Measuring Portfolio Diversication

Based on Optimized Uncorrelated Factors

Attilio Meucci
attilio.meucci@symmys.com

Alberto Santangelo
alberto.santangelo@hotmail.com

Romain Deguest
romain.deguest@edhec-risk.com

this revision: 08 October 2014

latest revision and code: symmys.com/node/599

Abstract1
We measure diversication in terms of the "Eective Number of Minimum-Torsion Bets", namely a
set of uncorrelated factors, optimized to closely track the factors used to allocate the portfolio. This
way we introduce a novel notion of "absolute risk contributions", which generalizes the "marginal
contributions to risk" in traditional risk parity. We discuss the advantages of the Minimum-Torsion
Bets over the traditional approach to diversication based on marginal contributions to risk. We
present a case study in the S&P 500.
Fully documented code is available at symmys.com/node/599.

JEL Classication: C1, G11

Keywords: Eective Number of Bets, PCA, Diversication Distribution, marginal risk contributions, Procrustes problem

1 The authors are grateful to Marcello Colasante, David Elliott, Bruno Dupire, Yashin Gopi, Lionel Martellini, Sergei
Polevikov, and Thierry Roncalli

Introduction

In recent years the practitioners and academic nancial community has witnessed a surge in interest in the concept of risk parity, as well as the broader concept of diversication management, see
[Roncalli, 2013] for a review and references.
In traditional risk parity, diversication is measured in terms of marginal risk contributions from
each individual risk factor. Such contributions are spurious, because in reality they contain eects
from all the factors at once. Furthermore, there exist no clear metric to quantify the diversication
represented by the marginal risk contributions.
In this article we propose an alternative approach to risk parity based on the Eective Number of
Bets in [Meucci, 2009a]: instead of the marginal contributions from correlated factors, we measure the
true contributions from uncorrelated bets. Then the Eective Number of Bets precisely quantify the
diversication level, summarizing in one number the ne structure of diversication contained in the
set of uncorrelated bets in our portfolio.
In the original paper, the uncorrelated bets are the markets principal components. The Principal
Components Bets have spurred interest and called for extensive empirical analysis, see [Frahm and Wiechers, 2011],
[DFine, 2011], [Lohre et al., 2011], [Lohre et al., 2012], [Deguest et al., 2013]. However, the principal
components are suboptimal, because they are purely statistical entities, not related to the investment
process.
In this article we introduce a natural set of uncorrelated bets to manage diversication, the
Minimum-Torsion Bets, which are the uncorrelated factors closest ("minimum-torsion") to the factors
used by the portfolio manager. The contributions to risk from the Minimum-Torsion Bets constitutes
a generalization of the marginal contributions to risk used in traditional risk parity.
The remainder of the paper is organized as follows: in Section 2 we revisit the general Eective
Number of Bets framework in the context of factor-based risk parity and diversication management. In
Section 3 we review the suboptimal implementation of the Eective Number of Bets, namely when the
bets are represented by the principal components. In Section 4 we introduce the natural implementation
of the Eective Number of Bets, namely when the bets are the Minimum-Torsion Bets. In Section 5 we
highlight the key advantages of our approach to risk parity, based on Eective Number of MinimumTorsion Bets, and the traditional approach to risk parity, based on marginal contributions to risk. In
Section 6 we test our approach in a practical case study in the S&P 500. In Section 7 we conclude. In
the Appendix A we detail all the technical proofs.

Eective Number of Bets

Here we review the Eective Number of Bets approach in [Meucci, 2009a], using a notation more
suitable for the generalizations to follow. Refer to the original paper for all the details.
Consider an arbitrary portfolio which gives rise to a yet to be realized projected return R. In assetbased portfolio management, a portfolio is a combination of m correlated assets (stocks, options,
Pm bonds,
futures, ...), and the portfolio return is a weighted average of the return of each asset R = m=1 wm Rm ,
where wm represent the weight of the m-th asset in the portfolio.
More in general, in factor-based portfolio management, and in factor-based risk parity, a portfolio
is a combination of n correlated factors, such as momentum, value, etc. Then, the portfolio return is
a combination of the factor returns
Pn
R = n=1 bn Fn ,
(1)
where bn represent the exposure of the portfolio to the n-th factor, and where the factors may include
a portfolio-specic residual.
Typically, but not necessarily m (the number of factors) is much smaller than n (the number of
assets). Furthermore, clearly, asset-based portfolio management represents a special case of factorbased management (1), where m = n, the factors are the asset returns Fm = Rm and the exposures
are the portfolio weights bm = wm .
2

