com/abstract=1314585
The BlackLitterman Model In Detail
Revised: February 16. 2009
Copyright 2009  Jay Walters. CFA
iwalters(blacklitterman.org
!"#$%&'$
(
In this paper we survey the literature on the BlackLitterman model. This paper provides a complete
description oI the model including Iull derivations Irom the underlying principles. The model is
derived both Irom Theil's Mixed Estimation model and Irom Bayes Theory. The various parameters oI
the model are also considered. along with inIormation on their computation or calibration. Further
consideration is given to several oI the key papers. with worked examples illustrating the concepts.
)*$%+,'$.+*
The BlackLitterman model was Iirst published by Fischer Black and Robert Litterman oI Goldman
Sachs in an internal Goldman Sachs Fixed Income document in 1990. Their paper was then published
in the Journal oI Fixed Income in 1991. A longer and richer paper was published in 1992 in the
Financial Analysts Journal (FAJ). The latter article was then republished by FAJ in the mid 1990's.
Copies oI the FAJ article are widely available on the Internet. It provides the rationale Ior the
methodology. and some inIormation on the derivation. but does not show all the Iormulas or a Iull
derivation. It also includes a rather complex worked example based on the global equilibrium. see
Litterman (2003) Ior more details on the methods required to solve this problem. UnIortunately.
because oI their merging the two problems. their results are diIIicult to reproduce.
The BlackLitterman model makes two signiIicant contributions to the problem oI asset allocation.
First. it provides an intuitive prior. the CAPM equilibrium market portIolio. as a starting point Ior
estimation oI asset returns. Previous similar work started either with the uninIormative uniIorm prior
distribution or with the global minimum variance portIolio. The latter method. described by Frost and
Savarino (1986). and Jorion (1986). took a shrinkage approach to improve the Iinal asset allocation.
Neither oI these methods has an intuitive connection back to the market.. The idea that one could use
'reverse optimization' to generate a stable distribution oI returns Irom the CAPM market portIolio as a
starting point is a signiIicant improvement to the process oI return estimation.
Second. the BlackLitterman model provides a clear way to speciIy investors views and to blend the
investors views with prior inIormation. The investor's views are allowed to be partial or complete. and
the views can span arbitrary and overlapping sets oI assets. The model estimates expected excess
returns and covariances which can be used as input to an optimizer. Prior to their paper. nothing
similar had been published. The mixing process (Bayesian and nonBayesian) had been studied. but
nobody had applied it to the problem oI estimating returns. No research linked the process oI
speciIying views to the blending oI the prior and the investors views. The BlackLitterman model
provides a quantitative Iramework Ior speciIying the investor's views. and a clear way to combine those
investor's views with an intuitive prior to arrive at a new combined distribution.
When used as part oI an asset allocation process. the BlackLitterman model leads to more stable and
1 The author grateIully acknowledges Ieedback and comments Irom Attilio Meucci and Boris Gnedenko.
Copyright 2009. Jay Walters 1
Electronic copy available at: http://ssrn.com/abstract=1314585 Electronic copy available at: http://ssrn.com/abstract=1314585
more diversiIied portIolios than plain meanvariance optimization. UnIortunately using this model
requires a broad variety oI data. some oI which may be hard to Iind.
First. the investor needs to identiIy their investable universe and Iind the market capitalization oI each
asset class. Then. they need to identiIy a time series oI returns Ior each asset class. and Ior the risk Iree
asset in order to compute a covariance matrix oI excess returns. OIten a proxy will be used Ior the asset
class. such as using a representative index. e.g. S&P 500 Index Ior US Domestic large cap equities.
The return on a short term sovereign bond. e.g US 13week treasury bill. would suIIice Ior most
United States investor's risk Iree rate.
Finding the market capitalization inIormation Ior liquid asset classes might be a challenge Ior an
individual investor. but likely presents little obstacle Ior an institutional investor because oI their access
to index inIormation Irom the various providers. Given the limited availability oI market capitalization
data Ior illiquid asset classes. e.g. real estate. private equity. commodities. even institutional investors
might have a diIIicult time piecing together adequate market capitalization inIormation. Return data Ior
these same asset classes can also be complicated by delays and inconsistencies in reporting. Further
complicating the problem is the question oI how to deal with hedge Iunds or absolute return managers.
The question oI whether they should be considered a separate asset class is beyond the scope oI this
paper.
Next. the investor needs to quantiIy their views so that they can be applied and new return estimates
computed. The views can be derived Irom quantitative or qualitative processes. and can be complete or
incomplete. or even conIlicting.
Finally. the outputs Irom the model need to be Ied into a portIolio optimizer to generate the eIIicient
Irontier. and an eIIicient portIolio selected. Bevan and Winkelmann (1999) provide a description oI
their asset allocation process (Ior international Iixed income) and how they use the BlackLitterman
model within that process. This includes their approaches to calibrating the model and inIormation on
how they compute the covariance matrices. Both Litterman. et al (2003) and Litterman and
Winkelmann (1998) provide details on the process used to compute covariance matrices at Goldman
Sachs.
The standard BlackLitterman model does not provide direct sensitivity oI the prior to market Iactors
besides the asset returns. It is Iairly simple to extend BlackLitterman to use a multiIactor model Ior
the prior distribution. Krishnan and Mains (2005) have provided extensions to the model which allow
adding additional cross asset class Iactors which are not priced in the market. Examples oI such Iactors
are a recession. or credit. market Iactor. Their approach is general and could be applied to other Iactors
iI desired.
Most oI the BlackLitterman literature reports results using the closed Iorm solution Ior unconstrained
optimization. They also tend to use nonextreme views in their examples. I believe this is done Ior
simplicity. but it is also a testament to the stability oI the outputs oI the BlackLitterman model that
useIul results can be generated via this process. As part oI a investment process. it is reasonable to
conclude that some constraints would be applied at least in terms oI restricting short selling and
limiting concentration in asset classes. Lack oI a budget constraint is also consistent with a Bayesian
investor who may not wish to be 100 invested in the market due to uncertainty about their belieIs in
the market. This is normally considered as part oI a two step process. Iirst compute the optimal
portIolio. and then determine position along the Capital Market Line.
For the ensuing discussion. we will describe the CAPM equilibrium distribution as the prior
distribution. and the investors views as the conditional distribution. This is consistent with the original
Copyright 2009. Jay Walters 2
Black and Litterman (1992) paper. It also is consistent with our intuition about the outcome in the
absence oI a conditional distribution (no views in BlackLitterman terminology.) This is the opposite
oI the way most examples oI Bayes Theorem are deIined. they start with a nonstatistical prior
distribution. and then add a sampled (statistical) distribution oI new data as the conditional distribution.
The mixing model we will use. and our use oI normal distributions. will bring us to the same outcome
independent oI these choices.
/0123141%1*'125+,16
The reIerence model Ior returns is the base upon which the rest oI BlackLitterman is built. It includes
the assumptions about which variables are random. and which are not. It also deIines which parameters
are modeled. and which are not modeled. Most importantly. many authors oI papers on the Black
Litterman model use an alternative reIerence model. not the one which was initially speciIied in Black
and Litterman (1992). or He and Litterman (1999).
We start with normally distributed expected returns
(1) E!r "#N !$. %"
The Iundamental goal oI the BlackLitterman model is to model these expected returns. which are
assumed to be normally distributed with mean and variance L. Note that we will need both oI these
values. the expected returns and covariance matrix later as inputs into a MeanVariance optimization.
We deIine . the mean return. as a random variable itselI distributed as
$#N !&. %
&
"
a is our estimate oI the mean and L
a
is the variance oI our estimate Irom the mean return . Another
way to view this simple linear relationship is shown in the Iormula below.
(2) &'$()
Formula (2) may seem to be incorrect with a on the leIt hand side. however our estimate (a) varies
around the actual value () with a disturbance value (c). so the Iormula is correctly speciIied.
c is normally distributed with mean 0 and variance L
a
. c is assumed to be uncorrelated with . We can
complete the reIerence model by deIining L
r
as the variance oI our estimate a. From Iormula (2) and
the assumption above that c and are not correlated. then the Iormula to compute L
r
is
(3)
%
r
'%(%
&
Formula (3) tells us that the proper relationship between the variances is (L
r
> L. L
a
).
We can check the reIerence model at the boundary conditions to ensure that it is correct. In the absence
oI estimation error. e.g. c 0 . then L
r
L. As our estimate gets worse. e.g. L
a
increases. then L
r
increases as well.
The reIerence model Ior the BlackLitterman model expected return is
(4)
E!r "#N !&. %
r
"
A common misconception about the BlackLitterman reIerence model is that Iormula (1) is the
reIerence model. and that is not random. We will address this model later. in the section entitled the
Alternate ReIerence Model. Many authors approach the problem Irom this point oI view so we cannot
neglect it. When considering results Irom BlackLitterman implementations it is important to
Copyright 2009. Jay Walters 3
understand which reIerence model is being used in order to understand how the various parameters will
impact the results.
7+89$.*:2$0127!;52<=.6."%.8231$%*#
As previously discussed. the prior distribution Ior the BlackLitterman model is the estimated mean
excess return Irom the CAPM equilibrium. The process oI computing the CAPM equilibrium excess
returns is straight Iorward.
CAPM is based on the concept that there is a linear relationship between risk (as measured by standard
deviation oI returns) and return. Further. it requires returns to be normally distributed. This model is oI
the Iorm
(5)
E!r "'r
f
(*r
m
(+
Where
r
f
The risk Iree rate.
r
m
The excess return oI the market portIolio.
* A regression coeIIicient computed as
*',

p

m
+ The residual. or asset speciIic (idiosyncratic) excess return.
Under the CAPM theory the idiosyncratic risk associated with an asset's u is uncorrelated with the u
Irom other assets and this risk can be reduced through diversiIication. Thus the investor is rewarded
Ior the systematic risk measured by . but is not rewarded Ior taking idiosyncratic risk associated with
u.
The Two Fund Separation Theorem. closely related to CAPM theory states that all investors should
hold two assets. the CAPM market portIolio and the risk Iree asset. The line drawn in standard
deviation/return space between the risk Iree rate and the CAPM market portIolio is called the Capital
Market Line. Depending on their risk aversion all investors will hold a portIolio on this line. with an
arbitrary Iraction oI their wealth in the risky asset. and the remainder in the riskIree asset. All
investors share the same risky portIolio. the CAPM market portIolio. The CAPM market portIolio is
on the eIIicient Irontier. and has the maximum Sharpe Ratio
2
oI any portIolio on the eIIicient Irontier.
All investors should hold a portIolio on this line. because they hold a mix oI the risk Iree asset and the
market portIolio. Because all investors hold only the market portIolio Ior their portIolio oI risky assets.
at equilibrium the market capitalizations oI the various assets will determine their weights in the market
portIolio.
Since we are starting with the market portIolio. we will be starting with a set oI weights which
naturally sum to 1. The market portIolio only includes risky assets. because by deIinition investors are
rewarded only Ior taking on systematic risk. In the CAPM model. the risk Iree asset with 0 will not
be in the market portIolio.
We will constrain the problem by asserting that the covariance matrix oI the returns. L. is known. In
practice. this covariance matrix is computed Irom historical return data. It could also be estimated.
however there are signiIicant issues involved in estimating a consistent covariance matrix. There is a
2 The Sharpe Ratio is the excess return divided by the excess risk. or (E(r) rI) / o.
Copyright 2009. Jay Walters 4
rich body oI research which claims that mean variance results are less sensitive to errors in estimating
the variance and that the population covariance is more stable over time than the returns. so relying on
historical covariance data should not introduce excessive model error. By computing it Irom actual
data we know that the resulting covariance matrix will be positive deIinite. It is possible when
estimating a covariance matrix to create one which is not positive deIinite. and thus notrealizable.
For the rest oI this section. we will use a common notation. similar to that used in He and Litterman
(1999) Ior all the terms in the Iormulas. Note that this notation is diIIerent. and conIlicts. with the
notation used in the section on Bayesian theory.
Here we derive the equations Ior 'reverse optimization' starting Irom the quadratic utility Iunction
(6)
U'w
T
./!
0
2
" w
T
%w
U Investors utility. this is the obiective Iunction during portIolio optimization.
w Vector oI weights invested in each asset
H Vector oI equilibrium excess returns Ior each asset
o Risk aversion parameter oI the market
L Covariance matrix Ior the assets
U is a concave Iunction. so it will have a single global maxima. II we maximize the utility with no
constraints. there is a closed Iorm solution. We Iind the exact solution by taking the Iirst derivative oI
(6) with respect to the weights (w) and setting it to 0.
dU
dw
'./0%w'0
Solving this Ior H (the vector oI excess returns) yields:
(7) H oLw
In order to use Iormula (7) we need to have a value Ior o . the risk aversion coeIIicient oI the market.
Most oI the authors speciIy the value oI o that they used. Bevan and Winkelmann (1998) describe their
process oI calibrating the returns to an average sharpe ratio based on their experience. For global Iixed
income (their area oI expertise) they use a sharpe ratio oI 1.0. Black and Litterman (1992) use a Sharpe
ratio closer to 0.5 in the example in their paper.
We can Iind o by multiplying both sides oI (7) by w
T
and replacing vector terms with scalar terms.
(E(r) r
I
) oo
2
(8) o (E(r) r
I
) / o
2
E(r) Total return on the market portIolio (E(r) w
T
H r
I
)
r
I
Risk Iree rate
o
2
Variance oI the market portIolio (o
2
w
T
Lw)
As part oI our analysis we must arrive at the terms on the right hand side oI Iormula (8); E(r). r
I
. and
o
2.
in order to calculate a value Ior o. Once we have a value Ior o. then we plug w. o and L into Iormula
(7) and generate the set oI equilibrium asset returns. Formula (7) is the closed Iorm solution to the
reverse optimization problem Ior computing asset returns given an optimal meanvariance portIolio in
the absence oI constraints. We can rearrange Iormula (7) to yield the Iormula Ior the closed Iorm
calculation oI the optimal portIolio weights in the absence oI constraints.
Copyright 2009. Jay Walters 5
(9) w'!0%"
/1
.
II we Ieed H. o. and L back into the Iormula (9). we can solve Ior the weights (w). II we instead used
historical excess returns rather than equilibrium excess returns. the results will be very sensitive to
changes in H. With the BlackLitterman model. the weight vector is less sensitive to the reverse
optimized H vector. This stability oI the optimization process. is one oI the strengths oI the Black
Litterman model.
Herold (2005) provides insights into how implied returns can be computed in the presence oI simple
equality constraints such as the budget or Iull investment (Lw 1) constraint.
The only missing piece is the variance oI our estimate oI the mean. Looking back at the reIerence
model. we need L
a
. Black and Litterman made the simpliIying assumption that the structure oI the
covariance matrix oI the estimate is proportional to the covariance oI the returns L. They created a
parameter. t. as the constant oI proportionality. Given that assumption.
L
a
tL. then the prior
distribution is:
(10) P(A) ~ N(H. tL)
This is the prior distribution Ior the BlackLitterman model. It represents our estimate oI the mean oI
the distribution oI excess returns.
