Professional Documents
Culture Documents
Omer CAYIRLI2
1
This working paper was written as a part of author’s studies in PhD in Finance at University of South Florida
Business School.
2
Former Assistant Specialist at Central Bank of Republic of Turkey, Markets Department, Foreign Exchange
Transactions Division, University of North Carolina-Chapel Hill Kenan-Flagler MBA 2007
The Black-Litterman Model (BLM), created by Fischer Black and Robert Litterman, is a
sophisticated portfolio construction method that overcomes the problem of unintuitive, highly-
concentrated portfolios, input-sensitivity, and estimation error maximization. The BLM uses a
Bayesian approach to combine the subjective views of an investor regarding the expected
returns of one or more assets with the market equilibrium vector of expected returns (the prior
distribution) to form a new, mixed estimate of expected returns. The resulting new vector of
returns (the posterior distribution), leads to intuitive portfolios with sensible portfolio weights.
(Idzorek, 2004).
BLM was first introduced in Black and Litterman (1990) and further explained in Black
and Litterman (1991) and Black and Litterman (1992) It is an asset allocation model which has
its roots in mean-variance (MV) optimization model and capital asset pricing model (CAPM).
Model builds on MV optimization and CAPM by using a Bayesian framework that allows
investors to incorporate their views on markets effectively into asset allocation process.
without using unnecessary constraints which also reflect their views on markets. It is well known
to both academics and practitioners that standard MV optimization is very sensitive to expected
returns and often generates extreme portfolios (concentration in very few assets, large long and
short positions). BLM overcomes these issues by choosing a neutral reference point, CAPM
equilibrium. It also allows investors to express their views with varying confidence levels and
As a result, since its introduction BLM has been increasingly used among practitioners
and allowed asset allocation decisions to be in more of a quantitative nature rather than
Foundations of BLM
Black and Litterman (1991) explains the intuition behind BLM as follows:
• There are two distinct sources of information about future excess returns:
• Both sources of information are uncertain and are best expressed as probability
distributions.
• Hence, choose expected excess returns that are as consistent as possible with
In standard MV optimization, the user inputs a complete set of expected returns and the
portfolio optimizer generates the optimal portfolio weights. Because there is a complex mapping
between expected returns and the portfolio weights, and because there is no natural starting
point for the expected return assumptions, users of the standard portfolio optimizers often find
their specification of expected returns produces output portfolio weights which do not seem to
make sense. In the BLM the user inputs any number of views, which are statements about the
expected returns of arbitrary portfolios, and the model combines the views with equilibrium,
producing both the set of expected returns of assets as well as the optimal portfolio weights (He
BLM is an equilibrium approach to asset allocation since its starting point is market
portfolio and CAPM equilibrium. As Black and Litterman (1991) puts forth, a neutral reference is
a critically important input in making use of MV optimization model, and equilibrium provides the
appropriate neutral reference. Purpose of the equilibrium is to give investor an appropriate point
when expected returns move away from their equilibrium values, imbalances in markets tend to
Model3
Assume there are N assets in the market and returns of these assets (r) are normally
distributed with expected return (μ) and a known constant covariance matrix (Σ), r ∼ N(μ, Σ)
Assuming the average risk tolerance of the world is represented by the risk aversion
Π = δΣweq , (1)
The Bayesian prior is that the expected returns (μ) are normally distributed with the
mean of Π,
μ = Π+ϵ(e) , (2)
where ϵ(e) is a normally distributed random vector with zero mean and covariance matrix τΣ,
ϵ(e) ∼ N(0, τΣ) . (τ is a scalar indicating the uncertainty of CAPM and we provide more
information on later sections) As a result, BLM expected excess returns are random variables
These results are obtained by solving the following unconstrained maximization problem,
𝛿
max 𝑤 ′ 𝛱 − 2
𝑤 ′Σ w (3)
𝑤
In addition to CAPM prior, the investor also has a number of views on the market
returns. A view is expressed as a statement that the expected return of a portfolio p has a
normal distribution with mean equal to q and a standard deviation given by ω, p(r)~N(q,𝜔2 ). Let
3
Model presented in this section is mostly based on the one presented in (He and Litterman, 1999)
where ϵ(v) is an unobservable normally distributed random vector with zero mean and a
given by
̅ −1 is given by
and the covariance matrix 𝑀
̅ −1
𝛴̅ = 𝛴 + 𝑀 (9)
Usual definition of δ is market risk premium divided by market variance as pointed out in
δ characterizes the expected risk-return tradeoff. It is the rate at which an investor will
forego expected return for less variance. In the reverse optimization process, the risk aversion
coefficient acts as a scaling factor for the reverse optimization estimate of excess returns; the
weighted reverse optimized excess returns equal the specified market risk premium (Idzorek,
̂ ∗ can be obtained by
For an investor with a different risk tolerance, the optimal portfolio 𝑤
̂ ∗ = (𝛿/𝛿̂ )𝑤 ∗
𝑤 (10)
where 𝛿̂ is the risk aversion parameter for the investor (He and Litterman, 1999).