Let us assume for now an ideal, apparently non-realistic scenario, where we can express the portfolio
return as a combination of n Bets, or uncorrelated factors
Pn
R = n=1 bn Fn ,
(2)
where CrfFn ; Fm g = 0 if n 6= m. Then, we can compute the Diversication Distribution, namely true
relative contributions to total risk from each bet
pn

Vfbn Fn g
;
V fRg

n = 1; : : : ; n,

(3)

where V denotes the variance. Notice that, as for any distribution, the masses pn sum to one and are
non-negative
Pn
pn 0; n = 1; : : : ; n,
(4)
n=1 pn = 1;

The Diversication Distribution (3) provides a detailed picture of the portfolio concentration structure. A portfolio is well diversied among the n factors, and thus achieves risk parity, if the masses
p1 ; : : : ; pn are equal, or equivalently if the Diversication Distribution is uniform. To quantify diversication precisely, we use the exponential of the entropy, a tool from information theory, that measures
the uniformity of a distribution. Accordingly, we dene the Eective Number of Bets as follows
Pn
n=1 pn ln pn .
N e
(5)

In the case of full concentration, i.e. when all the risk loads on one single factor, that factors risk
contribution (3) is 1, while all the other contributions are 0, and therefore N = 1, the lowest possible
value. At the opposite extreme, in the case of full diversication, the contributions to risk (3) from all
the factors are equal, and N = n, the maximum possible value. For the intermediate cases 1 N n.

Principal Components bets

To measure diversication via the Eective Number of Bets, we need to express our portfolio returns as
a combination of uncorrelated terms, as in (2). To do so, one option is to use the principal components
of the original factors F in (1), as suggested in [Meucci, 2009a].
Accordingly, we compute the covariance matrix of the factors F
Cv fF g. Next, we perform
the principal component decomposition e 2 e0 = F , where is the diagonal matrix of the singular
values (square-root of eigenvalues) of F and e is the matrix whose columns are the eigenvectors of
0
0
F , which are orthogonal, and are normalized with length one ee = e e = I, the identity matrix.
The eigenvectors generate n uncorrelated factors F PC and n new portfolio exposures bPC , which
allow us to express the portfolio return (1) in the uncorrelated format (2), as follows
F PC

e0 F ;

bPC

e0 b;

R = b0 F = bPC F PC ,

(6)

where the last equality follows from b0 F = b0 ee0 F .

Then given the exposures b, we can compute the Eective Number of Bets (5)
pPC (b) =

(e0 b) (e0
b0 F b

F b)

NPC (b) = e

pPC (b)0 ln pPC (b)

(7)

where denotes the term-by-term product, see Appendix A.1.

The principal components approach provides a formal set of uncorrelated factors from which to
compute the Eective Number of Bets. However, it presents several problems.
First, the principal components bets tend to be statistically unstable, especially those relative to
the lowest eigenvalues.
3

Second, the principal components bets are not invariant under simple scale transformations, as
provided by a diagonal matrix d with positive entries
F
?
?
scaling y

dF

PCA decorrelation

F PC
?
?scaling
y

PCA decorrelation

(8)

! (dF )PC 6= dF PC

To illustrate the above diagram, suppose that we want to measure the uncorrelated sources of risk in a
portfolio in terms of P&L, rather than in terms of returns. The returns are the P&L normalized by a
constant, so we expect the uncorrelated sources of risk to be the same. However, the Eective Number
of Bets based on principal components of the P&L and of the return are dierent. To illustrate with a
second example, suppose that we wish to measure some returns in basis points, instead of percentage
points. The ensuing principal components bets will change dramatically, which is unacceptable.
Third, the principal components bets are not unique. Indeed, as discussed in [Deguest et al., 2013],
if en is one of the n eigenvectors, so is its opposite en , and thus there are basically 2n possible
combinations of principal components bets.
Fourth, the principal components bets are in general not easy to interpret, and hence disconnected
from the decision process. In particular, in a dynamic setting, the meaning of the PCA factors changes
from one date to another date, except possibly for the very rst few factors.
Fifth, the principal components bets give rise to counter-intuitive results.