>91'.4?.*:2$012@.1A#
This section will describe the process oI speciIying the investors views on the estimated mean excess
returns. We deIine the combination oI the investors views as the conditional distribution. First. by
construction we will require each view to be unique and uncorrelated with the other views. This will
give the conditional distribution the property that the covariance matrix will be diagonal. with all oII
diagonal entries equal to 0. We constrain the problem this way in order to improve the stability oI the
results and to simpliIy the problem. Estimating the covariances between views would be even more
complicated and error prone than estimating the view variances. Second. we will require views to be
Iully invested. either the sum oI weights in a view is zero (relative view) or is one (an absolute view).
We do not require a view on all assets. In addition it is actually possible Ior the views to conIlict. the
mixing process will merge the views based on the conIidence in the views and the conIidence in the
prior.
We will represent the investors k views on n assets using the Iollowing matrices
1 P. a kn matrix oI the asset weights within each view. For a relative view the sum oI the
weights will be 0. Ior an absolute view the sum oI the weights will be 1. DiIIerent authors
compute the various weights within the view diIIerently. He and Litterman (1999) and Idzorek
(2005) use a market capitalization weighed scheme. whereas Satchell and ScowcroIt (2000) use
an equal weighted scheme.
1 Q. a k1 matrix oI the returns Ior each view.
1 O a kk matrix oI the covariance oI the views. O is generally diagonal as the views are
required to be independent and uncorrelated. O
1
is known as the conIidence in the investor's
views. The ith diagonal element oI O is represented as e
i
.
We do not require P to be invertible. Meucci (2006) describes a method oI augmenting the matrices to
make the P matrix invertible while not changing the net results.
Copyright 2009. Jay Walters 6
O is symmetric and zero on all nondiagonal elements. but may also be zero on the diagonal iI the
investor is certain oI a view. This means that O may or may not be invertible. At a practical level we
can require that e ~ 0 so that O is invertible. but we will reIormulate the problem so that O is not
required to be invertible.
As an example oI how these matrices would be populated we will examine some investors views. Our
example will have Iour assets and two views. First. a relative view in which the investor believes that
Asset 1 will outperIorm Asset 3 by 2 with conIidence. e
1
. Second. an absolute view in which the
investor believes that Asset 2 will return 3 with conIidence e
2
. Note that the investor has no view on
Asset 4. and thus it's return should not be directly adiusted. These views are speciIied as Iollows:
P'
2
1 0 /1 0
0 1 0 0
3
; Q'
2
2
3
3
;
4'
2
5
11
0
0 5
22
3
Given this speciIication oI the views we can Iormulate the conditional distribution mean and variance
in view space as
P(B'A) ~ N(Q. O)
and in asset space as
(11) P(B'A) ~ N(P
1
Q. P
T
O
1
P
1
)
Remember that P may not be invertible. and even iI P is invertible P
T
O
1
P is probably not invertible.
making this expression impossible to evaluate in practice. Luckily. to work with the BlackLitterman
model we don't need to evaluate Iormula (11). It is interesting to see how the views are proiected into
the asset space.
O. the variance oI the views is inversely related to the investors conIidence in the views. however the
basic BlackLitterman model does not provide an intuitive way to quantiIy this relationship. It is up to
the investor to compute the variance oI the views O.
There are several ways to calculate O.
6 Proportional to the variance oI the prior
6 Use a conIidence interval
6 Use the variance oI residuals in a Iactor model
6 Use Idzorek's method to speciIy the conIidence along the weight dimension
;%+9+%$.+*&62$+2$012@&%.&*'12+42$012;%.+%
We can iust assume that the variance oI the views will be proportional to the variance oI the asset
returns. iust as the variance oI the prior distribution is. Both He and Litterman (1999). and Meucci
(2006) use this method. though they use it diIIerently. He and Litterman (1999) set the variance oI the
views as Iollow:
(12)
5
ii
'p!7%" p
T
8i' i
5
ii
'0 8i9 i
or
4'diag ! P!7%" P
T
"
Copyright 2009. Jay Walters 7
This speciIication oI the variance. or uncertainty. oI the views essentially equally weights the investor's
views and the market equilibrium weights. By including t in the expression. the Iinal solution becomes
independent oI t as well. This seems to be the most common method used in the literature.
Meucci (2006) doesn't bother with the diagonalization at all. and iust sets
4'
1
c
P %P
t
He sets c 1. and one obvious choice Ior c is t
1
.We will see later that this Iorm oI the variance oI the
views lends itselI to some simpliIications oI the BlackLitterman Iormulas.
B#12&27+*4.,1*'12)*$1%C&6
The investor can actually compute the variance oI the view. This is most easily done by deIining a
conIidence interval around the estimated mean return. e.g. Asset 2 has an estimated 3 mean return
with the expectation it is 68 likely to be within the interval (2.0.3.0). Knowing that 68 oI the
normal distribution Ialls within 1 standard deviation oI the mean. allows us to translate this into a
variance oI (0.010)
2
.
Here we are speciIying our uncertainty in the estimate oI the mean. we are not speciIying the variance
oI returns about the mean. This Iormulation oI the variance oI the view is consistent with using t 1
because the scale oI the uncertainty oI the prior will be on the same order as the uncertainty in the
view. For example. the standard deviation above is 1.0. and given t ~ 0.025 and a standard deviation
oI the prior returns oI 15. this would lead to the weight on the view being 6x the weight on the prior.
II t was 1. the views would be weighted 225x more than the weight on the prior. Understanding the
interplay between the selection oI t and the speciIication oI the variance oI the views is critical.
B#12$012@&%.&*'12+4231#.,&6#24%+82&2D&'$+%25+,16
II the investor is using a Iactor model to compute the views. they can use the variance oI the residuals
Irom the model to drive the variance oI the return estimates. The general expression Ior a Iactor model
oI returns is:
(13) E!r "'
:
i'1
n
*
i
f
i
()
Where
E!r " is the return oI the asset
*
i
is the Iactor loading Ior Iactor (i)
f
i
is the return due to Iactor (i)
) is an independent normally distributed residual
And the general expression Ior the variance oI the return Irom a Iactor model is:
(14) J ! r"'BJr ! F " B
T
(J !)"
B is the Iactor loading matrix
F is the vector oI returns due to the various Iactors
Given Iormula (13). and the assumption that c is independent and normally distributed. then we can
compute the variance oI c directly as part oI the regression. While the regression might yield a Iull
Copyright 2009. Jay Walters 8
covariance matrix. the mixing model will be more robust iI only the diagonal elements are used.
Beach and Orlov (2006) describe their work using GARCH style Iactor models to generate their views
Ior use with the BlackLitterman model. They generate the precision oI the views using the GARCH
models.
B#12),E+%1FG#251$0+,
Idzorek (2005) describes a method Ior speciIying the conIidence in the view in terms oI a percentage
move oI the weights on the interval Irom 0 conIidence to 100 conIidence. We will look at
Idzorek's algorithm in the section on extensions.
/012<#$.8&$.+*25+,16
The original BlackLitterman paper reIerences Theil's Mixed Estimation model rather than a Bayesian
estimation model. though we can get similar results Irom both methodologies. I chose to start with
Theil's model because it is simpler and cleaner. I also work through the Bayesian version oI the
derivations Ior completeness.
With either approach. we will be using the reIerence model to estimate the mean returns. not the
distribution oI returns. This is important in understanding the values used Ior t and O. and Ior the
computations oI the variance oI the prior and posterior distributions oI returns.
Another way to think oI the estimation model and the reIerence model is that while the estimated return
is more accurate the variance oI the distribution does not change because we have more data. The
prototypical example oI this would be to blend the distributions. P(A) ~ N(10. 20) and P(B'A) ~
N(12. 20). II we apply Bayes Iormula in a straightIorward Iashion. P(A'B) ~ N(11. 10).
Clearly with Iinancial data we did not really cut the variance oI the return distribution in iust because
we have a slightly better estimate oI the mean.
However. iI the mean is the random variable. and not the distribution. then our result oI P(A'B) ~
N(11. 10) makes sense. By blending these two estimates oI the mean. we have an estimate oI the
mean with much less uncertainty (less variance) than either oI the estimates.
Theil's Mixed Estimation Model
Theil's mixed estimation model was created Ior the purpose oI estimating parameters Irom a mixture oI
complete prior data and partial conditional data. This is a good Iit with our problem as it allows us to
express views on only a subset oI the asset returns. there is no requirement to express views on all oI
them. The views can also be expressed on a single asset. or on arbitrary combinations oI the assets.
The views do not even need be consistent. the estimation model will take each into account based on
the investors conIidence.
Theil's Mixed Estimation model starts Irom a linear model Ior the parameters to be estimated. We can
use Iormula (2) Irom our reIerence model as a starting point.
Our simple linear model is shown below:
(15) &'x *(u
Where
& is the n x 1 vector oI CAPM equilibrium returns Ior the assets.
Copyright 2009. Jay Walters 9
x is the n x n matrix I
n
which are the Iactor loadings Ior our model.
* is the n x 1 vector oI unknown means Ior the asset return process.
u is a n x n matrix oI residuals Irom the regression where E!u"'0, J !u"'E !u u"';
and is nonsingular.
The BlackLitterman model uses a very simple linear model. the expected return Ior each asset is
modeled by a single Iactor which has a coeIIicient oI 1. Thus. x. is the identity matrix. Given that
and u are independent. and x is constant. then we can model the variance oI a as:
J !&"'x J !*" x(J !u"
Which can be simpliIied to:
(16) J !&"'%(;
This ties back to Iormula (3) in the reIerence model. The total variance oI the estimated return is the
sum oI the variance oI the actual return process plus the variance oI the estimate oI the mean. We will
come back to this relation again later in the paper.
We will pragmatically compute L Irom historical data Ior the asset returns.
Next we consider some additional inIormation which we would like to combine with the prior. This
inIormation can be subiective views or can be derived Irom statistical data. We will also allow it to be
incomplete. meaning that we might not have an estimate Ior each asset return.
(17) q'p*(v
Where
q is the k x 1 vector oI returns Ior the views.
p is the k x n vector mapping the views onto the assets.
* is the n x 1 vector oI unknown means Ior the asset return process.
v is a k x k matrix oI oI residuals Irom the regression where
E!v"'0 , J ! v"'E !v v"'4 and O is nonsingular.
We can combine the prior and conditional inIormation by writing:
2
&
q
3
'
2
x
p
3
<
*(
2
u
v
3
Where the expected value oI the residual is 0. and the expected value oI the variance oI the residual is
J
!2
u
v
3"
'E
!2
u
v
3
2 u v 3
3
'
2
; 0
0 4
3
We can then apply the generalized least squares procedure. which leads to estimating
<
* as
<
*'
2
2 x p 3
2
; 0
0 4
3
/1
2
x
p
33
/1
2 x p 3
2
; 0
0 4
3
/1
2
&
q
3
This can be rewritten without the matrix notation as
(18) <
*'2 x ;
/1
x(p 4
/1
p3
/1
2 x ;
/1
&(p 4
/1
q3
Copyright 2009. Jay Walters 10
For the BlackLitterman model which is a single Iactor per asset. we can drop x as it is the identity
matrix. II one wanted to use a multiIactor model Ior the equilibrium. then x would be the equilibrium
Iactor loading matrix.
(19) <
*'2 ;
/1
(p 4
/1
p3
/1
2 ;
/1
&(p 4
/1
q3
This new
<
* is the weighted average oI the estimates. where the weighting Iactor is the inverse oI the
variance oI the estimate. It is also the best linear unbiased estimate given the data. and has the property
that it minimizes the variance oI the residual. Note that given a new
<
* . we also should have an
updated expectation Ior the variance oI the residual.
II we were using a Iactor model Ior the prior. then we would keep x the Iactor weightings in the
Iormulas. This would give us a multiIactor model. where all the Iactors will be priced into the
equilibrium.
We can reIormulate our combined relationship in terms oI our estimate oI
<
* and a new residual = u
as
2
&
q
3
'
2
x
p
3
<
*(= u
Once again E! = u"'0 . so we can derive the expression Ior the variance oI the new residual as:
(20)
J ! = u"'E! = u = u"'2;
/1
(p 4
/1
p3
/1
and the total variance is
J !2 v &3"'J !
<
*"(J ! = u"
We began this section by asserting that the variance oI the return process is a known quantity derived
Irom historical estimates. Improved estimation oI the quantity
<
* does not change our estimate oI the
variance oI the return distribution. L. Because oI our improved estimate. we do expect that the
variance oI the estimate (residual) has decreased. thus the total variance has changed.. We can simpliIy
the variance Iormula (16) to
(21) J !2 v &3"'%(J ! = u"
This is a clearly intuitive result. consistent with the realities oI Iinancial time series. We have
combined two estimates oI the mean oI a distribution to arrive at a better estimate oI the mean. The
variance oI this estimate has been reduced. but the actual variance oI the underlying process has not
changed. Given our uncertain estimate oI the process. the total variance oI our estimated process has
also improved incrementally. but it has the asymptotic limit that it cannot be less than the variance oI
the actual underlying process.
He and Litterman (1999) adopt this same convention Ior computing the variance oI the posterior
distribution.
In the absence oI views.Iormula (21) simpliIies to
J !2 v &3"'%(;
Which is the variance oI the prior distribution oI returns.
Appendix A contains a more detailed derivation oI Iormulas (19) and (20).
Copyright 2009. Jay Walters 11
A Quick Introduction to Bayes Theory
This section provides a quick overview oI the relevant portion oI Bayes theory in order to create a
common vocabulary which can be used in analyzing the BlackLitterman model Irom a Bayesian point
oI view.
Bayes theory states
(22) P! A>B"'
P! B>A" P! A"
P! B"
P(A'B) The conditional (or ioint) probability oI A. given B Also known as the posterior
distribution. We will call this the posterior distribution Irom here on.
P(B'A) The conditional probability oI B given A. Also know as the sampling
distribution. We will call this the conditional distribution Irom here on.
P(A) The probability oI A. Also known as the prior distribution. We will call this the
prior distribution Irom here on.
P(B) The probability oI B. Also known as the normalizing constant.
When actually applying this Iormula and solving Ior the posterior distribution. the normalizing constant
will disappear into the constants oI integration so Irom this point on we will ignore it..
A general problem in using Bayes theory is to identiIy an intuitive and tractable prior distribution. One
oI the core assumptions oI the BlackLitterman model (and MeanVariance optimization) is that asset
returns are normally distributed. For that reason we will conIine ourselves to the case oI normally
distributed conditional and prior distributions. Given that the inputs are normal distributions. then it
Iollows that the posterior will also be normally distributed. When the prior distribution and the
posterior have the same structure. the prior is known as a coniugate prior. Given interest there is
nothing to keep us Irom building variants oI the BlackLitterman model using diIIerent distributions.
however the normal distribution is generally the most straight Iorward.
Another core assumption oI the BlackLitterman model is that the variance oI the prior and the
conditional distributions about the actual mean are known. but the actual mean is not known. This
case. known as Unknown Mean and Known Variance is well documented in the Bayesian literature.
This matches the model which Theil uses where we have an uncertain estimate oI the mean. but know
the variance.
We deIine the signiIicant distributions below:
The prior distribution
(23) P(A) ~ N(x.S/n)
where S is the sample variance oI the distribution about the mean. with n samples then S/n is the
variance oI the estimate oI x about the mean.
The conditional distribution
(24) P(B'A) ~ N(.O)
O is the uncertainty in the estimate oI the mean. it is not the variance oI the distribution about the
mean.