In BLM, views represent the subjective feelings of the investor about relative values
offered in different markets. An investor is not required to state a view about a given market
unless he has a view. In addition, if some of views are more strongly held than others investor is
able to express the differences (Black and Litterman, 1992). This is one of the major
assets or asset classes included in the investment universe, which is a daunting task. Moreover
in MV there is no efficient way of incorporating relative views whereas BLM efficiently generates
posterior expected returns based prior returns, views, view uncertainty and covariance matrix.
BLM not only updates the expected returns of assets that investor has a view but all assets
BLM allows investors to use absolute views as well as relative views. In Exhibit 2 first
view portfolio (p1) represents an absolute view on Asset 2. Second view portfolio (p2)
represents a relative view in which Asset 4 overperforms Asset 1. Last view portfolio (p3)
represents a relative view in which a market cap-weighted portfolio of Asset 2 and Asset 3
outperforms Asset 5.
basically the expected excess return of Asset 1. q2 represents the expected return on p2, and
how it differs from the equilibrium. To state more clearly, for an absolute view as represented by
in p1 above, if the expected return is greater than the equilibrium return than the asset under
consideration will have higher allocation and vice versa. For a relative view, ending portfolio
weights will depend on the difference between view expected returns and equilibrium expected
returns. If views point a higher relative expected return compared with relative equilibrium
expected returns, resulting portfolio will suggest holding higher shares of assets which are
terms, ε(v), from expressed views. In Black and Litterman (1992) and He and Litterman (1999) it
is assumed that Ω is diagonal. A diagonal Ω corresponds to the assumption that the views
represent independent draws from future distribution of returns, or that the deviations of
expected returns from the means of the distribution representing each view are independent
Diagonal elements of Ω are ω1, ω2, …, ωk where ωk represents the confidence level of
kth view. Normally diagonal elements of Ω are different than zero. However, as mentioned in
Walters (2009) they can be zero if investor is 100% confident in his view.