Example 1 To illustrate the counter-intuitive results obtained with the principal component framework, consider the equal-load (equal-weight, in asset-based allocation) portfolio beq ( n1 ; : : : ; n1 )0 . Also,
consider an idealized, though non-realistic, market where all the factors have equal volatility and equal,
positive pair-wise correlation
2

F ]m;n

Cv fFm ; Fn g =

for all m 6= n
for m = n

(9)

In such a homogeneous market, if the correlation > 0 is very small, we would expect the equal-load
portfolio to be highly diversied, giving rise to a number of uncorrelated bets close to the number of
factors NPC (beq ) n. Instead, the equal-load portfolio always displays maximum concentration, i.e.
only one (!) bet
NPC (beq ) = 1.
(10)
The counter-intuitive full-concentration e ect (10) follows because the equal-load portfolio is fully exposed to the rst principal component and not exposed to any other principal component, see the proof
in Appendix A.2.

Minimum-Torsion Bets

The Eective Number of Bets approach to risk parity builds on the uncorrelated decomposition (2).
Hence, unlike the standard approach to risk parity based on marginal contributions to risk, the Eective
Number of Bets approach highlights the contributions from truly separate sources of risk.
However, if the uncorrelated portfolio decomposition (2) is achieved via the principal components
bets (6), we obtain suboptimal results for the several reasons highlighted in Section 3.
Fortunately, the principal components bets are not the only zero-correlation transformation of the
original factors F that allows to express the portfolio as in the uncorrelated decomposition (2). There
4

exist several alternative linear transformations F = tF , or torsions, of the original factors F , that are
uncorrelated, and that are represented by a suitable n n decorrelating torsion matrix t.
For instance, one could think of independent component analysis, see [Back and Weigend, 1997],
or more simply we could use the lower-triangular Cholesky decomposition F
Cv fF g ll0 , where
1
1
1 0 10
1
t
l . Indeed Cvfl F g = l ll l
= I, and thus F = l F are uncorrelated. However, such
transformations display the same problems as the principal component approach. Most notably, the
resulting uncorrelated factors F are not interpretable, as in general they bear no relationship with the
original factors F that are used to manage the portfolio.

4.1

Problem formulation

Here, we propose a natural, interpretable denition for the de-correlating transformation and the
resulting uncorrelated factors: we choose the minimum torsion linear transformation that least disrupts
the original factors F . More precisely, among all the torsions t that ensure that the new factors are
uncorrelated, we select the one that minimizes the tracking error with respect to the original factors
tMT

argmin
CrftF g=I n

NTE ftF kF g,

where NTE denotes the multi-entry normalized tracking error

s
P
Zn Fn
1
g.
NTE fZkF g
n Vf
n
SdfFn g

(11)

(12)

Figure 1: Minimum-Torsion Bets are the uncorrelated (orthogonal) factors closest to the original
factors, Principal Component Bets are uncorrelated (orthogonal) factors with no clear connection to
the original factors
Suppose that we can solve the minimum torsion optimization (11), which we do further below in
this section. Then we introduce the Minimum-Torsion Bets F MT and the respective Minimum-Torsion
Exposures bMT , and use them to express the portfolio return (1) in the uncorrelated format (2), as
follows
0
0 1
F MT tMT F ;
bMT tMT b;
R = bMT F MT .
(13)
The interpretation of the Minimum-Torsion Bets F MT is easily visualized geometrically in Figure 1,
by equating factors to vectors, and no-correlation to orthogonality. Whereas the Minimum-Torsion
5

Bets F MT are the closest to the original factors F used to manage the portfolio, the PCA bets F PC
dened in (6) bear no close relationship to the management factors F .
Then, the Minimum-Torsion Bets and loadings (13) allow us to compute the Eective Number of
Minimum-Torsion Bets
pMT (b) =

0 1

(tMT b) (tMT
b0 F b

F b)

NMT (b) = e

pMT (b)0 ln pMT (b)

(14)

see Appendix A.1.