Then the posterior distribution is speciIied by
Copyright 2009. Jay Walters 12
(25) P(A'B) ~ N(O
1
?@?nS
A!
x
T
O
1
nS
1
A!
.(O
1
nS
1
)
1
)
The variance term in (25) is the variance oI the estimated mean about the actual mean.
In Bayesian statistics the inverse oI the variance is known as the precision. We can describe the
posterior mean as the weighted mean oI the prior and conditional means. where the weighting Iactor is
the respective precision. Further. the posterior precision is the sum oI the prior and conditional
precision. Formula (25) requires that the precisions oI the prior and conditional both be noninIinite.
and that the sum is nonzero. InIinite precision corresponds to a variance oI 0. or absolute conIidence.
ero precision corresponds to inIinite variance. or total uncertainty.
A Iull derivation oI Iormula (25).using the PDF based Bayesian approach is shown in Appendix B.
As a Iirst check on the Iormulas we can test the boundary conditions to see iI they agree with our
intuition. II we examine Iormula (25) in the absence oI a conditional distribution. it should collapse
into the prior distribution.
~ N(nS
A!
x nS
1
A!
.(nS
1
)
1
)
(26) ~ N(x.S/n)
As we can see in Iormula (26). it does indeed collapse to the prior distribution. Another important
scenario is the case oI 100 certainty oI the conditional distribution. where S. or some portion oI it is
0. and thus S is not invertible. We can transIorm the returns and variance Irom Iormula (25) into a Iorm
which is more easy to work with in the 100 certainty case.
(27) P(A'B) ~ N(x (S/n)O?@?SBn
A!
x.(S/n) (S/n)(O S/n)
1
(S/n))
This transIormation relies on the result that (A
1
B
1
)
1
A A(AB)
1
A. It is easy to see that when S
is 0 (100 conIidence in the views) then the posterior variance will be 0. II O is positive inIinity (the
conIidence in the views is 0) then the posterior variance will be (S/n).
We will revisit equations (25) and (27) later in this paper where we transIorm these basic equations into
the various parts oI the BlackLitterman model. Appendices C and D contain derivations oI the
alternate BlackLitterman Iormulas Irom the standard Iorm. analogous to the transIormation Irom (25)
to (27).
Using Bayes Theorem for the Estimation Model
One oI the maior assumptions made by the BlackLitterman model is that the covariance oI the
estimated mean is proportional to the covariance oI the actual returns. The parameter t will serve as
the constant oI proportionality. It takes the place oI 1/n in Iormula (23). The new prior is:
(28) P(A) ~ N(H. tS)
This is the prior distribution Ior the BlackLitterman model.
We can now apply Bayes theory to the problem oI blending the prior and conditional distributions to
create a new posterior distribution oI the asset returns. Given equations (25). (28) and (11) we can
apply Bayes Theorem and derive our Iormula Ior the posterior distribution oI asset returns.
Substituting (28) and (11) into (25) we have the Iollowing distribution
(29) P(A'B) ~ N((tL)
1
H P
T
O
1
Q(tL)
1
P
T
O
1
P
1
.((tL)
1
P
T
O
1
P)
1
).
Copyright 2009. Jay Walters 13
This is sometimes reIerred to as the BlackLitterman master Iormula. A complete derivation oI the
Iormula is shown in Appendix B. An alternate representation oI the same Iormula Ior the mean returns
<
. and covariance (M) is
(30)
<
.'.(7 %P
T
2! P7 %P
T
"(43
/1
2 Q/P .3
(31) M'!!7%"
/1
(P
T
4
/1
P"
/1
The derivation oI Iormula (30) is shown in Appendix D. Remember that M. the posterior variance. is
the variance oI the posterior mean estimate about the actual mean. It is the uncertainty in the posterior
mean estimate. and is not the variance oI the returns. In order to compute the variance oI the returns so
that we can use it in a meanvariance optimizer we need to apply Iormula (0). This is mentioned in He
and Litterman (1999) but not in any oI the other papers.
(32) L
p
L M
Substituting the posterior variance Irom (31) we get
L
p
L ((tL)
1
P
T
O
1
P)
1
In the absence oI views this reduces to
(33) L
p
L (tL) (1 t)L
Thus when applying the BlackLitterman model in the absence oI views the variance oI the estimated
returns will be greater than the prior distribution variance. We see the impact oI this Iormula in the
results shown in He and Litterman (1999). In their results. the investor's weights sum to less than 1 iI
they have no views.
3
Idzorek (2005) and most other authors do not compute a new posterior variance.
but instead use the known input variance oI the returns about the mean. Several oI the authors set t 1
along with using the variance oI returns. We will consider these diIIerent reIerence models in a later
section.
Once we dig into the topic a little more we realize that iI we have only partial views. views on some
assets. then by using a posterior estimate oI the variance we will tilt the posterior weights towards
assets with lower variance (higher precision oI the estimated mean) and away Irom assets with higher
variance (lower precision oI the estimated mean). Thus the existence oI the views and the updated
covariance will tilt the optimizer towards using or not using those assets. This tilt will not be very large
iI we are working with a small value oI t. but it will be measurable.
Since we are building the covariance matrix. L. Irom historical data we can compute t Irom the number
oI samples. We can also estimate t based on our conIidence in the prior distribution. Note that both oI
these techniques provide some intuition Ior selecting a value oI t which is closer to 0 than to 1. Black
and Litterman (1992). He and Litterman (1999) and Idzorek (2005) all indicate that in their calculations
they used small values oI t. on the order oI 0.025 0.050. Satchell and ScowcroIt (2000) state that
many investors use a t around 1 which does not seem to have any intuitive connection here.
We can check our results by seeing iI the results match our intuition at the boundary conditions.
Given Iormula (30) it is easy to let O 0 showing that the return under 100 certainty oI the views is
(34)
<
.'.(%P
T
2 P % P
T
3
/1
2 Q/P.3
Thus under 100 certainty oI the views. the estimated return is insensitive to the value oI t used.
3 This is shown in table 4 and mentioned on page 11 oI He and Litterman (1999).
Copyright 2009. Jay Walters 14
Furthermore. iI P is invertible which means that it we have also oIIered a view on every asset. then
<
.'P
/1
Q
II the investor is not sure about their views. so O . then Iormula (30) reduces to
<
.'.
Finding an analytically tractable way to express and compute the posterior variance under 100
certainty is a challenging problem. Formula (25) above works only iI (P
T
O
1
P) is invertible which is not
usually the case because the posterior variance in asset space is also not usually tractable.
The alternate Iormula Ior the posterior variance derived Irom (27) is
(35) M'7 %/7% P
T
! P7% P
T
(4"
/1
P 7%
II O 0 (total conIidence in views. and every asset is in at least one view) then Iormula (35) can be
reduced to M 0. II on the other hand the investor is not conIident in their views. O . then
Iormula (35) can be reduced to M tL.
Meucci. 2005 describes the transIormation Irom (27) to (35). but does not show the Iull derivation. I
have included that derivation in Appendix B.
7&6."%&$.*:2H
This section will discuss some empirical ways to select and calibrate the value oI t.
The Iirst method to calibrate t relies on Ialling back to basic statistics. When estimating the mean oI a
distribution the uncertainty (variance) oI the mean estimate will be proportional to the number oI
samples. Given that we are estimating the covariance matrix Irom historical data. then
7'
1
T
The maximum likelihood estimator
7'
1
T /k
The best quadratic unbiased estimator
T The number oI samples
k The number oI assets
There are other estimators. but usually. the Iirst deIinition above is used. Given that we usually aim Ior
a number oI samples around 60 (5 years oI monthly samples) then t is on the order oI 0.02. This is
consistent with several oI the papers which indicate they used values oI t on the range (0.025. 0.05).
This is probably also consistent with the model to compute the variance oI the distribution. L.
We could instead calibrate t to the amount invested in the risk Iree asset given the prior distribution.
Here we see that the portIolio invested in risky assets given the prior views will be
w'.20!1(7" %3
/1
Thus the weights allocated to the assets are smaller by 1/(1t) than the CAPM market weights. This
is because our Bayesian investor is uncertain in their estimate oI the prior. and they do not want to be
100 invested in risky assets.
Copyright 2009. Jay Walters 15
The Alternative Reference Model
This section will discuss the most common alternative reIerence model used with the BlackLitterman
estimation model.
The most common alternative reIerence model is the one used in Satchell and ScowcroIt (2000). and in
the work oI Meucci.
E!r "#N !$. %"
In this reIerence model. is normally distributed with variance L. We estimate . but the mean itselI.
. is not considered a random variable. This is commonly described as having a t 1. but more
precisely we have eliminated t as a parameter. In this model O becomes the covariance oI the views
around the investor's estimate oI the mean return. iust as L is the covariance oI the prior about it's
mean. Given these speciIications Ior the covariances oI the prior and conditional. it is inIeasible to
have an updated covariance Ior the posterior. For example. it would require that the covariance oI the
posterior is the covariance oI the prior and conditional iI they had equal covariances. This conIlicts
with our earlier statements that higher precision in our estimate oI the mean return doesn't cause the
covariance oI the return distribution to shrink by the same amount.
The primary artiIacts oI this new reIerence model are. Iirst t is nonexistant. and second. the investor's
portIolio weights in the absence oI views equal the CAPM portIolio weights. Finally at
implementation time there is no need or use oI Iormulas (31) or (32).
Note that none oI the authors prior to Meucci (2008) except Ior Black and Litterman (1992). and He
and Litterman (1999) make any mention oI the details oI the BlackLitterman reIerence model. or oI
the diIIerent reIerence model used by most authors. It is unclear to this author why this has occurred.
In the BlackLitterman reIerence model. the updated posterior variance oI the mean estimate will be
lower than either the prior or conditional variance oI the mean estimate. indicating that the addition oI
more inIormation will reduce the uncertainty oI the model. The variance oI the returns Irom Iormula
(32) will never be less than the prior variance oI returns. This matches our intuition as adding more
inIormation should reduce the uncertainty oI the estimates. Given that there is some uncertainty in this
value (M). then Iormula (32) provides a better estimator oI the variance oI returns than the prior
variance oI returns.
/012)89&'$2+42H
The meaning and impact oI the parameter t causes a great deal oI conIusion Ior many users oI the
BlackLitterman model. In the literature we appear to see authors divided into two groups over t. In
reality. t is not the signiIicant diIIerence. the reIerence model is the diIIerence and the author's
speciIication oI t iust an artiIact oI the reIerence model they use.
The Iirst group thinks t should be a small number on the order oI 0.0250.05. and includes He and
Litterman (1999). Black and Litterman (1992) and Idzorek(2004). The second group thinks t should be
near 1. or eliminates it Irom the model. and includes Satchell and ScowcroIt (2000). Meucci (2005) and
others.
A better division oI the authors has to do with the reIerence model. The Goldman Sachs papers. He
and Litterman (1999) and Black and Litterman (1992) use the BlackLitterman reIerence model with t.
All the other authors use the alternative reIerence model described above. and either eliminate t
(Meucci). set it to 1 (Satchell and ScowcroIt. or calibrate the model to it (Idzorek). Satchell and
Copyright 2009. Jay Walters 16
ScowcroIt (2000) describe a model with a stochastic t. but this is really stochastic variance oI returns.
Given the BlackLitterman reIerence model we can still perIorm an exercise to understand the impact
oI t on the results. We will start with the expression Ior O similar to the one used by He and Litterman
(1999). Rather than using only the diagonal. we will retain the entire structure oI the covariance matrix
to make the math simpler and more clear.
(36) 4'P!7 %" P
T
We can substitute this into Iormula (30) as
<
.'.(7 %P
T
2! P7 %P
T
"(43
/1
2 Q/P .
T
3
'.(7 %P
T
2! P7 %P
T
"(! P 7% P
T
"3
/1
2Q/P .
T
3
'.(7 %P
T
2 2! P 7% P
T
"3
/1
2Q/P .
T
3
'.(!
1
2
"7 %P
T
! P
T
"
/1
2 P7 %3
/1
2 Q/P .
T
3
'.(!
1
2
"7 %!7%"
/1
P
/1
2Q/P .
T
3
'.(!
1
2
" P
/1
2 Q/P.
T
3
(37)
<
.'.(!
1
2
"2 P
/1
Q/.
T
3
Clearly using Iormula (36) is iust a simpliIication and does not do iustice to investors views. but we can
still see that setting O proportional to t. will eliminate t Irom the Iinal Iormula Ior
<
. . In the
general Iorm iI we Iormulate O as
(38) 4'P!+7%" P
T
Then we can rewrite Iormula (37) as
(39)
<
.'.(!
1
1(+
"2 P
/1
Q/.3
We can see a similar result iI we substitute Iormula (36) into Iormula (35).
M'7%/7% P
T
2 P7%P
T
(43
/1
P 7%
'7%/7%P
T
2 P 7% P
T
(P7% P
T
3
/1
P7%
'7%/7%P
T
2 2! P7%P
T
"3
/1
P7%
'7%/!
1
2
"!7%"! P
T
"! P
T
"
/1
!7%"
/1
! P"
/1
! P"!7%"
'7%/!
1
2
"7%
(40)
M'!
1
2
" 7%
Note that t is not eliminated Irom Iormula (40). We can also observe that iI t is on the order oI 1 and
we were to use Iormula (32) that the uncertainty in the estimate oI the mean would be a signiIicant
Iraction oI the variance oI the returns. With the alternative reIerence model. no posterior variance
calculations are perIormed and the mixing is weighted by the variance oI returns.
Copyright 2009. Jay Walters 17
In both cases. our choice Ior O has evenly weighted the prior and conditional distributions in the
estimation oI the posterior distribution. This matches our intuition when we consider we have blended
two inputs. Ior both oI which we have the same level oI uncertainty. The posterior distribution will be
the average oI the two distributions.
II we instead solve Ior the more useIul general case oI O uP(tL)P
T
where u > 0. substituting into
(30) and Iollowing the same logic as used to derive (40) we get
(41)
<
.'.(
1
!1(+"
2 P
/1
Q/.3
This parameterization oI the uncertainty is speciIied in Meucci (2005) and it allows us an option
between using the same uncertainty Ior the prior and views. and having to speciIy a separate and
unique uncertainty Ior each view. Given that we are essentially multiplying the prior covariance matrix
by a constant this parameterization oI the uncertainty oI the views does not have a negative impact on
the stability oI the results.
Note that this speciIication oI the uncertainty in the views changes the assumption Irom the views
being uncorrelated. to the views having the same correlations as the prior returns.
In summary. iI the investor uses the alternative reIerence model and makes O proportional to tL. then
they need only calibrate the constant oI proportionality. u. which indicates their relative conIidence in
their views versus the equilibrium. II they use the BlackLitterman reIerence model and set O
proportional to tL. then return estimate will not depend on the value oI t. but the posterior covariance
oI returns will depend on the proper calibration oI t.
An Asset Allocation Process
The BlackLitterman model is iust one part oI an asset allocation process. Bevan and Winkelmann
(1998) document the asset allocation process they use in the Fixed Income Group at Goldman Sachs.