𝜔1 0 0
Ω= [0 ⋱ 0]
0 0 𝜔𝑛
He and Litterman (1999) sets ωk/ 𝜏 to be equal to individual variances of view portfolios
(𝑝1 𝛴𝑝1′ ) ∗ 𝜏 0 0
Ω= [ 0 ⋱ 0 ]
0 0 (𝑝𝑛 𝛴𝑝𝑛′ ) ∗ 𝜏
𝜏 is a measure of the investor’s confidence in the prior estimates (Walters 2010). Black
and Litterman (1992) argues that since the uncertainty in the mean is much smaller than the
uncertainty return itself, τ will be close to zero. He and Litterman (1999) assumes 𝜏 to be 0.05,
corresponding to the confidence level of the CAPM prior mean if it was estimated using 20
years of data. Satchell and Scowcroft (2000) sets 𝜏 to 1 with assumption that the equilibrium
excess returns conditional upon the individual’s forecasts equals the individual forecast on
average. Meucci (2010) argues that in practice, a tailor-made calibration that spans the interval
Walters (2010) considers three methods to select a value for 𝜏: Estimating 𝜏 from the
standard error of the equilibrium covariance matrix, using confidence intervals, and examining
BLM APPLICATION
includes monthly hedged USD returns of 10 broad Bank of America Merrill Lynch bond indices
monthly excess returns which is basically the monthly return minus risk-free rate. Risk-free rate
is obtained from Kenneth R. French’s data library. Market capitalizations are obtained from
SIFMA (US), US Treasury, Office for National Statistics (UK), European Central Bank, and
Ministry of Finance (Japan). Market capitalizations are converted to USD using representative
Exhibit 5 displays the optimal risky portfolios based on standard MV optimization and
BLM without views. It is apparent that MV optimization produces portfolios that are not plausible
in practice. Without short-sale constraint MV results imply a huge long position in EGB0 and
short positions in US Treasury Bill and US Agencies. Adding a short sale constraint does not
help much as we now have a huge long position in EGB0 accompanied with small positions in
US mortgages and Euro corporates. In both cases, deviations from the market portfolio are
significant.
backed securities by 10 basis points and US Treasury notes and bonds will outperform EMU
direct government bonds by 10 basis points. These view portfolios can be represented as
follows:
1
Given these views and setting 𝜏 = 12 we calculate the BLM combined view returns
(posterior returns) and run standard MV optimizations to obtain the optimal risky portfolio 1
(without short-sale constraint) and optimal risky portfolio 2 (with short-sale constraint). Results
are presented in Exhibit 7. First observation is that for unconstrained optimization, only weights
of assets that investor has view(s) change. Another observation is that even if the investor’s
view is that C0A0 outperforming M0A0 by 10 basis points, in the resulting portfolio M0A0 has
increased weight and C0A0’s weight decreases. Optimal portfolio tilts towards M0A0 because
investor’s view implies that excess return spread between M0A0 and C0A0 will decrease from
23 basis points to 17 basis points given the covariance structure and implied equilibrium returns.
We should also note that excess return spread is still higher than the view. This is because of
the uncertainty in views. More uncertainty in views will cause combined view vector to be closer
Once an investor has established his objectives, an asset allocation model establishes a
correspondence between views and optimal portfolios. It is often useful to start an analysis by
using a model to find the implied investor views for which an existing portfolio is optimal. The
weights relative to benchmark or market portfolio define the directions of the investor’s views.
By assuming the investor’s degree of risk aversion one can find the excess returns for which the
10
Implied views are particularly important in terms of strategic asset allocation. In strategic
asset allocation process it is not unusual for some asset classes to be left out of the portfolio or
to be assigned less weights compared to market weights despite being in investment universe.
BLM implied returns provide a solid method for assesing optimality of these decisions. It is
interesting that neither academicians nor practicians paid little attention to this aspect of BLM in
Equations (8) and (9) determine the posterior covariance matrix when investor has
view(s).
