European Journal of Operational Research 258 (2017) 564–572
Contents lists available at ScienceDirect
European Journal of Operational Research
journal homepage: www.elsevier.com/locate/ejor
Stochastics and Statistics
On the Bayesian interpretation of Black–Litterman
Petter Kolm
a
, Gordon Ritter
a , b , ∗
a
Courant Institute of Mathematical Sciences, New York University, 251 Mercer St., New York, NY 10012, United States
b
Rutgers University and Baruch College, United States
a r t i c l e i n f o
Article history:
Received 21 April 2016
Accepted 15 October 2016
Available online 20 October 2016
Keywords:
Finance
Investment analysis
Bayesian statistics
Black–Litterman
Portfolio optimization
a b s t r a c t
We present the most general model of the type considered by Black and Litterman (1991) after fully
clarifying the duality between Black–Litterman optimization and Bayesian regression. Our generalization
is itself a special case of a Bayesian network or graphical model. As an example, we work out in full detail
the treatment of views on factor risk premia in the context of APT. We also consider a more speculative
example in which the portfolio manager specifies a view on realized volatility by trading a variance swap.
©2016 Elsevier B.V. All rights reserved.
1. Introduction
The topic of portfolio optimization in the style of Black and Lit-
terman (1992, 1991) seems to have generated more than its share
of confusion over the years, as evidenced by articles with titles
such as “A demystification of the Black–Litterman model” ( Satchell
& Scowcroft, 20 0 0 ), etc. The method itself is often described as
“Bayesian” but the original authors do not elaborate directly on
connections with Bayesian statistics.
In language universally familiar to statisticians ( Robert, 2007 ), a
Bayesian statistical model consists of:
1. A vector-valued random variable x ∈ X ⊆ R
d
distributed accord-
ing to f ( x | θ), where realizations of x have been observed and
only the parameter θ (which belongs to a real vector space
⊆ R
) is unknown, and
2. A prior density π( θ) on .
The function f ( x | θ) is called the likelihood and, after condition-
ing on θ, forms a density on the data space X ⊆ R
d
. The posterior
is the density on proportional to f ( x | θ) π ( θ), and the normaliza-
tion factor drops out of certain calculations. In Bayesian statistics,
all statistical inference is based on the posterior.
The paper by Litterman and He (1999) contains many references
to a “prior” but only one mention of a “posterior” without details,
and no mention of a “likelihood.”
In
the present note, we clarify the exact nature of the Bayesian
statistical model to which Black–Litterman optimization corre-
∗
Corresponding author.
E-mail addresses: ritter@post.harvard.edu , gordon.ritter@gsacapital.com (G. Rit-
ter).
sponds, in terms of the prior, likelihood, and posterior. In the pro-
cess we also lay out the full set of assumptions made, some of
which are glossed over in other treatments.
2. Black, Litterman, and Bayes
Consider a view such as “the German equity market will out-
perform a capitalization-weighted basket of the rest of the Euro-
pean equity markets by 5%,” which is an example presented in
Litterman and He (1999) . Let p ∈ R
n
denote a portfolio which is
long one unit of the DAX index, and short a one-unit basket which
holds each of the other major European indices (UKX, CAC40, AEX,
etc.) in proportion to their respective aggregate market capitaliza-
tions, so that
i
p
i
= 0 . Let q = 0 . 05 in this example. This view
may be equivalently expressed as
E [ p
r ] = q ∈ R (1)
where r is the random vector of asset returns over some subse-
quent interval, and q denotes the expected return, according to the
view. If there are multiple such views, say
E [ p
i
r ] = q
i
, i = 1 . . . k
then the portfolios p
i
are more conveniently arranged as rows of a
matrix P , and the statement of views becomes
E [ P r ] = q for q ∈ R
k
. (2)
In the language of statistics, the core idea of Black and Lit-
terman (1991) is to treat the portfolio manager’s views as noisy
observations which are useful for performing statistical inference
concerning the parameters in some underlying model for r . For ex-
ample, if
r ∼ N( θ, ) (3)
http://dx.doi.org/10.1016/j.ejor.2016.10.027
0377-2217/© 2016 Elsevier B.V. All rights reserved.