At a minimum. a BlackLitterman oriented investment process would have the Iollowing steps:
6 Determine which assets constitute the market
6 Compute the historical covariance matrix Ior the assets
6 Determine the market capitalization Ior each asset class.
6 Use reverse optimization to compute the CAPM equilibrium returns Ior the assets
6 SpeciIy views on the market
6 Blend the CAPM equilibrium returns with the views using the BlackLitterman model
6 Feed the estimates (estimated returns. covariances) generated by the BlackLitterman model
into a portIolio optimizer.
6 Select the eIIicient portIolio which matches the investors risk preIerences
A Iurther discussion oI each step is provided below.
The Iirst step is to determine the scope oI the market. For an asset allocation exercise this would be
identiIying the individual asset classes to be considered. For each asset class the weight oI the asset
class in the market portIolio is required. Then a suitable proxy return series Ior the excess returns oI
the asset class is required. Between these two requirements it can be very diIIicult to integrate illiquid
Copyright 2009. Jay Walters 18
asset classes such as private equity or real estate into the model. Furthermore. separating public real
estate holdings Irom equity holdings (e.g. REITS in the S&P 500 index) may also be required. Idzorek
(2006) provides an example oI the analysis required to include commodities as an asset class.
Once the proxy return series have been identiIied. and returns in excess oI the risk Iree rate have been
calculated. then a covariance matrix can be calculated. Typically the covariance matrix is calculated
Irom the highest Irequency data available. e.g. daily. and then scaled up to the appropriate time Irame.
Investor's oIten use an exponential weighting scheme to provide increased weights to more recent data
and less to older data. Other Iiltering (Random Matrix Theory) or shrinkage methods could also be
used in an attempt to impart additional stability to the process.
Now we can run a reverse optimization on the market portIolio to compute the equilibrium excess
returns Ior each asset class. Part oI this step includes computing a o value Ior the market portIolio. This
can be calculated Irom the return and standard deviation oI the market portIolio. Bevan and
Winkelmann (1998) discuss the use oI an expected Sharpe Ratio target Ior the calibration oI o. For
their international Iixed income investments they used an expect Sharpe Ratio oI 1.0 Ior the market.
The investor then needs to calibrate t in some manner. This value is usually on the order oI 0.025 ~
0.050.
At this point almost all oI the machinery is in place. The investor needs to speciIy views on the market.
These views can impact one or more assets. in any combination. The views can be consistent. or they
can conIlict. An example oI conIlicting views would be merging opinions Irom multiple analysts.
where they may not all agree. The investor needs to speciIy the assets involved in each view. the
absolute or relative return oI the view. and their uncertainty in the return Ior the view consistent with
their reIerence model and measured by one oI the methods discussed previously.
Appendix E shows the process oI cranking through Iormulas (30). (35) and (32) to compute the new
posterior estimate oI the returns and the covariance oI the posterior returns. These values will be the
inputs to some type oI optimizer. a meanvariance optimizer being the most common. II the user
generates the optimal portIolios Ior a series oI returns. then they can plot an eIIicient Irontier/
31#6$#
This section oI the document will step through a comparison oI the results oI the various authors. The
iava programs used to compute these results are all available as part oI the akutan open source Iinance
proiect at sourceIorge.net. All oI the mathematical Iunctions were built using the Colt open source
numerics library Ior Java. Any small diIIerences between my results and the authors reported results
are most likely the result oI rounding oI inputs and/or results.
When reporting results most authors have iust reported the portIolio weights Irom an unconstrained
optimization using the posterior mean and variance. Given that the vector H is the excess return vector.
then we do not need a budget constraint (Lw
i
1) as we can saIely assume any 'missing' weight is
invested in the risk Iree asset which has expected return 0 and variance 0. This calculation comes Irom
Iormula (9).
w H(oL
p
)
1
As a Iirst test oI our algorithm we veriIy that when the investor has no views that the weights are
correct. substituting Iormula (33) into (9) we get
w
nv
H(o(1t)L)
1
Copyright 2009. Jay Walters 19
(42) w
nv
w/(1t)
Given this result. it is clear that the output weights with no views will be impacted by the choice oI t
when the BlackLitterman reIerence model is used. He and Litterman (1999) indicate that iI our
investor is a Bayesian. then they will not be certain oI the prior distribution and thus would not be Iully
invested in the risky portIolio at the start. This is consistent with Iormula (42).
5&$'0.*:2$01231#6$#2+42I12&*,2J.$$1%8&*
First we will consider the results shown in He and Litterman (1999). These results are the easiest to
reproduce and they seem to stay close to the model described in the original Black and Litterman
(1992) paper. As Robert Litterman coauthored both this paper and the original paper. it makes sense
they would be consistent.
He and Litterman (1999) set
(43) O diag(P
T
(tL)P)
This essentially makes the uncertainty oI the views equivalent to the uncertainty oI the equilibrium
estimates. They select a small value Ior t (0.05). and they use the BlackLitterman reIerence model
and the updated posterior variance oI returns as calculated in Iormulas (35) and (32).
Table 1 These results correspond to Table 7 in He and Litterman. 1999.
Asset P0 P1 weq/(1t) w w  weq/(1t)
Australia 0.0 0.0 4.3 16.4 1.5 .0
Canada 0.0 1.0 8.9 2.1 53.9 51.8
France 0.295 0.0 9.3 5.0 .5 5.4
Germany 1.0 0.0 10.6 5.2 23.6 18.4
Japan 0.0 0.0 4.6 11.0 11.0 .0
UK 0.705 0.0 6.9 11.8 1.1 13.0
USA 0.0 1.0 7.1 58.6 6.8 51.8
q 5.0 4.0
e/t .043 .017
.193 .544
Table 1 contains results computed using the akutan implementation oI BlackLitterman and the input
data Ior the equilibrium case and the investor's views Irom He and Litterman (1999). The values
shown Ior w exactly match the values shown in their paper.
5&$'0.*:2$01231#6$#2+42),E+%1F
This section oI the document describes the eIIorts to reproduce the results oI Idzorek (2005). In trying
Copyright 2009. Jay Walters 20
to match Idzorek's results I Iound that he used the alternative reIerence model. which leaves L. the
known variance oI the returns Irom the prior distribution. as the variance oI the posterior returms. This
is a signiIicant diIIerence Irom the algorithm used in He and Litterman (1999). but in the end given that
Idzorek used a small value oI t. the diIIerences amounted to approximately 50 basis points per asset.
Tables 2 and 3 below illustrate computed results with the data Irom his paper and how the results diIIer
between the two versions oI the model.
Table 2 contains results generated using the data Irom Idzorek (2005) and the BlackLitterman model
as described by He and Litterman (1999). Table 3 shows the same results as generated by the
alternative reIerence model variant oI the algorithm. This alternative model variant appears to be the
method used by Idzorek.
Table 2 He and Litterman version oI BlackLitterman Model with Idzorek data (Corresponds with
data in Idzorek's Table 6).
Asset Class
w
eq
w BlackLitterman
ReIerence Model
Idzorek's
Results
US Bonds .07 18.87 28.96 10.09 10.54
Intl Bonds .50 25.49 15.41 10.09 10.54
US LG 6.50 11.80 9.27 2.52 2.73
US LV 4.33 11.80 14.32 2.52 2.73
US SG 7.55 1.31 1.03 .28 0.30
US SV 3.94 1.31 1.59 .28 0.30
Intl Dev 4.94 23.59 27.74 4.15 3.63
Intl Emg 6.84 3.40 3.40 .0 0
Note that the results in Table 2 are close. but Ior several oI the assets the diIIerence is about 50 basis
points. The values shown in Table 3 are within 4 basis points. essentially matching the results reported
by Idzorek.
Copyright 2009. Jay Walters 21
Table 3 Alternative ReIerence Model version oI the BlackLitterman Model with Idzorek data
(Corresponds with data in Idzorek's Table 6).
Country
w
eq
w Alternative
ReIerence Model
Idzorek's
Results
US Bonds .07 19.34 29.89 10.55 10.54
Intl Bonds .50 26.13 15.58 10.55 10.54
US LG 6.50 12.09 9.37 2.72 2.73
US LV 4.33 12.09 14.81 2.72 2.73
US SG 7.55 1.34 1.04 .30 0.30
US SV 3.94 1.34 1.64 .30 0.30
Intl Dev 4.94 24.18 27.77 3.59 3.63
Intl Emg 6.84 3.49 3.49 .0 0
K*25&*F1%$G#2>&896.*:2/01+%1$.'2!*&6?#.#
Mankert (2006) derives the BlackLitterman 'master Iormula' using a Sampling Theory approach which
yields t as a ratio oI the conIidence in the prior to the conIidence in the sampling distribution. She also
provides a Iull derivation oI Iormula (30) Irom Iormula (29).
Her derivation is valid Ior the the Alternative ReIerence Model. which is the model most used in the
literature. Further. many authors set O is proportional to P(L)P
T
which is consistent with her work. We
will show that her thesis holds Ior the alternative reIerence model. but does not Ior the BlackLitterman
reIerence model.
Given the derivation oI the posterior distribution created by the mixing oI a normally distributed
statistical prior and a subiective conditional shown in Appendix A. we can by replacing the subiect
conditional with a statistical conditional distribution arrive at Iormula (25) by the same logic.
Given
P! A"#N ! x
p
.
S
p
n
" . givenn samples
P! B>A"#N ! x
c
.
S
c
m
" . givenm samples
then
(44)
P! A>B"#N!
2
!
S
p
n
"
/1
(!
S
c
m
"
/1
3
/1
2
x
p
!
S
p
n
"
/1
(x
c
!
S
c
m
"
/1
3
.
2
!
S
p
n
"
/1
(!
S
c
m
"
/1
3
/1
"
Next we want to attempt to move the (m) terms out oI each term so we can eliminate them yielding:
Copyright 2009. Jay Walters 22
P! A>B"#N!
2
m
2
! S
p
!
m
n
""
/1
(! S
c
"
/1
33
/1
2
m
2
x
p
! S
p
!
m
n
""
/1
(x
c
! S
c
"
/1
33
.
2
m
2
!S
p
!
m
n
""
/1
(! S
c
"
/1
33
/1
"
In the mean term the m's cancel out and we see the Iormula Mankert used to support her thesis
2
!S
p
!
m
n
""
/1
(! S
c
"
/1
3
/1
2
x
p
!S
p
!
m
n
""
/1
(x
c
! S
c
"
/1
3
Thus iI we set L S
p
and O S
c
then we can let t m/n. the ratio oI uncertainty in the views to the
uncertainty in the prior. For the alternative reIerence model. her approach is valid.
However in the posterior variance oI the estimate an extra m term remains. this makes the covariance
term
!
1
m
"
2
!
m
n
S
p
"
/1
(!S
c
"
/1
3
/1
This does not line up with Iormula (32) as we would expect. given the substitutions in the paragraph
above. Her approach is not consistent with the BlackLitterman reIerence model.
!,,.$.+*&62L+%F
This section provides a brieI discussion oI eIIorts to reproduce results Irom some oI the maior research
papers on the BlackLitterman model.
OI the maior papers on the BlackLitterman model. there are two which would be very useIul to
reproduce. Satchell and ScowcroIt (2000) and Black and Litterman (1992). Satchell and ScowcroIt
(2000) does not provide enough data in their paper to reproduce their results. They have several
examples. one with 11 countries equity returns plus currency returns. and one with IiIteen countries.
They don't provide the covariance matrix Ior either example. and so their analysis cannot be
reproduced. It would be interesting to conIirm that they use the alternative reIerence model by
reproducing their results.
Black and Litterman (1992) do provide what seems to be all the inputs to their analysis. however they
chose a nontrivial example including partially hedged equity and bond returns. This requires the
application oI some constraints to the reverse optimization process which I have been unable to
Iormulate as oI this time. I plan on continuing this work with the goal oI veriIying the details oI the
BlackLitterman implementation used in Black and Litterman (1992).
Extensions to the BlackLitterman Model
In this section I will cover the extensions to the BlackLitterman model proposed in Idzorek (2005).
Fusai and Meucci (2003) and Krishnan and Mains (2006).
Idzorek (2005) presents a means to calibrate the conIidence or variance oI the investors views in a
simple and straightIorward method. Fusai and Meucci (2003) propose a way to measure how
consistent a posterior estimate oI the mean is with regards to the prior. or some other estimate. Braga
and Natale (2007) describe how to use Tracking Error to measure the distance Irom the equilibirum to
the posterior portIolio. Krishnan and Mains (2006) present a method to incorporate additional Iactors
into the model.
Copyright 2009. Jay Walters 23
),E+%1FG#2<M$1*#.+*
Idzorek's apparent goal was to reduce the complexity oI the BlackLitterman model Ior non
quantitative investors. He achieves this by allowing the investor to speciIy the investors conIidence in
the views as a percentage (0100) where the conIidence measures the change in weight oI the
posterior Irom the prior estimate (0) to the conditional estimate (100). This linear relation is shown
below
(45)
confidence'! < w/w
mkt
"B!w
100
/w
mkt
"
w
100
is the weight oI the asset under 100 certainty in the view
w
mkt
is the weight oI the asset under no views
w is the weight oI the asset under the speciIied view.
He provides a method to back out the value oI e required to generate the proper tilt (change in weights
Irom prior to posterior) Ior each view. These values are then combined to Iorm O. and the model is
used to compute posterior estimates. A side eIIect oI this method is that it is insensitive to the choice
oI t.
In his paper he discusses solving Ior e using a least squares method. We can actually solve this
analytically. The next section will provide a derivation oI the Iormulas required Ior this solution.
First we will use the Iollowing Iorm oI the uncertainty oI the views
(46) 4'+ P %P
T
u. the coeIIicient oI uncertainty. is a scalar quantity in the interval 0. . When the investor is 100
conIident in their views. then u will be 0. and when they are totally uncertain then u will be . Note
that Iormula (46) is exact. it is identical to Iormula (43) the O used by He and Litterman (1999)
because it is a 1x1 matrix. This allows us to Iind a closed Iorm solution to the problem oI O Ior
Idzorek's conIidence.
First we substitute Iormula (41)
<
.'.(
1
!1(+"
2 P
/1
Q/.3
into Iormula (9).
< w'
<
.!0%"
/1
Which yields
(47)
< w'
2
.(
1
!1(+"
2 P
/1
Q/.3
3
!0%"
/1
Now we can solve Iormula (47) at the boundary conditions Ior u.
lim
+CD
. w
mk
'.!0%"
/1
lim
+C0
. w
100
'P
/1
Q!0 %"
/1
And recombining some oI the terms in (47) we arrive at
Copyright 2009. Jay Walters 24
(48)
< w'.!0%"
/1
(
2
1
!1(+"
3
2 P
/1
Q!0%"
/1
/.!0%"
/1
3
Substituting w
mk
and w
100
back into (48) we get
< w'w
mk
(
2
1
!1(+"
3
2
w
100
/w
mk
3
And comparing the above with Iormula (45)
confidence'! < w/w
mk
"B !w
100
/w
mk
"
We see that
confidence'
1
!1(+"
And iI we solve Ior u
(49) +'!1/confidence"B confidence
Using Iormulas (49) and (46) the investor can easily calculate the value oI e Ior each view. and then
roll them up into a single O matrix. To check the results Ior each view. we then solve Ior the posterior
estimated returns using Iormula (30) and plug them back into Iormula (45). Note that when the
investor applies all their views at once. the interaction amongst the views will pull the posterior away
Irom the results generated when the views were taken one at a time.