̅ −1
𝛴̅ = 𝛴 + 𝑀 (9)
Solving equation (7) for implied views (Q) using equation (8), yields to equation (11):
̅ 𝜇̅ − (𝜏𝛴)−1 𝛱]
𝑄 = [(𝑃′ 𝛺 −1 )′ (𝑃′ 𝛺 −1 )]−1 (𝑃′ 𝛺 −1 )′ [𝑀 (11)
11
implied views of an investor using known prior covariance matrix, posterior covariance matrix, 𝜏,
Assume an investor decides to allocate his neutral portfolio assets in M0A0 and G0P0 to
G0Q0 without stating any views, just the desired portfolio allocations. Still we can form a view
weighted portfolio of G0P0 and M0A0 and long position in G0Q0. Combined view returns based
on this information are presented in the following exhibit. Implied view for our investor is
0.239%, which means that he expects G0Q0 to outperform market-cap weighted portfolio of
G0P0 and M0A0 by 24 basis points. Only after checking this implied view and assesing its
12
In BLM, usual approach to confidence level of the views is using the variance of the view
portfolio and scaling it by using 𝜏. Idzorek (2004) argues that there may be other sources of
information in addition to the variance of the view portfolio that affect an investor’s confidence in
a view and these factors should be combined with the variance of the view portfolio to produce
Idzorek (2004) starts with finding the implied confidence levels in views. His approach in
finding implied confidence levels is solving for BLM combined view vector for 100% confidence
level for a view portfolio as well as variance of the view portfolio. He then calculates the
deviations from market portfolio based on these two approaches and defines the implied
̂ −𝑤𝑒𝑞
𝑤
𝐼𝑚𝑝𝑙𝑖𝑒𝑑 𝐶𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒 𝐿𝑒𝑣𝑒𝑙 = 𝑤̂
100% −𝑤𝑒𝑞
̂ is weight vector based on the covariance matrix of the error term (Ω), and 𝑤
where 𝑤 ̂100%is
He, then, proposes the diagonal elements of Ω be derived in a manner that is based on
the user-specified confidence levels and that results in portfolio tilts of,
where 𝑇𝑖𝑙𝑡𝑘 is the approximate tilt caused by the kth view, and 𝐶𝑘 is the confidence in the kth
view. In this setting, in the absence of other views, the approximate recommended weight vector
is:
where 𝑤𝑘,% is the target weight vector based on the tilt caused by the kth view.
Idzorek (2004)’s procedure for determining Ω and combined view vector based on user-
specified confidence levels is as follows: Calculate combined view vector for each view under
13
maximum departures from the market capitalization weights caused by 100% confidence in the
kth view by using 𝐷𝑘,100% = (𝑤𝑘,100% − 𝑤𝑒𝑞 ) where 𝐷𝑘,100% is the departure from the market
capitalization weight based on 100% confidence in kth view, multiply the N elements of 𝐷𝑘,100%
by the user- specified confidence in the kth view to estimate the desired tilt caused by the k th
view (𝑇𝑖𝑙𝑡𝑘 = 𝐷𝑘,100% ∗ 𝐶𝑘 ), estimate the target weight vector based on the tilt (𝑤𝑘,% = 𝑤𝑒𝑞 +
𝑇𝑖𝑙𝑡𝑘 ), find the value of ωk, that minimizes the sum of the squared differences between wk,% and
wk subject to the constraint ωk>0, and finally repeat all previous steps for K views to build a KxK
Representing the uncertainty of views and specifying the diagonal elements of Ω in BLM
is not straightforward and not intuitive as the model itself. However, Idzorek (2004)’s method
provides a tool that can be used to estimate Ω from the confidence levels provided by investors
Portfolio construction method presented in (Herold, 2003) has two basic aspects: A
Bayesian model similar to BLM which is used for refining alphas, and diagnostic tools for
analyzing the views and a given portfolio. In this approach views are not about a single asset or
asset class since model is aimed at active risk-return space. Instead of taking the market
portfolio as the starting point, neutral reference is benchmark portfolio and incorporates
qualitative views of investors in a framework that maximizes the following alpha/tracking error
objective function:
𝜆𝐴 2 ′ 𝜆𝐴 ′
max (𝛼𝑝 − 𝜓𝑃 ) = max (ℎ𝑃𝐴 𝛼𝐵𝑎𝑦𝑒𝑠 − ℎ 𝛴ℎ𝑃𝐴 )
ℎ𝑃𝐴 2 ℎ𝑃𝐴 2 𝑃𝐴
where hPA denotes vector of active portfolio weights, λA is the investor’s aversion to active risk,
14
α ~ N(0, 𝜏𝛴). Managers will diverge from the benchmark portfolio if they expect alphas to be
different from zero and the levels of conviction a manager has in those views determine the
extent of this deviation. In this context views are summarized by Pα ~ N(q, 𝛺).Views are stated
as a long/short portfolio as it is the case for relative BLM views and tracking error variances of
The prior distributions of the alphas and the views are combined to obtain the posterior
(Herold, 2003) uses view correlation matrix, view risk contributions, and view implied
View correlation matrix serves the purpose that views are consistent with each other
according to the risk model. View correlations are extracted from the covariance matrix of the
views.
