors) in an effort to
add alpha to the portfolio through active positioning of the asset classes. The exceptional
market forecast component of expected return,
Q
f A or in vector form,
q
f , is the input
that would drive any such intentional beta mis-weights.
Page 29; 3/3/2013
This of course, is as active as active gets, by any definition. Each time the
optimization is run, the forecast exceptional market return might well be different, so the
resulting active beta position will also be different during that period. If these forecasts
are skillful, value will be added over a time series of such forecasts. That value will be
experienced in realization as a regression alpha over and above the policy beta return.
In todays practice, this particular active management discipline is growing to be
unusual, often disparaged even, though in the hands of the skillful it really is an
appropriate investment discipline. The search for alpha today is more often through hired
money managers expected to add alpha through their carefully selected residual holdings
of securities (and occasionally of styles and factors) relative to the managers underlying
benchmark.
Ive organized our work here with the goal of supporting the sponsors beta
timing activity, processing the sponsors exceptional market forecasts appropriately into
the current asset allocation policy. While many scoff, nonetheless sponsors often want
their portfolio to reflect a view that value will beat growth over the next period, or
bonds will beat stocks, or domestic equities will beat international equities, or whatever.
The exceptional returns vector allows a sponsor to readily incorporate such views.
This has the quite important side benefit of taking some of the forecasting angst
out of the basic SAA decision. Using consensus expected returns for the asset classes that
are the elements of Portfolio Q is pretty easy reverse optimization is quickly and readily
available and can usually be applied sensibly. Modifications to that set of forecasts can be
understood for what they are, exceptional forecasts, and processed in a manner and with a
risk aversion term appropriate to their active nature.
G. ADAPTI NG THE OPTI MI ZER TO DI FFERENT SOLUTI ON
NEEDS: A SI DE NOTE FOR I NTERNAL DESI GN USE
The biggest problem facing anyone implementing an optimizer with these talents
may not be the development of the mathematics, now set forth above. The biggest thing
may be the design of the user interface that allows a user to control such an optimizer in
actual use. Seldom will a strategist use all the talents of this optimizer in a single grand
Page 30; 3/3/2013
optimization, although that might happen. Can we identify the tasks that it might be used
for, in an effort to support the design of the user interface?
There are four major inclusion/exclusion questions that the strategist must answer
before setting up this very generalized optimizer to solve a particular problem:
1. Are we solving for the underlying SAA policy, or has it already been
determined so that we can take it as a given benchmark?
a. And if we are solving for the SAA policy, are we doing so on an
asset-only basis, or on a surplus basis relative to a liability?
2. Will there be exceptional market forecasts, for beta timing (portfolio
active beta)?
3. Will there be manager structure (alpha) decisions, or are we simply
optimizing for SAA and portfolio active beta (beta decisions) across the
asset classes?
Depending on which of these decisions are to be included in the optimization,
there are implications for the inputs that will be needed, the outputs that will be
interesting, and for which ones of the five terms of the surplus utility function will be
used. It is conceivable that one might do all of these tasks together in well-coordinated
total return optimization, or that we take some one or more of them one at a time, as we
have done for years in optimizing manager structure. There are many different
combinations, and this presents many design challenges for the user interface, for the
schedules of input parameters, and for the type of output graphs and data tables that will
be useful. Well only figure this out over time.
1. Question 1: Is the strategic asset allocation policy to be
determined, or not?
If, in answer to the first question, the policy asset allocation, or SAA policy, is to
be determined by the optimization, then
PA A SAA
= , i.e., the SAA beta vector is a
variable and will be determined in the process; it is the optimization change vector (either
directly, or through the manager betas and the manager holdings, see below). If, on the
other hand, the SAA policy has already been set, it is not going to be up for change, then
Page 31; 3/3/2013
it can be expressed as a fixed benchmark beta vector, and
PA A b
= . The vector is
frozen.
2. Question 1A: Is a liability to be considered?
If the SAA policy is to be determined, then it will be determined either with, or
without, the inclusion of a liability (the present value of the future consumption for which
the assets are being held). If the liability is to be considered, then the utility function in
equation (40) is OK, as is. If not, then the values for the vector
L
and the scalars
, A L
e ,
and
L
e are all set to zeroes, and the value for the A/L ratio
0 0
A L is set to 1, except
where it modifies the risk free rate term, where it is also set to zero. With this, equation
(40) is identical to the asset-only utility function, equation (33).
If the answer to the first question was that the SAA was not to be determined,
then the answer to question 2 is also determined, since liability considerations would only
seldom be valuable where the SAA policy were not at issue. The exception that disproves
the rule is for cash balance plans: in the rare occasion where the liability is set by
reference to a third-party managed active fund, there is an alpha that is hedgeable, since
the plan can actually invest in that fund if it chooses to do so. In that case, we want the
liability terms even if not doing SAA policy and just doing manager structure.
3. Question 3: Is beta timing included?
If the optimizer is going to be used to decide whether to over- or under-weight
some of the asset classes in the beta vectors, then it would be done by putting non-zero
entries in one or more of the elements of the exceptional market forecast vector,
Q
f .
This will have the effect of altering the
PA
vector to reflect those biases. This is an easy
one to handle: the vector is by default a zero vector.
If the
Q
f vector is made to have any non-zero elements, then we may want to
display output data for the
PA
vector, and the risk and return impacts of this portfolio
active holding, perhaps separately for Parts 1 and 2 of the utility function.
Page 32; 3/3/2013
4. Question 4: Are managers being allocated, is this in whole or in
part a manager structure optimization process?