Idzorek's method greatly simpliIies the investor's process oI speciIying the uncertainty in the views
when the investor does not have a quantitative model driving the process. In addition. this model does
not add meaningIul complexity to the process.
!*2<M&89612+42),E+%1FG#2<M$1*#.+*
Idzorek describes the steps required to implement his extension in his paper. but does not provide a
worked example. In this section I will work through his example Irom where he leaves oII in the
paper.
Idzorek's example includes 3 views:
1 International Dev Equity will have absolute excess return oI 5.25. ConIidence 25.0
1 International Bonds will outperIorm US bonds by 25bps. ConIidence 50.0
1 US Growth Equity will outperIorm US Value Equity by 2. ConIidence 65.0
In his paper Idzorek deIines the steps in his method which include calculations oI w
100
and then the
calculation oI e Ior each view given the desired change in the weights. From the previous section. we
can see that we only need to take the investor's conIidence Ior each view. plug it into Iormula (49) and
compute the value oI alpha. Then we plug u. P and L into Iormula (46) and compute the value oI e Ior
each view. At this point we can assemble our O matrix and proceed to solve Ior the posterior returns
using Iormula (29) or (30).
In working the example. I will show the results Ior each view including the w
mkt
and w
100
in order to
make the workings oI the extension more transparent. Tables 4. 5 and 6 below each show the results
Ior a single view.
Copyright 2009. Jay Walters 25
Table 4 Calibrated Results Ior View 1
Asset e w
mkt
w w
100
Implied ConIidence
Intl Dev Equity .002126625 24.18 25.46 29.28 25.00
Table 5 Calibrated Results Ior View 2
Asset e w
mkt
w w
100
Implied ConIidence
US Bonds .000140650 19.34 29.06 38.78 50.00
Intl Bonds .000140650 26.13 16.41 6.69 50.00
Table 6 Calibrated Results Ior View 3
Asset e w
mkt
w w
100
Implied ConIidence
US LG .000466108 12.09 9.49 8.09 65.00
US LV .000466108 12.09 14.69 16.09 65.00
US SG .000466108 1.34 1.05 .90 65.00
US SV .000466108 1.34 1.63 1.78 65.00
Then we use the Ireshly computed values Ior the O matrix with all views speciIied together and arrive
at the Iollowing Iinal result shown in Table 7 blending all 3 views together.
Table 7 Final Results Ior Idzorek's ConIidence Extension Example
Asset View 1 View 2 View 3 o w
mkt
Posterior
Weight
change
US
Bonds
0.0 1.0 0.0 .1 3.2 19.3 29.6 10.3
Intl
Bonds
0.0 1.0 0.0 .5 8.5 26.1 15.8 10.3
US LG 0.0 0.0 0.9 6.3 24.5 12.1 8.9 3.2
US LV 0.0 0.0 0.9 4.2 17.2 12.1 15.2 3.2
US SG 0.0 0.0 0.1 7.3 32.0 1.3 1.0 .4
US SV 0.0 0.0 0.1 3.8 17.9 1.3 1.7 .4
Copyright 2009. Jay Walters 26
Asset View 1 View 2 View 3 o w
mkt
Posterior
Weight
change
Intl Dev 1.0 0.0 0.0 4.8 16.8 24.2 26.0 1.8
Intl Emg 0.0 0.0 0.0 6.6 28.3 3.5 3.5 .0
Total 101.8
Return 5.2 .2 2.0
Omega/
tau
.08507 .00563 .01864
Lambda .002 .006 .002
51&#%.*:2$012)89&'$2+42$012@.1A#
This section will discuss several methods used in the literature to measure the impact oI the views on
the posterior distribution. In general we can divide these measures into two groups. The Iirst group
allows us to test the hypothesis that the views or posterior contradict the prior. The second group
allows us to measure a distance or inIormation content between the prior and posterior.
Theil (1971). and Fusai and Meucci (2003) describe measures that are designed to allow a hypothesis
test to ensure the views or the posterior does not contradict the prior estimates. Theil (1971) describes
a method oI perIorming a hypothesis test to veriIy that the views are compatibile with the prior. We
will extend that work to measure compatibility oI the posterior and the prior. Fusai and Meucci (2003)
describe a method Ior testing the compatibility oI the posterior and prior when using the alternative
reIerence model.
He and Litterman (1999). and Braga and Natale (2007) describe measures which can be used to
measure the distance between two distributions. or the amount oI tilt between the prior and the
posterior. These measures don't lend themselves to hypothesis testing. but they can potentially be used
as constraints on the optimization process. He and Litterman (1999) deIine a metric. . which
measures the tilt induced in the posterior by each view. Braga and Natale (2007) use Tracking Error
Volatility (TEV) to measure the distance Irom the prior to the posterior. We will also introduce the
concept oI relative entropy Irom inIormation theory. and use the KullbackLeibler distance
4
to measure
the relative entropy between the prior and the posterior.
Theil's Measure of Compatibility Between the Views and the Prior
Theil (1971) describes this as testing the compatibility oI the views with the prior inIormation. Given
the linear mixed estimation model. we have the prior (15) and the conditional (17).
(15) &'x *(u
(17) q'p*(v
The mixed estimation model deIines u as a random vector with mean 0 and covariance tL. and v as a
4 It is not really a true distance as it is not symmetric. but we will reIer to it as a distance in this paper.
Copyright 2009. Jay Walters 27
random vector with mean 0 and covariance O.
The approach we will take is very similar to the approach taken when analyzing a prediction Irom a
linear regression.
We can deIine an estimator Ior . we will call this estimator
<
* . computed only Irom the inIormation
in the views. We can measure the estimation error between the prior and the views as:
(50) E'!&/x
<
*"'/x!
<
*/*"(u
The vector has mean 0 and variance V(). We will Iorm our hypothesis test using the Iormulation
(51) F'E!E"J !E"
/1
E!E"
The quantity. . is known as the Mahalanobis distance (multidimensional analog oI the zscore) and is
distributed as
2
(n). In order to use this Iorm we need to solve Ior the E() and V().
II we consider only the inIormation in the views. the estimator oI is:
<
*'! P
T
4
/1
P"
/1
P
T
4
/1
Q
Note that since P is not required to be a matrix oI Iull rank that we might not be able to evaluate this
Iormula as written. We work in return space here (as opposed to view space) as it seems more natural.
Later on we will transIorm the Iormula into view space to make it computable.
We then substitute the new estimator into the Iormula (50) and eliminate x as it is the identity matrix in
the BlackLitterman application oI mixedestimation.
(52) E'/! P
T
4
/1
P"
/1
P
T
4
/1
Q(*(u
Next we substitute Iormula (17) Ior Q.
E'/! P
T
4
/1
P"
/1
P
T
4
/1
! P *(v"(*(u
E'/! P
T
4
/1
P"
/1
! P
T
4
/1
P" *(! P
T
4
/1
P"
/1
P
T
4
/1
v(*(u
E'/! P
T
4
/1
P"
/1
P
T
4
/1
v(u
Give our estimator. we want to Iind the variance oI the estimator.
J !E"'E !EE
T
"
J !E"'E !! P
T
4
/1
P"
/1
P
T
4
/1
v v
T
4
/1
P! P
T
4
/1
P"
/1
/2! P
T
4
/1
P"
/1
P
T
4
/1
v u
T
(uu
T
"
But E(vu) 0 so we can eliminate the cross term. and simpliIy the Iormula.
J !E"'E 2! P
T
4
/1
P"
/1
P
T
4
/1
v v
T
4
/1
P! P
T
4
/1
P"
/1
(uu
T
3
J !E"'E 2! P
/1
v v
T
! P
T
"
/1
"(uu
T
3
J !E"'! P
T
4
/1
P"
/1
(7%
The last step is to take the expectation oI Iormula (52). At the same time we will substitute the prior
estimate (H) Ior .
E!E"'./! P
T
4
/1
P"
/1
P
T
4
/1
Q
Now substitute the various values into (51) as Iollows
F'! ./! P
T
4
/1
P"
/1
P
T
4
/1
Q"2 ! P
T
4
/1
P"
/1
(7 %3
/1
!./! P
T
4
/1
P"
/1
P
T
4
/1
Q"
T
UnIortunately. under the usual circumstances we cannot compute . Because P does not need to
Copyright 2009. Jay Walters 28
contain a view on every asset. several oI the terms are not always computable as written. However. we
can easily convert it to view space by multiplying by P and P
T
.
(53)
<
F'! P./Q"24(P7 %P
T
3
/1
! P./Q"
T
This new test statistic statistic. . in Iormula (53) is distributed as
2
(q) where q is the number oI views.
We can use this test statistic to determine iI our views are consistent with the prior by means oI a
standard conIidence test.
P!q"'1/F !F!q""
Where F() is the CDF oI
2
(q) distribution.
We can also compute the sensitivities oI this measure to the views using the chain rule.
GP
G q
'
GP
G F
GF
Gq
Substituting the various terms
(54)
GP
Gq
'/f !F"2/2!!4(P7% P
T
"
/1
! P ./Q"
T
"3
Where I() is the PDF oI the
2
(q) distribution.
Fusai and Meucci's Measure of Consistency
Next we will look at the work oI Fusai and Meucci (2003). In their paper they present a way to
quantiIy the statistical diIIerence between the posterior return estimates and the prior estimates. This
provides a way to calibrate the uncertainty oI the views and ensure that the posterior estimates are not
extreme when viewed in the context oI the prior equilibrium estimates.
In their paper they use the Alternative ReIerence Model. Their measure is analagous to Theil's
Measure oI Compatibility. but because the alternative reIerence model uses the prior variance oI
returns Ior the posterior they do not need any derivation oI the variance. We can apply a variant oI
their measure to the BlackLitterman ReIerence Model as well.
They propose the use oI the Mahalanobis distance oI the posterior returns Irom the prior returns. I
include t here to match the BlackLitterman reIerence model. but their work does not include t as they
use the Alternative ReIerence Model.
(54) M!q"'!$
BL
/$"!7 %"
/1
!$
BL
/$"
It is essentially measuring the distance Irom the prior. . to the estimated returns.
BL
. normalized by
the uncertainty in the estimate. We use the covariance matrix oI the prior distribution as the
uncertainty. The Mahalanobis distance is distributed as a chisquare distribution with n degrees oI
Ireedom (n is the number oI assets). and thus allows us to use it in a hypothesis test. Thus the
probability oI this event occurring can be computed as:
(55) P!q"'1/F !M !q""
Where F(M(q)) is the CDF oI the chi square distribution oI M(q) with n degrees oI Ireedom.
Finally. in order to identiIy which views contribute most highly to the distance away Irom the
equilibrium. we can also compute sensitivities oI the probability to each view. We use the chain rule to
compute the partial derivatives
Copyright 2009. Jay Walters 29
GP!q"
Gq
'
GP
GM
GM
G$
BL
G$
BL
Gq
(56)
GP!q"
Gq
'/f ! M "2 2!$
BL
/$"32 ! P!7%" P(4"
/1
P3
Where I(M) is the PDF oI the chi square distribution with n degrees oI Ireedom Ior M(q).
They work an example in their paper which results in an initial probability oI 94 that the posterior is
consistent with the prior. They speciIy that their investor desires this probability to be no less than
95 (a commonly used conIidence level in hypothesis testing). and thus they would adiust their views
to bring the probability in line. Given that they also compute sensitivities. their investor can identiIy
which views are providing the largest marginal increase in their measure and they investor can then
adiust these views. These sensitivities are especially useIul since some views may actually be pulling
the posterior towards the prior. and the investor could strengthen these views. or weaken views which
pull the posterior away Irom the prior. This last point may seem nonintuitive. Given that the views
are indirectly coupled by the covariance matrix. one would expect that the views only push the
posterior distribution away Irom the prior. However. because the views can be conIlicting. either
directly or via the correlations. any individual view can have a net impact pushing the posterior closer
to the prior. or pushing it Iurther away.
They propose to use their measure in an iterative method to ensure that the posterior is consistent with
the prior to the speciIied conIidence level.
With the BlackLitterman ReIerence Model we could rewrite Iormula (54) using the posterior variance
oI the return instead oI tL yielding:
(57) M!q"'!$
BL
/$"!!7 %"
/1
(P
T
4
/1
P"
/1
!$
BL
/$"
Otherwise their Consistency Measure and it's use is the same Ior both reIerence models.
He and Litterman's Lambda
He and Litterman (1999) use a measure. . to measure the impact oI each view on the posterior. They
deIine the BlackLitterman unconstrained posterior portIolio as a blend oI the equilibrium portIolio
(prior) and a contribution Irom each view. that contribution is measured by .
Deriving the Iormula Ior we will start with Iormula (9) and substitute in the various values Irom the
posterior distribution.
w'!0%"
/1
<
.
We substitute the return Irom Iormula (25) Ior
<
.
(58)
< w'
1
0
=
%
/1
M
/1
2!7%"
/1
.(P
T
4
/1
Q3
We will Iirst simpliIy the covariance term
Copyright 2009. Jay Walters 30
=
%
/1
M
/1
'! %(M
/1
"
/1
M
/1
=
%
/1
M
/1
'! %M(I "
/1
=
%
/1
M
/1
'%
/1
! %
/1
(M"
/1
=
%
/1
M
/1
'%
/1
! %
/1
(!7%"
/1
(P4
/1
P
T
"
/1
=
%
/1
M
/1
'%
/1
!
1(7
7
%
/1
(P4
/1
P
T
"
/1
=
%
/1
M
/1
'! P
T
7 %P"
/1
2!1(7"! P
T
%P"
/1
(74
/1
3
/1
=
%
/1
M
/1
'! P
T
7 %P"
/1
2
! P
T
% P"
!1(7"
/
! P
T
7 %P"
!1(7"
! P
T
%P"
!1(7"
2
4
7
(
P
T
%P
1(7
3
/1
3
=
%
/1
M
/1
'
7
1(7
2 I /
! P
T
%P"
1(7
!
4
7
(
P
T
%P
1(7
"
/1
3
Then we can deIine
(59) A'2
4
7
(
P
T
%P
1(7
3
And Iinally rewrite as
(60)
=
%
/1
M
/1
'
7
1(7
2 I /! P
T
A
/1
P"!
%
1(7
"3
Our goal is to simpliIy the Iormula to the Iorm
< w'
1
1(7
H
w
eq
(P
T
I
J
In order to Iind we substitute Iormula (60) into (58) and then gather terms.
< w'
1
0
7
1(7
2 I /! P
T
A
/1
P"!
%
1(7
"32 !7 %"
/1
.(P
T
4
/1
Q3
< w'
1
0
7
1(7
2!7 %"
/1
./! P
T
A
/1
P"!
7
1(7
".(P
T
4
/1
Q/! P
T
A
/1
P"!
%
1(7
" P
T
4
/1
Q3
< w'
1
1(7
H
w
eq
(P
T
2
/A
/1
P .!
1
0!1(7"
"(
74
/1
Q
0
/! A
/1
P"!
%
0!1(7"
" P
T
4
/1
Q
3J
< w'
1
1(7
H
w
eq
(P
T
2
7
0
4
/1
Q/
A
/1
P %w
eq
1(7
/A
/1
7
1(7
! P
T
%P"
4
/1
Q
0
3
J
So we can see that along with (59). the Iollowing Iormula deIines .