The Minimum-Torsion Bets (13) address all the problems of the principal components bets (6). In
particular, the normalization in the tracking error (12) allows us not to worry about non-homogenous
factors F measured in completely dierent units, such as interest rates and implied volatilities. In
other words, unlike in the case of uncorelated bets dened by the principal components (8), for the
Minimum Torsion Bets the following diagram holds
F
?
?
scaling y

dF

M T decorrelation

F MT
?
?scaling
y

M T decorrelation

(15)

! (dF )MT = dF MT

Example 2 To provide a rst taste of the intuitive nature of the Minimum-Torsion Bets, let us consider again the idealized homogeneous market with equal volatilities and arbitrary small homogeneous
correlations (9). Furthermore, let us consider again the equal-load (equal-weight, in asset-based allocation) portfolio beq
( n1 ; : : : ; n1 )0 . Unlike with Principal Component Bets (10), the Number of
Minimum-Torsion E ective Bets (14) for the equal-load portfolio is, as intuition suggests, the largest
possible
NMT (beq ) = n,
(16)
see Appendix A.2. This result is very intuitive: in an uncorrelated market each position of equal size
represents a separate bet.

4.2

Problem solution

The solution of the minimum torsion optimization (11) is a special instance of a quadratically constrained quadratic program [W], related to the solution of the orthogonal Procrustes problem [W].
Adapting from [Everson, 1997], we obtain the solution by rst computing analytically a starting guess,
and then perturbing the starting guess via an e cient recursive algorithm, as follows.
Let us denote by F the vector of the factors standard deviations, extracted with the correlation
matrix C F from the covariance matrix F Cv fF g, as follows
F

dg(

F ) CF

dg(

F ).

(17)

Then, let us factor as in [Meucci, 2009b] the correlation matrix via its Riccati root c, namely the
symmetric positive denite matrix such that
C F = cc0 = c2 .

(18)

The Riccati root of the correlation matrix C F is easily computed in terms of the PCA decomposition
C F = g 2 g 0 , where is the diagonal matrix of the singular values (square-root of eigenvalues) of the
6

correlation matrix C F and g is the matrix whose columns are the respective eigenvectors, which are
orthogonal, and are normalized with length one gg 0 = g 0 g = I, the identity matrix. Then the Riccati
root of the correlation matrix (18) reads explicitly
c = g g0 .
Next, we compute a perturbation matrix
A.3.2 for the rationale)

0.
1.
2.
3.
4.
5.

(19)

recursively with the algorithm below (refer to Appendix

= Perturb(c)
Minimum-Torsion recursion
Initialize
d
I
1
Riccati root
u
(dc2 d) 2
Rotation
q
u 1 dc
Stretching
d
dg(dg 1 (qc))
Perturbation
dq
If convergence, output ; else go to 1

(20)

where the operator dg 1 (m) extracts the n 1 vector on the principal diagonal of the n n matrix
m; and the operator dg(v) embeds the n 1 vector v into the principal diagonal of a square matrix
which is zero anywhere else.

Example 3 For instance from a correlation

1 0:5 0:3
1 0:1 )
0:3 0:1 1

C F = ( 0:5
we obtain

0:9535 0:0016 0:0026

0:9661
0:0001 )
0:0027 0:0001 0:9886

0:9544 0:2580 0:1503

c = ( 0:2580 0:9656 0:0313 );

= ( 0:0017

0:1503 0:0313 0:9881

The algorithm (20) converges extremely fast, within a dozen iterations, or fractions of a second,
even when has n2
150,000 entries (!), the dimensions required to handle the S&P case study in
Section 6.
The solution of the minimum torsion problem (11) then reads
tMT = dg(

F)

dg(

F)

(21)

where we emphasize that the factors cannot be collinear, or else the inverse Riccati root c 1 is not
dened.
The minimum torsion transformation (21) denes the Minimum-Torsion Bets and the MinimumTorsion Exposures, as in (13); and the Minimum-Torsion Diversication Distribution and the Eective
Number of Minimum-Torsion Bets, as in (14).
If we switch o the numerical perturbation term, which amounts to setting
I in the minimum
torsion transformation (21), we obtain an approximate minimum-torsion
t

dg(

F)c

dg(

F)

tMT .

(22)

As we show in see Appendix A.3, the approximate minimum-torsion (22) solves the original minimumtorsion problem (11) under the additional constraint that the volatilities of the new factors be the
same as the volatilities of the original factors
SdftF g =
7

F.