View risk contributions are used to analyze how much each view contributes to total
tracking error for a given portfolio. Marginal contributions to tracking error (MCTE) on a view
level obtained by differentiating the following equation that defines the tracking error with
respect to hPA,
′
𝜓𝑃 = ( ℎ𝑃𝐴 𝛴ℎ𝑃𝐴 )0.5
Implied alphas are obtained by reverse optimization and they are proportional to the
MCTEs. The implied alpha for asset i (i=1, …, N) and implied alpha for view j (j=1, …, K) are
given by,
𝑖𝑚𝑝𝑙𝑖𝑒𝑑 𝑖𝑚𝑝𝑙𝑖𝑒𝑑
𝛼𝑖 = 𝜆𝐴 𝑀𝐶𝑇𝐸𝑖 𝜓𝑃 and 𝛼𝑗 = 𝜆𝐴 𝑀𝐶𝑇𝐸𝑗 𝜓𝑃 .
Implied alphas, along with the risk contributions, allow a judgment as to whether the risk
profile of the portfolio is accordance with the conviction that managers have in their views.
15
By determining the shrinkage in information ratio, it is possible to assess how far the
In BLM and the underlying CAPM it is assumed that risk can be completely
characterized by covariance. Krishnan and Norman (2005) focuses on this assumption and
argues that the expected return of an asset class should increase with its volatility and its beta
to alternative risk factors based on the vast research on multifactor models. Following Cochrane
(1999), Krishnan and Norman (2005) defines alternative risk factors as recession risk and
incorporates the recession risk into the BLM using the performance of distressed bonds as a
proxy. The BLM assumes that investors make allocation decision relative to the utility function
𝛿
𝑈(𝑤, 𝑟) = 𝑤 ′ 𝑟 − ( ) 𝑤 ′ 𝛴𝑤 which is the same as the quadratic utility function used in standard
2
MV optimization.
Krishnan and Norman (2005) assumes that views are specified with 100% confidence.
Hence, instead of the usual BLM combined return vector formula they use the following:
This is similar to the formula Idzorek (2004) uses when computing the implied
To incorporate the recession risk into the BLM, Krishnan and Norman (2005) prefers to
16
where β is a vector measuring the sensitivity of each asset class to a recession factor. In this
𝛱∗ = 𝛿𝛴𝑤 + 𝛾𝑗 𝛽 𝑗
where 𝛽 𝑗 is a vector containing the beta of each asset class to the factor 𝑓 𝑗 . 𝛾𝑗 is a risk aversion
parameter that is defined as follows assuming that over the long term an investor collects 𝑟𝑗 for
𝑟𝑗 −𝛾𝑤𝑗′ 𝛴𝑤
𝛾𝑗 ≈ 𝑤𝑗′ 𝛽 𝑗
,
where 𝑤𝑗 are simply the coefficients of a linear regression and define the closest combination of
Using a market-neutral version of the NYU Altman Defaulted Public Bond index as a
proxy for recession risk, Krishnan and Norman (2005) finds that the two-factor BLM allocates
more conservatively to asset classes that have a positive alternative beta, as expected.
Giacometti et al. (2007) has its motivation from the fact that the computation of both the
equilibrium portfolio and the posterior distribution rely on the assumption of a normal market and
investigate whether BLM can be enhanced by using stable Paretian distributions as a statistical
tool for asset returns. They generalize the procedure of the BLM allowing introduction of
dispersion matrices obtained from multivariate Gaussian distribution, symmetric t-Student and
Giacometti et al. (2007) considers different problems of optimal allocation among risky
assets using variance, value at risk (VaR), and conditional value at risk (CVaR) as risk
17
assuming that all portfolios (𝑟 ′ 𝑤) are uniquely determined by the neutral mean 𝑤 ′ 𝛱 and the risk
measure ρ( ) that is defined alternatively as the dispersion 𝑤 ′ 𝑉𝑤, the Varφ (𝑟 ′ 𝑤), and the
CVarφ (𝑟 ′ 𝑤) where V is a dispersion matrix and 𝜑 is the percentile of probability density function.