(61) I'
7
0
4
/1
Q/
A
/1
P %w
eq
1(7
/A
/1 7
1(7
! P
T
%P"
4
/1
Q
0
He and Litterman's represents the weight on each oI the view portIolios on the Iinal posterior weight.
As a result. we can use as a measure oI the impact oI our views.
Copyright 2009. Jay Walters 31
Braga and Natale and Tracking Error Volatility
Braga and Natale (2007) propose the use oI tracking error between the posterior and prior portIolios as
a measure oI distance Irom the prior. Tracking error is commonly used by investors to measure risk
versus a benchmark. and can be used as an investment constraint. As it is so commonly used. most
investors have an intuitive understanding and ga level oI comIort with TEV. Tracking error volatility is
deIined as
(62) TEJ'
K
w
actv
T
%w
actv
wherew
actv
'< w/w
r
Where
w
actv
Active weights. or active portIolio
< w Weight in the investor's portIolio
w
r
Weight in the reIerence portIolio
% Covariance matrix oI returns
They also derive the Iormula Ior tracking error sensitivities as Iollows: Given that
TEJ' f ! w
actv
"
and we can Iurther reIine
w
actv
'g! q" where q representsthe views
Then we can use the chain rule to decompose the sensitivity oI TEV to the views
(63)
GTEJ
Gq
'
GTEJ
Gw
actv
Gw
actv
Gq
We can solve Ior the Iirst term oI Iormula (63) directly.
GTEJ
Gw
actv
'G
!
K
w
actv
T
%w
actv
"
Gw
actv
Let x'w
actv
T
%w
actv
. then applv the chain rule
GTEJ
Gw
actv
'
GTEJ
G x
G x
Gw
actv
GTEJ
Gw
actv
'2
1
2
K
w
actv
T
%w
actv
32 2 %w
actv
3
GTEJ
Gw
actv
'
%w
actv
K
w
actv
T
%w
actv
Solving Ior the second term oI Iormula (63) is slightly more complicated.
Copyright 2009. Jay Walters 32
Gw
actv
Gq
'
G! < w/w
ref
"
Gq
Gw
actv
Gq
'
G!!0%"
/1
E! r"/!L%"
/1
."
Gq
Gw
actv
Gq
'!0%"
/1
G! E! r"/."
Gq
Gw
actv
Gq
'!0%"
/1
G
7 %P
T
2! P 7% P
T
"(43
/1
2Q/P .3
GQ
Gw
actv
Gq
'!0%"
/1
7% P
T
2! P7 %P
T
"(43
/1
Gw
actv
Gq
'
7
0
P
T
2 ! P7 %P
T
"(43
/1
This result is somewhat diIIerent Irom that Iound in the Braga and Natale (2007) paper because we use
the Iorm oI the BlackLitterman model which requires less matrix inversions. The Iormula Ior the
sensitivities is
(64)
GTEJ
Gq
'
%w
actv
K
w
actv
T
%w
actv
7
0
P
T
2 ! P7% P
T
"(43
/1
Braga and Natale work through a Iairly simple example in their paper. I have been unable to reproduce
either their equilibrium or their mixing results. Given their posterior distribution as presented in the
paper one can easily reproduce their TEV results.
One advantage oI the TEV is that most investors are Iamiliar with it. and so they will have some
intuition as to what it represents. The consistency metric introduced by Fusai and Meucci (2003) will
not be as Iamiliar to investors.
Relative Entropy and the KullbackLeibler Information Criteria
For a third measure oI the impact oI the views on the posterior distribution. we can turn to InIormation
Theory. The most common approach in InIormation Theory Ior measuring the distance between two
distributions is to use a relative entropy measure. A widely used relative entropy is the Kullback
Leibler InIormation Criteria (KLIC). It measures the incremental amount oI inIormation required to
represent the posterior given the prior. As such it can be a useIul tool Ior measuring the impact oI the
views in the BlackLitterman model.
The continuous Iorm oI the KLIC is shown below:
(65)
D
PQ
'/
M
x
log!
dQ
dP
"dP'
M
x
log!
dP
dQ
" dP
We can apply it to the BlackLitterman Model where P and Q are the probability density Iunctions
(PDF) Ior the posterior and prior distributions respectively.
From Iormula (65) we can see why the KLIC is not a true distance measure. It is not symmetric. the
measure Irom p to q will in general not be equal to the measure Irom q to p. One approach to making it
symmetric is to Iormulate it
Copyright 2009. Jay Walters 33
D
S
'
1
2
2 D
PQ
(D
QP
3
It is oIten used in it's discrete Iorm. shown below
(66) KLIC! p. q"'
:
i'1
n
p
i
ln!
p
i
q
i
"
p
i
Posterior weight Ior asset (i)
q
i
Prior weight Ior asset (i)
In the discrete Iorm we could use it to measure the distance Irom the prior weights to the posterior
weights. However. there are some drawbacks to using the discrete KLIC to measure the added
inIormation content oI the posterior versus the prior. One drawback is that in order to use the discrete
KLIC we require p
i
> 0 and q
i
~ 0 i. II we are using constrained optimization and a noshort selling
constraint then we can work around this restriction. but in the general case it is not a good metric.
II we use the continuous distributions oI P and Q in Iormula (65). we will not suIIer Irom the problems
which aIIect the discrete KLIC. In order to proceed. we Iirst need to deIine P and Q. Given that both P
and Q are the multivariate normal distribution. their probability density Iunction is:
(67) P . Q'
1
!2&"
NB 2
det !%"
1B2
e
!/
1
2
! x/$"
T
%
/1
! x/$""
Substituting Iormula (67) into (65) Ior both the prior and posterior distribution. and evaluating the
integral we arrive at the Iormula Ior the KLIC between two multivariate normal distributions.
(68)
D
KL
'
1
2 2
log
!
det! %
post
"
det ! %
pri
"
"
(tr ! %
post
/1
%
pri
"(!$
post
/$
pri
"
T
%
post
/1
!$
post
/$
pri
"/N
3
II we set L
post
L
pre
then we can simpliIy Iormula (68) dramatically. The Iirst term goes to 0. the
second and Iourth terms cancel out as the tr !%
/1
%"/N'tr ! I
N
"/N'0 . That leaves only the third
term which is the Mahalanobis Distance. same as Iormula (54). Thus. we can see that the Consistency
Metric oI Fusai and Meucci (2003) is related to the continuous KLIC oI the distributions.
The KLIC is a relative measure. this means that while we can test the impact oI diIIering views on the
posterior distribution Ior a given problem. we cannot easily compare the impact across problems using
the KLIC. Further. because the measure is not absolute like TEV or the Mahalanobis Distance. it will
be harder to develop intuition based on the scale oI the KLIC.
A Demonstration of the Measures
Now we will work a sample problem to illustrate all oI the metrics. and to provide some comparison oI
their Ieatures. We will start with the equilibrium Irom He and Litterman (1999) and Ior Example 1 use
the views Irom their paper. Germany will outperIorm other European markets by 5 and Canada will
outperIorm the US by 4.
Table 8 Example 1 Returns and Weights. equilibrium Irom He and Litterman. (1999).
Copyright 2009. Jay Walters 34
Asset P0 P1 eq weq/(1t) w w  weq/(1t)
Australia 0.0 0.0 4.45 3.9 16.4 1.5 14.9
Canada 0.0 1.0 9.06 6.9 2.1 53.3 51.2
France 0.295 0.0 9.53 8.4 5.0 3.3 8.3
Germany 1.0 0.0 11.3 9 5.2 33.1 27.9
Japan 0.0 0.0 4.65 4.3 11.0 11.0 0.0
UK 0.705 0.0 6.98 6.8 11.8 7.8 19.6
USA 0.0 1.0 7.31 7.6 58.6 7.3 51.3
q 5.0 4.0
e/t 0.02 .017
Table 9  Impact Measures Ior Example 1
Measure Value (ConIidence Level) Sensitivity (V1) Sensitivity (V2)
Theil's Measure 1.67 (0.041) 1.15 2.21
Fusai and Meucci's Measure 0.87 (0.00337) 0.18 0.33
0.292 0.538
TEV 8.20 0.697 1.309
KLIC 1.222
Table (8) illustrates the results oI applying the views and Table (9) displays the various impact
measures. II we examine the change in the estimated returns vs the equilibrium. we see where the USA
returns decreased by 29 bps. but the allocation decreased 51.3 caused by the optimizer Iavoring
Canada whose returns increased by 216 bps and whose allocation increased by 51.2. This shows that
what appear to be moderate changes in return Iorecasts can cause very large swings in the weights oI
the assets. a common issue with mean variance optimization.
Next looking at the impact measures. Theil's measure indicates that we can be conIident at the 5 level
that the views are consistent with the prior. Fusai and Meucci's Mahalanobis Distance is less than 1. so
in return space the new Iorecast return vector is less than one standard deviation Irom the prior
estimates. The consistency measure is 0.30. which means we can be very conIident that the posterior
agrees with the prior. The measure oI Fusai and Meucci is much more conIident that the posterior is
consistent with the prior than Theil's measure is.
He and Litterman's Lambda indicates that the second view has a relative weight ~ which means it is
impacting the posterior more signiIicantly than the Iirst view.
The TEV oI the posterior portIolio is 8.20 which is signiIicant in terms oI how closely the posterior
portIolio will track the equilibrium portIolio. Given the examples Irom their paper this scenario seems
Copyright 2009. Jay Walters 35
to have a very large TEV.
Next we change our conIidence in the views by dividing the variance by 4. this will increase the change
Irom the prior to the posterior and allow us to make some iudgements based on the impact measures.
Table 10 Example 2 Returns and Weights. equilibrium Irom He and Litterman. (1999).
Asset P0 P1 weq/(1t) w w  weq/(1t)
Australia 0.0 0.0 4.72 16.4 1.5 14.9
Canada 0.0 1.0 10.3 2.1 83.9 81.8
France 0.295 0.0 10.2 5.0 7.7 12.7
Germany 1.0 0.0 12.4 5.2 48.1 32.9
Japan 0.0 0.0 4.84 11.0 11.0 0.0
UK 0.705 0.0 7.09 11.8 18.4 30.2
USA 0.0 1.0 7.14 58.6 23.2 81.8
q 5.0 4.0
e/t 0.01 0
Table 11  Impact Measures Ior Example 2
Measure Value (ConIidence Level) Sensitivity (V1) Sensitivity (V2)
Theil's Measure 2.607 (0.079) 3.32 6.66
Fusai and Meucci's Measure 2.121 (0.0470) 2.1 4.06
0.450 0.859
TEV 12.90 1.057 2.069
KLIC 8.090
Examining the updated results in Table (10) we see that the changes to the Iorecast returns have
increased and the changes to the asset allocation have become even more extreme. We now have an
80 increase in the allocation to Canada and an 80 decrease in the allocation to the USA. From
Table (11) we can see that Theil's measure has increased and we are no longer conIident at the 5 level
that the views are consistent with the prior estimates. Fusai and Meucci's measure now is close to the
5 conIidence level that the posterior is consistent with the prior. It is unclear in practice what bound
we would want to use. but 5 is a very common conIidence level to use Ior statistical tests.
Once again we see the He and Litterman's Lambda shows the second view having more oI an impact
on the Iinal weights. In this scenario it is now up to 86.
The TEV has increased. and is now 12.90 which seems very high and likely outside the tolerance.
The KLIC has increased 8x indicating that the posterior is signiIicantly diIIerent Irom the prior.
Copyright 2009. Jay Walters 36
Next we change our conIidence in the views by multiplying the variance by 4. this will decrease the
change Irom the prior to the posterior and allow us to make some iudgements based on the impact
measures.
Table 12 Example 3 Returns and Weights. equilibrium Irom He and Litterman. (1999).
Asset P0 P1 weq/(1t) w w  weq/(1t)
Australia 0.0 0.0 4.15 16.4 1.5 14.9
Canada 0.0 1.0 7.8 2.1 22.7 20.6
France 0.295 0.0 8.85 5.00 1.6 3.5
Germany 1.0 0.0 9.96 5.2 16.8 11.6
Japan 0.0 0.0 4.45 11.0 11.0 0.0
UK 0.705 0.0 6.86 11.8 3.7 8.1
USA 0.0 1.0 7.47 58.6 38.0 20.6
q 5.0 4.0
e/t 0.09 0.07
Table 13  Impact Measures Ior Example 2
Measure Value (ConIidence Level) Sensitivity (V1) Sensitivity (V2)
Theil's Measure 0.687 (0.0073) 0.086 0.159
Fusai and Meucci's Measure 0.147 (0.0000) 0.000547 0.000933
0.120 0.220
TEV 3.40 0.272 0.531
KLIC 0.121
In this scenario we see that Theil's measure now shows us conIident at the 1 level that the views are
consistent with the prior estimates. Fusai and Meucci's Consistency Measure is very low. conIident to
more than 99.999 indicating this is a highly plausible scenario. Lambda Ior the second view is now
less than 25 and about that Ior the Iirst view. This indicates that neither view is having a large
impact on the posterior. The TEV is down to a manageable 3.4 and all the discrete KLIC measures
have decreased as the impact oI the views has lessened. The continuous KLIC measure seems to move
the most dramatically oI the KLIC measures. The continuous KLIC is now actually less than Fusai and
Meucci's Mahalanobis Distance which indicates that the changes in the posterior Covariance matrix are
reducing the Mahalanobis Distance calculation embedded in that metric.
Theil's test ranged Irom a high oI 99.27 conIident to a low oI 92.1 conIident that the views were
consistent with the prior estimates. This indicates that using a threshold oI 9598 Ior a conIidence
level would likely give good results. e.g. we can Iorce the test to Iail.
Fusai and Meucci's Consistency measure ranged Irom 99.99 conIident to 95.3 conIident.
Copyright 2009. Jay Walters 37
indicating the posterior was generally highly consistent with the prior by their measure. Fusai and
Meucci present that an investor may have a requirement that the conIidence level be 5. In light oI
these results that would seem to be a Iairly large value. The sensitivities oI the Consistency measure
scale with the measure. and Ior low values oI the measure the sensitivities are very low.
He and Litterman's Lambda ranged Irom a low oI 22 Ior the second view to a high oI 86. This is
consistent with the impact oI the second view on the weights. where in the Iirst two scenarios the
weights oI the United States and Canada were signiIicantly impacted by the view.
Across the three scenarios the TEV increased Irom 3.4 in the low conIidence case to 12.9 in the
high conIidence case. The latter value Ior the TEV is very large. It is not clear what a realistic
threshold Ior the TEV is in this case. but these values are likely towards the upper limit that would be
tolerated. The sensitivities oI the TEV scale with the TEV. so Ior posteriors with low TEV the
sensitivities are lower as well.
The KLIC ranged over one order oI magnitude Ior the 16x change in the conIidence in the views.
In analyzing these various measures oI the tilt caused by the views. the TEV and discrete KLIC oI the
weights measure the impact oI the views and the optimization process. which we can consider as the
Iinal outputs. II the investor is concerned about limits on TEV. they could be easily added as
constraints on the optimization process.
He and Litterman's Lambda measures the weight oI the view on the posterior weights. but only in the
case oI an unconstrained optimization. This makes it suitable Ior measuring impact and being a part oI
the process. but it cannot be used as a constraint in the optimization process.