(23)

The approximate minimum-torsion (22) is essentially the solution of the orthogonal Procrustes
problem rst derived by [Schoenemann, 1966]. Geometrically, the approximate minimum-torsion generates orthogonal bets tF by rotating the original factors F . However, we veried numerically that
the discrepancy between the approximate minimum-torsion bets tF and the true Minimum-Torsion
Bets tMT F can become relevant in highly correlated markets.
We note that the approximate minimum torsion transformation (22) resembles the decorrelating
solution in [Klein and Chow, 2010]. However, such solution i) applies to the empirical distribution of
the factors, rather than arbitrary distributions; ii) it applies to the covariance matrix, rather than the
correlation matrix, thereby being sensitive to the units in which the factors are measured; iii) it does
not account for the perturbation term as the exact solution (21) does.

Bets versus marginal contributions to risk

Here we reect on the main conceptual dierences between the traditional approach to risk parity,
based on marginal contributions to risk, and the present approach, based on the Eective Number of
Minimum-Torsion Bets, which we summarize in the table below

Risk contrib.

Traditional

Eective Number of Bets

Marginal Contributions

Diversication Distributions

Expression
Meaning
Properties

b (
b0

F b)

spurious contributions
from original factors

mn = 1;

(tMT b) (tMT
b0 F b

Fb

F b)

(24)

proper contributions
from Minimum-Torsion Bets

mn 7 0

pn = 1;

pn

The key to the traditional risk parity approach are the Marginal Contributions to Risk
m

b (
b0

F b)
Fb

(25)

see e.g. [Roncalli, 2013]. In the traditional risk parity approach, a portfolio is diversied if the
Marginal Contributions to Risk m are uniform. The key to our approach to risk parity based on
Eective Number of Bets is the Diversication Distribution of the Minimum-Torsion bets p, dened
in (14): a portfolio is diversied if p is uniform.
The main dierence, and at the same time the main weakness of the Marginal Contributions to
Risk m with respect to the Diversication Distribution p, is the fact that m represent dierential
contributions, i.e. they represent the sensitivity of risk to a small change in exposure to a given factor.
Hence, unlike the Diversication Distribution p, they do not represent the separate contributions to
risk from a given factor.
Furthermore, the Marginal Contributions to Risk m sum to one, just like the Diversication Distribution p. However, unlike the Diversication Distribution p, the Marginal Contributions to Risk m
are not necessarily positive, due to either negative correlations or the presence of negative exposures
to factors. As a result, when enforcing traditional risk parity, one focuses on the absolute values of the
Marginal Contributions to Risk jmn j, which no longer sum to one.
Finally, notice that if the factors F are uncorrelated, then the Minimum-Torsion transformation
tMT is the identity, and thus the Marginal Contributions to Risk and the Diversication Distribution
are one and the same, m = p.
To summarize, both traditional risk parity and Eective Number of Minimum-Torsion Bets apply
in full generality across markets with arbitrary factors. When the factors in the traditional risk parity
approach are uncorrelated, the Marginal Contributions to Risk display all the palatable features of
the Diversication Distribution, namely they sum to one, they are positive, and they are the true
contributors to risk. In the more general case where the factors are correlated, the traditional risk
parity approach incurs problems.
8

Hence, we can interpret the Eective Number of Minimum-Torsion Bets approach as a generalization of the traditional risk parity approach, which addresses the issues of the latter approach.

Case study

Figure 2: Diversication through time of an equal-weight portfolio of stocks

To illustrate, we consider an investment in n = 392 stocks in the S&P 500 Index (for simplicity, we
consider the stocks alive through the whole analysis period). In this case the factors F and respective
exposures b in (1) are the n stock returns and the n portfolio weights respectively.
Let us consider a simple equal-load portfolio beq
( n1 ; : : : ; n1 )0 . We estimate every month the
covariance matrix of the of the S&P stock returns F using a one-year rolling window of daily observations, and ltering the smallest eigenvalues to ensure positive deniteness. Using the portfolio
loadings and the returns covariance we compute the Minimum-Torsion Diversication Distribution
pMT (beq ) and the Minimum-Torsion Eective Number of Bets NMT (beq ) of the equal-load portfolio,
as in (14).
In Figure 2 we display the results. Each month, we sort the stocks in decreasing order of risk
contribution, as measured by the Minimum-Torsion Diversication Distribution pMT (beq ). The fairly
homogeneous structure of the S&P 500 stocks and the equal-load allocation provide a portfolio that
is not as diversied as one would expect. Unlike the portfolio weights, the Minimum-Torsion Diversication Distribution is not at: some stocks are riskier than others, a fact well known to portfolio
managers.
The ne structure of diversication represented by the Minimum-Torsion Diversication Distribution is summarized into the Minimum-Torsion Eective Number of Bets NPC (beq ). The Eective
Number of Bets is strictly less than the number of stocks n, because of the non-homogeneous contributions to risk from each stock. However, as intuition suggests, the Eective Number of Bets is of the