If managers are included in the optimization, we often refer to the active
return/active risk part as a manager structure optimization process (MSO). In this case,
the optimization change vector is
n
h , as it is determinative of
A
| through the handy
mechanism
A
| =
T
n n,q q
T
q q q
h B V
V
, which sums the managers raw beta vectors and calculates the
vector of CML betas that the manager holding vector thus implies.
Without managers being included in the optimization, we are simply determining
the asset allocation solution and not the simultaneous manager structure solution.
A
is
the optimization change vector.
So depending on whether the problem involves managers, the optimization
change vector changes its level of aggregation. A simple
A
when they are not included,
and the function =
T
A n n,q
h B translates the manager holdings vector into the total assets
beta vector when they are.
H. MI SCELANEOUS
1. Manager structure optimization:
If managers are included, even if the exceptional market forecast is a zero vector,
then in addition to the Part 4 treatment of their alphas and true active risks, the beta-
related residuals from Part 4 and the (sometimes zero) risk from Part 2 will be needed.
If we take out the liability terms (which might be very valuable to leave in for a
cash balance plan that has alphas in its crediting rates, for example), and if we combine
the beta-related residual risk from Part 4 with the portfolio active risk from part 3
(requiring us to set the lambda equal to each other), then this becomes a very
straightforward manager structure optimization utility function.
I . ALPHA-BETA SEPARATI ON
We are in the habit of presuming that alpha always has to come with beta in a pre-
packaged form. But the reality is that real alpha is definitionally uncorrelated to beta, and
Page 33; 3/3/2013
that observation frees the mind to consider unleashing alpha and beta from each other.
The new class of market-neutral long-short funds are good evidence of this reality
being pure alpha (plus the risk free rate) they can readily be ported onto any readily
available source of beta, from index funds to futures contracts, etc. The long-standing
practice of taking alpha from a fixed income portfolio by shorting Lehman Aggregate
and going long S&P 500 is another example.
But true alpha (regression alpha), being independent of beta, need not be ported
from one asset class to another. In the ideal world, one would build a portfolio of the best
alpha sources one could identify, regardless of their beta exposures or lack thereof. And
when done, one would fix the betas so as to make them look like the intended SAA
policy. We call this portable beta, and we see it as the fully generalized version of
portable alpha, full separation of alpha and beta.
There are costs and other frictions on this ideal process, so in reality it is not two
steps, but must be done as one, with costs on beta transfer and availability of beta
transfers being taken into account.
To do this with this optimizer, we only need to set up a system of constraints that
allows this to happen. We might (or might not it is potentially a strategists decision
variable) start with a budget constraint on the dollars of alpha exposure. And usually,
alphas sources are long-only with respect to the investor, because it is unfortunately
difficult to short a manager (although the alpha source itself may be short, long, or
mixed). These two constraints operate on the holdings of the managers. What about
betas?
The beta constraints would be separately stated from the alpha constraints. It has
been the practice to have both a budget constraint and a long-only constraint on the total
beta holdings of all asset classes and an implied constraint on the combined beta-alpha
exposure of each manager. Of course, to return the CML, cash must be available for use
in leveraging Portfolio Q, the reference portfolio, which implies that one must be able to
short cash. In practice, most strategists dont allow cash to be shorted, as one cant in the
real world borrow at the risk free rate. But this isnt enough. An example will illustrate
why.
Page 34; 3/3/2013
Thus unconstrained with respect to manager sources of alpha, we want the
optimizer to put the weight of the portfolio into the best combination of the best
managers without regard whatsoever to the beta of those alpha sources. One can imagine
the entire portfolio being invested in one small cap growth tech sector specialist if that
is where ones assumptions indicated that the best alpha was to be obtained. Of course,
then we would have to short most of that managers beta in order to go long the other
asset class betas to get back to the optimal beta portfolio.
The point is, we have to allow individual positions in beta that are short; it is only
the net beta positions, across short and long individual positions, that must be compliant
with the budget and long only constraints. So if we allow particular managers to go short,
most easily managers that are really just beta (indexes, futures, etc.), so long as the net of
all asset classes meets the budget and long-only constraint, the optimizer can short beta
where it needs to in order to favor one alpha source over another.
Costs for beta shorting and longing need to be estimated in the form of rates of
return, and incorporated into the short beta opportunity set.
The approach to actually conducting such an optimization is deceptively simple
merely include any available short beta sources (managers, futures, swaps, trusts,
whatever) as additional managers in the optimization, with their cost estimates. If the
costs of shorting and holding these short beta sources is included as a negative alpha
(take care to get the signs right its a short position, so the negative alpha will end up
being positive), then the optimizer will do the best it can within the constraints of costs
and availability to maximize utility across alpha and beta separately.
J . CONCLUSI ON
The idea of conducting a total return, total risk portfolio optimization across all
important variables is an appealing one. To date, those that have said they are doing this
have done so without really parsing the utility functions correctly. While the math is
tedious to work out, once set up it is no big deal; the computer takes care of it.
There is still an open question in my mind about whether we can do a one-step
total return optimization and get sensible results, or whether we need to first determine
the SAA benchmark (with the liability, if desired) and then move on to the other
Page 35; 3/3/2013
decisions. I suspect that we have something yet to learn about the interactions of strategic
and tactical inputs.
We also have a lot to learn about interface design and output design.
Page 36; 3/3/2013
To do list:
Capture the two fund (three fund) theorem page from my Q-Group presentation to
simplify some of the concepts here; citing Waring, M. Barton and Duane Whitney. 2009.
An AssetLiability Version of the Capital Asset Pricing Model with a Multi-Period
Two-Fund Theorem. Journal of Portfolio Management, Vol. 35 No. 4, Summer 2009,
http://www.iijournals.com/doi/pdfplus/10.3905/JPM.2009.35.4.111 .