Theil's Compatibility measure. Fusai and Meucci's Consistency measure. the KLIC measure the
posterior distribution. including the returns and the covariance matrix. The latter more directly
measures the impact oI the views on the posterior because it includes the impact oI the views on the
covariance matrix.
/A+ND&'$+%2O6&'FNJ.$$1%8&*
Krishnan and Mains (2005) developed an extension to the alternate reIerence model which allows the
incorporation oI additional uncorrelated market Iactors. The main point they make is that the Black
Litterman model measures risk. like all MVO approaches. as the covariance oI the assets. They
advocate Ior a richer measure oI risk. They speciIically Iocus on a recession indicator. given the thesis
that many investors want assets which perIorm well during recessions and thus there is a positive risk
premium associated with holding assets which do poorly during recessions. Their approach is general
and can be applied to one or more additional market Iactors given that the market has zero beta to the
Iactor and the Iactor has a nonzero risk premium.
They start Irom the standard quadratic utility Iunction (6). but add an additional term Ior the new
market Iactor(s).
(69) U'w
T
./!
0
0
2
" w
T
%w/
:
i '1
n
0
i
w
T
*
i
U is the investors utility. this is the obiective Iunction during portIolio optimization.
w is the vector oI weights invested in each asset
H is the vector oI equilibrium excess returns Ior each asset
L is the covariance matrix Ior the assets
Copyright 2009. Jay Walters 38
o
0
is the risk aversion parameter oI the market
o
i
is the risk aversion parameter Ior the ith additional risk Iactor
i
is the vector oI exposures to the ith additional risk Iactor
Given their utility Iunction as shown in Iormula (69) we can take the Iirst derivative with respect to w
in order to solve Ior the equilibrium asset returns.
(70) .'0
0
%w(
:
i'1
n
0
i
*
i
Comparing this to Iormula (7). the simple reverse optimization Iormula. we see that the equilibrium
excess return vector (H) is a linear composition oI (7) and a term linear in the
i
values. This matches
our intuition as we expect assets exposed to this extra Iactor to have additional return above the
equilibrium return.
We will Iurther deIine the Iollowing quantities:
r
m
as the return oI the market portIolio.
I
i
as the time series oI returns Ior the Iactor
r
i
as the return oI the replicating portIolio Ior risk Iactor i.
In order to compute the values oI o we will need to perIorm a little more algebra. Given that the
market has no exposure to the Iactor. then we can Iind a weight vector. v
i
. such that v
i
T
i
0. In order
to Iind v
i
we perIorm a least squares Iit oI >> f
i
/v
i
T
.>> subiect to the above constraint. v
0
will be
the market portIolio. and v
0
i
0 i by construction. We can solve Ior the various values oI o by
multiplying Iormula (70) by v and solving Ior o
0
.
v
0
T
.'0
0
v
0
T
%v
0
(
:
i'1
n
0
i
v
0
T
*
i
By construction v
0
i
0. and v
0
H r
m
. so
0
0
'
r
m
!v
0
T
%v
0
"
For any i > 1 we can multiply Iormula (70) by v
i
and substitute o
0
to get
v
i
T
.'0
0
v
i
T
%v
i
(
:
i'1
n
0
i
v
i
T
*
i
Because these Iactors must all be independent and uncorrelated. then v
i
i
0 i i so we can solve Ior
each o
i
.
0
i
'
!r
i
/0
0
v
i
T
%v
i
"
! v
i
T
*
i
"
The authors raise the point that this is only an approximation because the quantity >> f
i
/v
i
T
.>> may
not be identical to 0. The assertion that v
i
i
0 i i may also not be satisIied Ior all i and i. For the
case oI a single additional Iactor. we can ignore the latter issue.
In order to transIorm these Iormulas so we can directly use the BlackLitterman model. Krishnan and
Mains change variables. letting
Copyright 2009. Jay Walters 39
<
.'./
:
i'1
n
0
i
*
i
Substituting back into (69) we are back to the standard utility Iunction
U'w
T
<
./!
0
0
2
" w
T
%w
and Irom Iormula (11)
P
<
.'P! ./
:
i '1
n
0
i
*
i
"
P
<
.'P ./
:
i'1
n
0
i
P *
i
thus
<
Q'Q/
:
i'1
n
0
i
P *
i
We can directly substitute
<
. and
<
Q into Iormula (30) Ior the posterior returns in the Black
Litterman model in order to compute returns given the additional Iactors. Note that these additional
Iactor(s) do not impact the posterior variance in any way.
Krishnan and Mains work an example oI their model Ior world equity models with an additional
recession Iactor. This Iactor is comprised oI the Altman Distressed Debt index and a short position in
the S&P 500 index to ensure the market has a zero beta to the Iactor. They work through the problem
Ior the case oI 100 certainty in the views. They provide all oI the data needed to reproduce their
results given the set oI Iormulas in this section. In order to perIorm all the regressions. one would need
to have access to the Altman Distressed Debt index along with the other indices used in their paper.
D$%12P.%1'$.+*#
Future directions Ior this research include reproducing the results Irom the original papers. either Black
and Litterman (1991) or Black and Litterman (1992). These results have the additional complication oI
including currency returns and partial hedging.
Later versions oI this document should include more inIormation on process and a synthesized model
containing the best elements Irom the various authors. A Iull example Irom the CAPM equilibrium.
through the views to the Iinal optimized weights would be useIul. and a worked example oI the two
Iactor model Irom Krishnan and Mains (2005) would also be useIul.
Meucci (2006) and Meucci (2008) provide Iurther extensions to the BlackLitterman Model Ior non
normal views and views on parameters other than return. This allows one to apply the BlackLitterman
Model to new areas such as alternative investments or derivatives pricing. His methods are based on
simulation and do not provide a closed Iorm solution. Further analysis oI his extensions will be
provided in a Iuture revision oI this document.
J.$1%&$%12>%C1?
This section will provide a quick overview oI the reIerences to BlackLitterman in the literature.
Copyright 2009. Jay Walters 40
The initial paper. Black and Litterman (1991) provides some discussion oI the model. but does not
include signiIicant details and also does not include all the data necessary to reproduce their results.
They introduce a parameter. weight on views. which is used in a Iew oI the other papers but not clearly
deIined. It appears to be the Iraction P
T
O
1
P((tL)
1
P
T
O
1
P)
1
. This represents the weight oI the view
returns in the mixing. As O 0. then the weight on views 100.
Their second paper on the model. Black and Litterman (1992). provides a good discussion oI the model
along with the main assumptions. The authors present several results and most oI the input data
required to generate the results. however they do not document all their assumptions in any easy to use
Iashion. As a result. it is not trivial to reproduce their results. They provide some oI the key equations
required to implement the BlackLitterman model. but they do not provide any equations Ior the
posterior variance.
He and Litterman (1999) provide a clear and reproducible discussion oI BlackLitterman. There are
still a Iew Iuzzy details in their paper. but along with Idzorek (2005) one can recreate the mechanics oI
the BlackLitterman model. Using the He and Litterman source data. and their assumptions as
documented in their paper one can reproduce their results.
Idzorek (2005) provides his inputs and assumptions allowing his results to be reproduced. During this
process oI reproducing their results. I identiIied the Iact that Idzorek does not handle the posterior
variance the same way as He and Litterman.
Bevan and Winkelmann (1998) and the chapter Irom Litterman's book Litterman. et al. (2003) do not
shed any Iurther light on the details oI the algorithm. Neither provides the details required to build the
model or to reproduce any results they might discuss. Bevan and Winkelmann (1998) provide details
on how they use BlackLitterman as part oI their broader Asset Allocation process at Goldman Sachs.
including some calibrations oI the model which they perIorm. This is useIul inIormation Ior anybody
planning on building BlackLitterman into an ongoing asset allocation process.
Satchell and ScowcroIt (2000) claim to demystiIy BlackLitterman. but they don't provide enough
details to reproduce their results. and they seem to have a very diIIerent view on the parameter t than
the other authors do. I see no intuitive reason to back up their assertion that t should be set to 1. They
provide a detailed derivation oI the BlackLitterman 'master Iormula'..
Christadoulakis (2002). and Da and Jagnannathan (2005) are teaching notes Ior Asset Allocation
classes. Christadoulakis (2002) provides some details on the Bayesian mechanisms. the assumptions oI
the model and enumerates the key Iormulas Ior posterior returns. Da and Jagnannathan (2005)
provides some discussion oI an excel spreadsheet they build and work through a simple example in the
content oI their spreadsheet.
Herold (2003) provides an alternative view oI the problem where he examines optimizing alpha
generation. essentially speciIying that the sample distribution has zero mean.. He provides some
additional measures which can be used to validate that the views are reasonable.
Koch (2005) is a powerpoint presentation on the BlackLitterman model. It includes derivations oI the
'master Iormula' and the alternative Iorm under 100 certainty. He does not mention posterior
variance. or show the alternative Iorm oI the 'master Iormula' under uncertainty (general case).
Krishnan and Mains (2005) provide an extension to the BlackLitterman model Ior an additional Iactor
which is uncorrelated with the market. They call this the TwoFactor BlackLitterman model and they
show an example oI extending BlackLitterman with a recession Iactor. They show how it intuitively
impacts the expected returns computed Irom the model.
Copyright 2009. Jay Walters 41
Mankert (2006) provides a nice solid walk through oI the model and provides a detailed transIormation
between the two speciIications oI the BlackLitterman 'master Iormula' Ior the estimated asset returns.
She also provides some new intuition about the value t. Irom the point oI view oI sampling theory.
Meucci (2006) provides a method to use nonnormal views in BlackLitterman. Meucci (2008) extends
this method to any model parameter. and allow Ior both analysis oI the Iull distribution as well as
scenario analysis.
Braga and Natale (2007) describes a method oI calibrating the uncertainty in the views using Tracking
Error Volatility (TEV). This metric is a well known Ior it's use in benchmark relative portIolio
management.
Several oI the other authors reIer to a reIerence Firoozy and Blamont. Asset Allocation Model. Global
Markets Research. Deutsche Bank. July 2003. I have been unable to Iind a copy oI this document. I
will at times still reIer to this document based on comments by other authors. AIter reading other
authors reIerences to their paper. I believe my approach to the problem is somewhat similar to theirs.
Copyright 2009. Jay Walters 42
3141%1*'1#
Many oI these reIerences are available on the Internet. I have placed a BlackLitterman resources page
on my website. (www.blacklitterman.org) with links to many oI these papers.
Beach and Orlov (2006). An Application oI the BlackLitterman Model with EGARCHMDerived
views Ior International PortIolio Management. Steven Beach and Alexei Orlov. September 2006.
Bevan and Winkelmann (1998) Using the BlackLitterman Global Asset Allocation Model: Three
ears oI Practical Experience. Bevan and Winkelmann. June 1998. Goldman Sachs Fixed Income
Research paper.
Black and Litterman (1991) Global PortIolio Optimization. Fischer Black and Robert Litterman. 1991.
Journal oI Fixed Income. 1991.
Black and Litterman (1992) Global PortIolio Optimization. Fischer Black and Robert Litterman. 1992.
Financial Analysts Journal. Sept/Oct 1992.
Braga and Natale (2007) TEV Sensitivity to Views in BlackLitterman Model. Maria Debora Braga
and Francesco Paolo Natale. 2007.
Christadoulakis (2002) Bayesian Optimal PortIolio Selection: The BlackLitterman Approach.
Christadoulakis. 2002. Class Notes.
Da and Jagnannathan (2005) Teaching Note on BlackLitterman Model. Da and Jagnannathan. 2005.
Teaching notes.
DeGroot (1970) Optimal Statistical Decisions. 1970. Wiley Interscience.
Frost and Savarino (1986) An Empirical Bayes Approach to EIIicient PortIolio Selection. Peter Frost
and James Savarino. Journal oI Financial and Quantitative Analysuis. Vol 21. No 3. September 1986.
Fusai and Meucci (2003) Assessing Views. Risk Magazine. 16. 3. S18S21.
He and Litterman (1999) The Intuition Behind BlackLitterman Model PortIolios. He and Robert
Litterman. 1999. Goldman Sachs Asset Management Working paper.
Herold (2003). PortIolio Construction with Qualitative Forecasts. UlI Herold. 2003. J. oI PortIolio
Management. Fall 2003. p6172.
Herold (2005). Computing Implied Returns in a MeaningIul Way. UlI Herold. 2005. Journal oI Asset
Management. Vol 6. 1. 5364.
Idzorek (2005). A StepByStep guide to the BlackLitterman Model. Incorporating UserSpeciIied
ConIidence Levels. Thomas Idzorek. 2005. Working paper.
Idzorek (2006). Strategic Asset Allocation and Commodities. Thomas Idzorek. 2006. Ibbotson White
Paper.
Koch (2005) Consistent Asset Return Estimates. The BlackLitterman Approach. Cominvest
presentation.
Krishnan and Mains (2005). The TwoFactor BlackLitterman Model. Hari Krishnan and Norman
Mains. 2005. July 2005. Risk Magazine.
Litterman. et al (2003) Beyond Equilibrium. the BlackLitterman Approach. Litterman. 2003. Modern
Copyright 2009. Jay Walters 43
Investment Management: An Equilibrium Approach by Bob Litterman and the Quantitative Research
Group. Goldman Sachs Asset Management. Chapter 7.
Mankert (2006) The BlackLitterman Model Mathematical and Behavioral Finance Approaches
Towards its Use in Practice. Mankert. 2006. Licentiate Thesis.
Meucci (2005). Risk and Asset Allocation. 2005. Springer Finance.
Meucci (2006). Beyond BlackLitterman in Practice: A FiveStep Recipe to Input Views on non
Normal Markets. Attilio Meucci. Working paper.
Meucci (2008). Fully Flexible Views: Theory and Practice. Attilio Meucci. Working paper.
Qian and Gorman (2001). Conditional Distribution in PortIolio Theory. Edward Qian and Stephen
Gorman. Financial Analysts Journal. September. 2001.
Salomons (2007). The Black Litterman Model. Hype or Improvement Thesis.
Satchell and ScowcroIt (2000) A DemystiIication oI the BlackLitterman Model: Managing
Quantitative and Traditional PortIolio Construction. Satchell and ScowcroIt. 2000. J. oI Asset
Management. Vol 1. 2. 138150..
Theil (1971). Principles oI Econometrics. Henri Theil. Wiley.
Copyright 2009. Jay Walters 44
!991*,.M2!
This appendix includes the derivation oI the BlackLitterman master Iormula using Theil's Mixed
Estimation approach which is based on Generalized Least Squares.
/01.6G#25.M1,2<#$.8&$.+*2!99%+&'0
This approach is Irom Theil (1971) and is similar to the reIerence in the original Black and Litterman.
(1992) paper. Koch (2005) also includes a derivation similar to this.
II we start with a prior distribution Ior the returns. Assume a linear model such as
A.1 &'x *(u
Where a is the mean oI the prior return distribution. is the expected return and u is the normally
distributed residual with mean 0 and variance .
Next we consider some additional inIormation. the conditional distribution.
A.2 q'p*(v
Where q is the mean oI the conditional distribution and v is the normally distributed residual with mean
0 and variance .
Both and L are assumed to be nonsingular.