Figure 3: Intuitive diversication measure via Mimumum-Torsion versus counter-intuitive diversication measure via principal components
order of a few hundreds, given that the Minimum-Torsion Diversication Distribution is not too steep:
0

NMT (beq )
< 1.
n

(26)

The intuitive result (26) is the empirical counterpart of the similar theoretical full-diversication result
(16).
For comparison, we also compute the principal component Diversication Distribution pPC (beq )
and the principal component Eective Number of Bets NPC (beq ) of the equal-load portfolio, as in (7).
The Diversication Distribution is now too steep to display, loading basically in full on the rst entry
(which is no longer interpretable and not directly comparable with the portfolio weights). As a result,
as we see in Figure 3, the normalized Eective Number of Bets is almost zero
0

NPC (beq )
n

1.

(27)

The counter-intuitive result (27) is the empirical counterpart of the similar, extreme, theoretical fullconcentration result (10).
Finally, we compute the marginal risk contributions of the equal-load portfolio m(beq ), as in (25)
and display them in Figure 4. The overall risk prole is similar to the Diversication Distribution
pMT (beq ) in Figure 2. However, even with such a homogeneous market and portfolio, we notice
possibly spurious, sharp negative contributions at the beginning of the sample.

Conclusions

We have introduced the Minimum-Torsion Bets, the set of uncorrelated factors which closely tracks the
factors used in the allocation process. With the Minimum-Torsion Bets we have given new life to the
Eective Number of Bets approach to risk parity and, more in general, diversication management.
Indeed, unlike the Principal Component Bets originally used to measure the Eective Number of Bets,
the Minimum-Torsion Bets are easily interpretable and give rise to intuitive results.
We have highlighted the improvements of the Eective Number of Minimum-Torsion Bets over the
standard approach to risk parity, which relies on marginal, rather than total, contributions to risk.
We have illustrated how to use the Eective Number of Minimum-Torsion Bets to measure the
diversication of a portfolio of stocks in the S&P500.
10

Figure 4: Marginal contributions to risk through time of an equal-weight portfolio of stocks

References
[Back and Weigend, 1997] Back, D. A. and Weigend, A. (1997). A rst application of independent component analysis to extracting structure from stock returns. http://ssrn.com/abstract=1283022.
Working Paper.
[Deguest et al., 2013] Deguest, R., Martellini, L., and Meucci, A. (2013). Risk parity and beyond from asset allocation to risk allocation decisions. http://ssrn.com/abstract=2355778. EDHECRisk Working Paper.
[DFine, 2011] DFine, G. (2011). Asset allocation under a conditional diversication measure. https:
//www.maths.ox.ac.uk/system/files/attachments/300911.pdf.
[Everson, 1997] Everson, R. (1997). Orthogonal, but not orthonormal, procrustes problems. http://
citeseerx.ist.psu.edu/viewdoc/download?rep=rep1&type=pdf&doi=10.1.1.54.9741. Working Paper.
[Frahm and Wiechers, 2011] Frahm, G. and Wiechers, C. (2011). On the diversication of portfolios of risky assets. http://hdl.handle.net/10419/45354. Discussion papers in statistics and
econometrics.
[Klein and Chow, 2010] Klein, R. F. and Chow, V. (2010). Orthogonalized equity risk premia and
systematic risk decomposition. Working Paper 10-05, Dept. of Economics, West Virginia University.
[Lohre et al., 2011] Lohre, H., Neugebauer, U., and Zimmer, C. (2011). Diversifying risk parity. http:
//ssrn.com/abstract=1974446. Journal of Risk, Forthcoming.
[Lohre et al., 2012] Lohre, H., Neugebauer, U., and Zimmer, C. (2012). Diversied risk parity strategies for equity portfolio selection. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=
2049280&%3Ca%20href=. The Journal of Investing, 21, 111-128.
11

[Meucci, 2009a] Meucci, A. (2009a). Managing diversication. http://symmys.com/node/199. Risk,

74-79.
[Meucci, 2009b] Meucci, A. (2009b). Simulations with exact means and covariances. http://symmys.
com/node/162. Risk, 22(7), 89-91.
[Roncalli, 2013] Roncalli, T. (2013). Introduction to Risk Parity and Budgeting. Chapman and Hall.
[Schoenemann, 1966] Schoenemann, P. H. (1966). A generalized solution of the orthogonal procrustes
problem. Psychometrika, 31:110.