Applying the reverse optimization model, and using the three different measures of risk
along with Gaussian, t-student and stable distributions, Giacometti et al. (2007) obtain the
Variance: 𝛱 = 𝛿𝑉𝑤
𝛿 𝑉𝑤
CVar : 𝛱 = (𝐶𝑉𝑎𝑅𝜑 − 𝐸(𝑟))
2 √𝑤 ′ 𝑉𝑤
𝛿 𝑉𝑤
VaR : 𝛱 = (𝑉𝑎𝑅𝜑 − 𝐸(𝑟))
2 √𝑤 ′ 𝑉𝑤
Giacometti et al. (2007) provides a tool for using realistic models for asset returns and
enables investors to avoid the problems with the underlying normality assumption. In addition to
introducing α-stable distributions to BLM tool set; it also derives equilibrium excess returns for
dispersion-based risk measures, VaR and CVaR. However, we should note that although VaR
is used widely in practice it is not a coherent risk measure like CVaR since VaR doesn’t have
the subadditivity property [ρ(X + Y) ≤ ρ (X) + ρ(Y )], hence not convex (Artzner et al., 1999).
CONCLUSION
decisions. However, inherent problems with MV optimization hindered the efforts to make
quantitatively oriented asset allocation decisions. In practice, asset allocation decisions either
completely ignored MV optimization context or used it with very specific constraints. One of the
18
BLM also overcame most of the problems investors faced with MV optimization by
simply adding a neutral reference point (CAPM equilibrium) for expected returns. It also
provided the investors the ability incorporate their views effectively in the asset allocation
process.
As a result, BLM has received broad acceptance within the investment community and
19
Black, Fischer, and Robert Litterman. "Asset Allocation: Combining Investor Views with Market
Equilibrium." Goldman, Sachs & Co., 1990.
Black, Fischer, and Robert Litterman. "Global Asset Allocation with Equities, Bonds, and Currencies."
Goldman, Sachs & Company Fixed Income Research, 1991.
Black, Fischer, and Robert Litterman. "Global Portfolio Optimization." Financial Analysts Journal 48, no. 5
(Sep/Oct 1992): 28-43.
Black, Fisher. "Universal Hedging: Optimizing currency risk and reward in international equity
portfolios." Financial Analysts Journal 45, no. 4 (1989): 16-22.
Cochrane, John H. "Portfolio advice for a multifactor world." Economic Perspectives: Federal Reserve
Bank of Chicago, Q3 1999: 59-78.
Giacometti, M., I. Bertocchi, T.S. Rachev, and F. Fabozzi. "Stable distributions in the Black-Litterman
approach to asset allocation." Quantitative Finance 7, no. 4 (2007): 423-433.
He, Guangliang, and Robert Litterman. "The Intuition Behind Black-Litterman Model Portfolios."
Goldman Sachs Investment Management Series, December 1999.
Herold, Ulf. "Portfolio Construction with Qualitative Forecasts." Journal of Portfolio Management 30, no.
1 (2003): 61-72.
Idzorek, Thomas M. "A Step-by-Step Guide to the Black-Litterman Model: Incorporating user-specified
confidence levels." Zephyr Working Document, 2004.
Krishnan, Harl, and Mains Norman. "The Two-Factor Black Litterman Model." Risk 18, no. 7 (2005): 69-
73.
Litterman, Robert. Modern Investment Management: An Equilibrium Approach. New Jersey: John Wiley
& Sons, 2003.
Meucci, Attilio. "The Black-Litterman Model: Original Model and Extensions." In The Encyclopedia of
Quantitative Finance. Wiley, 2010.
Satchell, Stephen, and Alan Scowcroft. "A Demystification of the Black-Litterman Model: Managing
Quantitative and Traditional Construction." Journal of Asset Management 1, no. 2 (2000): 138-150.
20