We can combine the prior and conditional inIormation by writing:
A.3
2
&
q
3
'
2
x
p
3
*(
2
u
v
3
Where the expected value oI the residual is 0. and the expected value oI the variance is
E
!2
u
v
3
2 u v 3
3
'
2
; 0
0 4
3
We can then apply the generalized least squares procedure. which leads to estimating as
A.4
<
*'
2
2 x p3
2
; 0
0 4
3
/1
2
x
p
33
/1
2 x p 3
2
; 0
0 4
3
/1
2
&
q
3
This can be rewritten without the matrix notation as
A.5 <
*'2 x ;
/1
x (p4
/1
p 3
/1
2 x ;
/1
&(p 4
/1
q3
We can derive the expression Ior the variance using similar logic. Given that the variance is the
expectation oI !
<
*/*"
2
. then we can start by substituting Iormula A.3 into A.5
A.6
<
*'2 x ;
/1
x (p4
/1
p 3
/1
2 x ;
/1
! x *(u"(p 4
/1
! p *(v"3
This simpliIies to
<
*'2 x ;
/1
x(p4
1
p 3
/1
2 x *;
/1
x(p 4
/1
p *(x ;
/1
u(p4
/1
v3
Copyright 2009. Jay Walters 45
<
*'2 x ;
/1
x (p4
1
p 3
/1
2 x;
/1
x *(p4
/1
p *3(2 x ;
/1
x (p4
1
p 3
/1
2 x;
/1
u(p4
/1
v3
<
*'*(2 x;
/1
x (p4
1
p 3
/1
2 x ;
/1
u(p4
/1
v3
A.7 <
*/*'2 x;
/1
x (p4
1
p 3
/1
2 x ;
/1
u(p4
/1
v3
The variance is the expectation oI Iormula A.7 squared.
E!!
<
*/*"
2
"'!2 x ;
/1
x
T
(p4
1
p
T
3
/1
2 x;
/1
u
T
(p4
/1
v
T
3 "
2
E!!
<
*/*"
2
"'2 x;
/1
x
T
(p4
1
p
T
3
/2
2 x;
/1
u
T
u;
/1
x
T
(p4
/1
v
T
v 4
/1
p
T
(x ;
/1
u
T
v4
/1
p
T
(p4
/1
v
T
u;
/1
x
T
3
We know Irom our assumptions above that E!uu "'; . E!vv "'4 and E!uv "'0 because u
and v are independent variables. so taking the expectations we see the cross terms are 0
E!!
<
*/*"
2
"'2 x;
/1
x
T
(p4
/1
p
T
3
/2
2 ! x;
/1
;;
/1
x
T
"(! p4
/1
44
/1p
T
"(0(03
E!!
<
*/*"
2
"'2 x;
/1
x
T
(p4
/1
p
T
3
/2
2 x ;
/1
x
T
(p4
/1
p
T
3
And we know that Ior the BlackLitterman model. x is the identity matrix and ;'7% so aIter we
make those substitutions we have
A.8
E!!
<
*/*"
2
"'2!7%"
/1
(p4
/1
p
T
3
/1
Copyright 2009. Jay Walters 46
!991*,.M2O
This appendix contains a derivation oI the BlackLitterman master Iormula using the standard Bayesian
approach Ior modeling the posterior oI two normal distributions. One additional derivation is in
Mankert. 2006 where she derives the BlackLitterman 'master Iormula' Irom Sampling theory. and
also shows the detailed transIormation between the two Iorms oI this Iormula.
The PDF Based Approach
The PDF Based Approach Iollows a Bayesian approach to computing the PDF oI the posterior
distribution. when the prior and conditional distributions are both normal distributions. This section is
based on the prooI shown in DeGroot. 1970. This is similar to the approach taken in Satchell and
ScowcroIt. 2000.
The method oI this prooI is to examine all the terms in the PDF oI each distribution which depend on
E(r). neglecting the other terms as they have no dependence on E(r) and thus are constant with respect
to E(r).
Starting with our prior distribution. we derive an expression proportional to the value oI the PDF.
P(A) N(x.S/n) with n samples Irom the population.
So (x) the PDF oI P(A) satisIies
B.1
F! x"Nexp!! S B n"
/1
! E!r"/x"
2
"
Next. we consider the PDF Ior the conditional distribution.
P(B'A) N(.L)
So ('x) the PDF oI P(B'A) satisIies
B.2
F!$>x"Nexp! %
/1
! E!r"/$"
2
"
Substituting B.1 and B.2 into Iormula (1) Irom the text. we have an expression which the PDF oI the
posterior distribution will satisIy.
B.3
F! x>$"Nexp!/! %
/1
! E!r"/$"
2
(! S B n"
/1
! E!r"/x"
2
""
.
or F! x>$"Nexp!/;"
Considering only the quantity in the exponent and simpliIying
;'! %
/1
! E! r"/$"
2
(! SB n"
/1
! E! r"/x"
2
"
;'! %
/1
! E !r"
2
/2E !r "$($
2
"(! S B n"
/1
! E! r"
2
/2 E! r" x(x
2
""
;'E!r"
2
!%
/1
(!S B n"
/1
"/2 E!r"!$%
/1
(x! S B n"
/1
"(%
/1
$
2
(!S B n"
/1
x
2
II we introduce a new term y. where
Copyright 2009. Jay Walters 47
B.4
v'
!$%
/1
(x!S B n"
/1
"
!%
/1
(!S B n"
/1
"
and then substitute in the second term
;'E!r"
2
!%
/1
(!S B n"
/1
"/2 E!r" v!%
/1
(!S B n"
/1
"(%
/1
$
2
(!S B n"
/1
x
2
Then add 0'v
2
!%
/1
(!S B n"
/1
"/!$%
/1
(x! SB n"
/1
"
2
!%
/1
(!S B n"
/1
"
/1
;'E!r"
2
!%
/1
(!S B n"
/1
"/2 E!r" v!%
/1
(!S B n"
/1
"(%
/1
$
2
(!S B n"
/1
x
2
(v
2
!%
/1
(!S B n"
/1
"/!$%
/1
(x! SB n"
/1
"
2
!%
/1
(!S B n"
/1
"
/1
;'E!r"
2
!%
/1
(!S B n"
/1
"/2 E!r" v!%
/1
(!S B n"
/1
"(v
2
!%
/1
(!S B n"
/1
"
(%
/1
$
2
(! S B n"
/1
x
2
/!$%
/1
(x! SB n"
/1
"
2
!%
/1
(!S B n"
/1
"
/1
;'!%
/1
(!S B n"
/1
" 2 E!r"
2
/2 E!r" v(v
2
3(! %
/1
$
2
(!S B n"
/1
x
2
"
/!$%
/1
(x!S B n"
/1
"
2
! %
/1
(! SB n"
/1
"
/1
;'!%
/1
(!S B n"
/1
" 2 E!r"
2
/2 E!r" v(v
2
3/!$%
/1
(x! SB n"
/1
"
2
!%
/1
(!S B n"
/1
"
/1
(! %
/1
$
2
(!S B n"
/1
x
2
"! %
/1
(! SB n"
/1
"!%
/1
(!S B n"
/1
"
/1
;'! %
/1
(! S B n"
/1
" 2 E !r"
2
/2E !r" v(v
2
3
/!$
2
%
/2
(2$ x %
/1
! S B n"
/1
(x
2
! SB n"
/2
"!%
/1
(!S B n"
/1
"
/1
(! %
/2
$
2
(!S B n"
/1
%
/1
x
2
($
2
%
/1
! SB n"
/1
(x
2
!S B n"
/2
"!%
/1
(!S B n"
/1
"
/1
;'! %
/1
(! S B n"
/1
" 2 E !r"
2
/2E !r" v(v
2
3
(!!S B n"
/1
%
/1
x
2
/2$x %
/1
!S B n"
/1
($
2
%
/1
!S B n"
/1
"! %
/1
(! SB n"
/1
"
/1
;'!%
/1
(!S B n"
/1
" 2 E!r"
2
/2 E!r" v(v
2
3
(! %
/1
(! S B n"
/1
"
/1
! x/$"! %
/1
!S B n"
/1
"
The second term has no dependency on E(r). thus it can be included in the proportionalality Iactor and
we are leIt we
B.5
F! x>$"Nexp!/2 !%
/1
(!S B n"
/1
"
/1
! E ! R"/v"
2
3"
Thus the posterior mean is y as deIined in Iormula A.12. and the variance is
B.6
!%
/1
(!S B n"
/1
"
/1
Copyright 2009. Jay Walters 48
!991*,.M27
This appendix provides a derivation oI the alternate Iormat oI the posterior variance. This Iormat does
not require the inversion oI O. and thus is more stable computationally.
((tL)
1
P
T
O
1
P)
1
((tL)
1
P
T
O
1
P)
1
((tL)
1
P
T
O
1
P)
1
P
T
(P
T
)
1
((P
T
)
1
(tL)
1
(P
T
)
1
P
T
O
1
P)
1
(P
T
)
1
((tLP
T
)
1
O
1
P)
1
(P
T
)
1
((tLP
T
)
1
O
1
P)
1
(P
T
)
1
(((tLP
T
)
1
O
1
P)
1
(tL)(tL)
1
(P
T
)
1
((tLP
T
)
1
O
1
P)
1
(tL)(P
T
tL)
1
((tLP
T
)
1
O
1
P)
1
(tL)(P
T
tL)
1
(tL)(P
T
tL)
1
((tLP
T
)
1
O
1
P)((tL)
1
P
T
O
1
P)
1
(tL)(P
T
tL)
1
 (tLP
T
)
1
((tL)
1
P
T
O
1
P)
1
(O
1
P)((tL)
1
P
T
O
1
P)
1
(tL)(P
T
tL)
1
 (tLP
T
)
1
((tL)
1
P
T
O
1
P)
1
(O
1
P)((tL)
1
P
T
O
1
P)
1
(O
1
P)(P
1
O)(P
1
O)
1
((tL)
1
P
T
O
1
P)
1
(O
1
P)(P
1
O)((tL)
1
P
1
O P
T
O
1
PP
1
O )
1
(O
1
P)(P
1
O)((tL)
1
P
1
O P
T
)
1
(O
1
P)(P
1
O)((tL)
1
P
1
O P
T
)
1
(PtL)
1
(PtL)
(O
1
P)(P
1
O)((PtL)

(tL)
1
P
1
O (PtL)

P
T
)
11
(PtL)
(O
1
P)(P
1
O)(O PtLP
T
)
11
(PtL)
(O
1
P)(O
1
P)
1
(P(tL)P
T
O)
1
(PtL)
(P(tL)P
T
O)
1
P(tL)
(tL)(P
T
tL)
1
 (P(tL)P
T
O)
1
P(tL) (tLP
T
)
1
((tL)
1
P
T
O
1
P)
1
(tLP
T
)(tL)(P
T
tL)
1
 (tLP
T
)(P(tL)P
T
O)
1
P(tL) ((tL)
1
P
T
O
1
P)
1
(tL)(P
T
tL)(P
T
tL)
1
 (tLP
T
)(P(tL)P
T
O)
1
P(tL) ((tL)
1
P
T
O
1
P)
1
(tL)
 (tLP
T
)(P(tL)P
T
O)
1
(PtL) ((tL)
1
P
T
O
1
P)
1
Copyright 2009. Jay Walters 49
!991*,.M2P
This appendix presents a derivation oI the alternate Iormulation oI the BlackLitterman master Iormula
Ior the posterior expected return. Starting Irom Iormula (29) we will derive Iormula (30).
E!r "'2 !7%"
/1
(P
T
4
/1
P3
/1
2 !7%"
/1
.(P
T
4
/1
Q3
Separate the parts oI the second term
E!r "'2 2!7%"
/1
(P
T
4
/1
P3
/1
!7 %"
/1
.3(22 !7%"
/1
(P
T
4
/1
P3
/1
! P
T
4
/1
Q"3
Replace the precision term in the Iirst term with the alternate Iorm
E!r "'227 %/7% P
T
2 P 7% P
T
(43
/1
P 7%3!7 %"
/1
.3(22 !7%"
/1
(P
T
4
/1
P3
/1
! P
T
4
/1
Q"3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(22 !7%"
/1
(P
T
4
/1
P3
/1
! P
T
4
/1
Q"3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(2!7 %"!7%"
/1
2!7 %"
/1
(P
T
4
/1
P3
/1
! P
T
4
/1
Q"3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(
2
!7 %"
2
I
n
(P
T
4
/1
P7 %
3
/1
! P
T
4
/1
Q"
3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(
2
7%
2
I
n
(P
T
4
/1
P7 %
3
/1
! P
T
4
/1
"Q
3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(
2
7%
2
I
n
(P
T
4
/1
P7 %
3
/1
!4! P
T
"
/1
"
/1
Q
3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(2 7%2 4! P
T
"
/1
(P 7%3
/1
Q3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(2 7% P
T
! P
T
"
/1
2 4! P
T
"
/1
(P 7%3
/1
Q3
E!r "'2./27 %P
T
2 P7 %P
T
(43
/1
P.33(2 7% P
T
2 4(P 7% P
T
3
/1
Q3
Voila. the alternate Iorm oI the BlackLitterman Iormula Ior expected return.
E!r "'./2 7% P
T
2 P 7% P
T
(43
/1
32 Q/P .3
Copyright 2009. Jay Walters 50
!991*,.M2<
This section oI the document summarizes the steps required to implement the BlackLitterman model.
ou can use this road map to implement either the He and Litterman version oI the model. or the
Idzorek version oI the model. The Idzorek version oI the BlackLitterman model leaves out two steps.
Given the Iollowing inputs
w Equilibrium weights Ior each asset class. Derived Irom capitalization weighted CAPM Market
portIolio.
L Matrix oI covariances between the asset classes. Can be computed Irom historical data.
r
I
Risk Iree rate Ior base currency
o The risk aversion coeIIicient oI the market portIolio. This can be assumed. or can be computed
iI one knows the return and standard deviation oI the market portIolio.
t A measure oI uncertainty oI the equilibrium variance. Usually set to a small number oI the
order oI 0.025 0.050.
First we use reverse optimization to compute the vector oI equilibrium returns. H using Iormula (7).
(7) .'0%w
Then we Iormulate the investors views. and speciIy P. O
.
and Q. Given k views and n assets. then P is
a k n matrix where each row sums to 0 (relative view) or 1 (absolute view). Q is a k 1 vector oI the
excess returns Ior each view. O is a diagonal k k matrix oI the variance oI the views. or the
conIidence in the views. As a starting point. most authors call Ior the values oI e
i
to be set equal to
p
T
tL
i
p (where p is the row Irom P Ior the speciIic view).
Next assuming we are uncertain in all the views. we apply the BlackLitterman 'master Iormula' to
compute the posterior estimate oI the returns using Iormula (30).
(30)
<
.'.(7 %P
T
2! P 7% P
T
"(43
/1
2 Q/P.3
This Iollowing two steps are not needed when using the alternative reIerence model. In the alternative
reIerence model
%
p
'%
.
We compute the posterior variance using Iormula (35).
(35)
M'7%/7% P
T
2 P %P
T
(43
/1
P7 %
Closely Iollowed by the computation oI the sample variance Irom Iormula (32).
(32)
%
p
'%(M
And now we can compute the portIolio weights Ior the optimal portIolio on the unconstrained eIIicient
Irontier Irom Iormula (9).
(9) < w'
<
.!0%
p
"
/1
Copyright 2009. Jay Walters 51