12

Appendix

In this appendix we discuss technical results that can be skipped at rst reading.

A.1

Explicit expressions for Eective Number of Bets

The denominator in the Diversication Distribution (3) follows from the expression of the portfolio
return as a combination of factors (1) and reads
V fRg = Vfb0 F g = b0 VfF gb = b0

(28)

F b.

Assume that the Bets are an uncorrelated linear transformation of the factors
F

tF ;

0 1

(29)

b.

Then numerator of the Diversication Distribution (3) reads

Vfb F g = b VfF g b = b (CvfF gb)
0 1

= (t

b) (t

Ft

0 0

0 1

b) = (t

(30)
b) (t

F b)

Dividing the numerator (30) by the denominator (28) we obtain the explicit formula Diversication
Distribution (3), as follows
0 1
(t b) (t F b)
.
(31)
p=
b0 F b
The Principal Components expression (7) follows from setting t = e0 in the Diversication Distribution (31), as prescribed by the Principal Components transformation (6), which is a special case of
(29).
The Minimum-Torsion expression (14) follows from setting t = tMT in the Diversication Distribution (31), as prescribed by the Minimum-Torsion transformation (13), which is a special case of
(29).

A.2

Number of bets in homogeneous markets

Consider n factors F
(F1 ; : : : ; Fn )0 . Assume that the variances are all equal to 2 and all the
pair-wise correlations are equal to > 0, as in (9).
As proved in Appendix A.4 of [Meucci, 2009a], the equally-loading portfolio beq is the eigenvector
of the covariance matrix F relative to the largest eigenvalue 21 , or F beq = 21 beq (the eigenvectors
are dened modulo a scale factor, so beq need not satisfy b0eq beq = 1). As a result, the Diversication
0
Distribution (7) reads pPC (beq ) = (1; 0; : : : ; 0) and thus NPC (beq ) = 1, as in (10).
On the other hand, the minimum torsion transformation acts equally on all the factors, which are
0 1
indistinguishable. Hence, for symmetry reasons (tMT beq ) / 1 is a vector of equal entries, and so is
(tMT F beq ) / 1. Therefore the Diversication Distribution (14) reads pMT = ( n1 ; : : : ; n1 )0 and thus
NMT (beq ) = n, as in (16).

A.3

Minimum-torsion optimization
1=2

Let us denote the vector of the standard deviations in the covariance matrix F by
(dg 1 F ) ,
1
where the operator dg x extracts the diagonal from a matrix x. Let us denote by dg(v) a diagonal
matrix with the vector v on the diagonal.
Solving (11) is equivalent to solving
tMT

argmin tr (Cvfdg( )
CrftF g=I

13

(tF )

dg( )

F g).

(32)

Let us dene the normalized factors Z

dg( )

CvfZg = C F
Noting that Crftdg( )Zg = I , Crfdg( )
Crfdg( )

1 tdg(

F , whose covariance is the correlation matrix

dg( )
1

dg( )

(33)

tdg( )Zg = I, we write (32) as follows

argmin

tMT

tr (Cvfdg( )

t dg( )Z

Zg).

(34)

)Zg=I

Equivalently, we can write

tMT = dg( ) x dg( )

(35)

where x solves
argmin tr (Cvf(x

(36)

I)Zg)

CrfxZg=I

= argmin tr ((x

I)C F (x0

I))

CrfxZg=I

= argmin tr (xC F x0

C F x0 + C F )

xC F

CrfxZg=I

= argmin tr (xC F x0

2xC F ) + n

CrfxZg=I

where the last equality follows from the symmetry of C F .

Let us denote by D the set of diagonal matrices with full rank, and let us introduce the Riccati
root of the correlation matrix
1
(37)
c = c0 (C F ) 2 .
Then x in (35) is the solution of
x = argmin tr (xcc0 x0

2xcc)

(38)

xcc0 x0 2D

A.3.1

Constrained analytical solution

Let us impose the stronger constraint that volatilities are preserved (23), which amounts to
CvfxZg = xC F x0 = I.

(39)

We emphasize that the constraint (39) does restrict the solution, see (A.3.2) below. From the constraint
(39) we obtain tr (xC F x0 ) = tr (I) = n. As a result, the minimization (38) becomes
x = argmax tr (xcc),

(40)

xcc0 x0 =I

Dening y

xc, we can write

x

[argmax tr (yc)]c

=c

(41)

yy =I

where argmaxyy0 =I tr (yc) = I because c is symmetric with positive eigenvalues, see (64) below.
A.3.2

Unconstrained numerical solution

To solve the general problem (38) without the additional constraint on volatilities (39), let us dene
xc. Then
x = c 1,
(42)
where
argmin tr (
0 2D

14

2 c).

(43)

Adapting from [Everson, 1997], we can address the optimization (43) with an iterative algorithm
that solves two alternating steps. Let us write
(44)

dq,
where d is diagonal with full rank and q is orthonormal, in such a way that the constraint
0
dqq 0 d = d2 2 D is satised.
0
Step 1. Assume we know the diagonal matrix d 2 D such that
= d2 . Then (43) becomes
0

argmin tr (

2 c) = argmax tr (c ),

0 =d2

(45)

0 =d2

where we used the symmetry of the Riccati root c. The problem (45) is in the same format as (52)
below. The solution then follows from (63) and reads
1

= d((dc2 d) 2 )

dc.

(46)

Since d is invertible, from (44) we obtain

q=d

= ((dc2 d) 2 )

(47)

dc.

Step 2. Assume we know the orthogonal matrix q such that qq 0 = I and

becomes
d argmin tr (dqq 0 d 2dqc) = argmin tr (d2 2dqc).
dqq 0 d2D

= dq. Then (43)

(48)

d2 2D

In order to solve (48), we dierentiate its objective function

f (d)

tr (d2 )

2tr (dqc) = tr (d2 )

2tr (cdq)

(49)

with respect to each entries on the diagonal of d = dg(d1 ; : : : ; dn )

P
@f
@ P 2
=
( m dm 2 m;k cm;k dk qk;m )
@dn
@dn
P
= 2(dn
m qn;m cm;n )
= 2(dn

[qc]n;n ),

(50)

Setting to zero the derivatives we obtain dn = [qc]n;n for all n = 1; : : : ; n, or

d = dg(dg

(qc)):

Alternating (47) and (51) we arrive at the algorithm (20), which we initialized with d

A.4

(51)
I.

The constrained Procrustes problem

Adapting from [Schoenemann, 1966], here we present the solution to the orthogonal Procrustes problem
argmax tr (kz),

(52)

zz 0 =d2

where k is a real matrix and d2 is diagonal with full rank.

Consider the singular value decomposition of the product
kd

pdg( )s0 ,

(53)

where p and s are orthonormal matrices and is a vector with nonnegative entries. Using (53) and
since d is invertible, we can dene the change of variables
y

s0 d
15

zp.

(54)

Then the optimization target in (52) reads

dsyp0 ) = tr (pdg( )yp0 )
Pn
= tr (dg( )yp0 p) = tr (dg( )y) = n=1 n;n yn;n ,
1

tr (kz) = tr (pdg( )s0 d

(55)

and the constraint in (52) is equivalent to

yy 0 = s0 d

zpp0 z 0 d

s = s0 d

zz 0 d

s = s0 d

1 2

d d

s = s0 s = I.

(56)

Hence the solution of (52) is

z = dsyp0 ,
where
y

argmax
yy 0 =I

(57)

Pn

n=1 n;n yn;n .

(58)

Since n;n
0 for all n = 1; : : : ; n, the maximum is attained by yn;n = 1 for all n = 1; : : : ; n, which
implies y = I. Substituting this in (57) we obtain the solution to (52)
z = dsp0 .

(59)

Notice that if k is invertible, then we can simplify the solution (59). Indeed, we can write
sp0 = sdg( )

(53)

s0 s dg( )p0 = sdg( )

|{z}

(61)

s dk0 = u

1 0

dk0 ,

(60)

where

sdg( )s0 .

(61)

It is easy to see that u is the Riccati root of

U

(53)

dk0 kd = sdg( )p0 pdg( )s0 = sdg( )dg( )s0 = sdg( )|{z}
s0 s dg( )s0
I

= uu0 = u2 .

(62)

Hence, from the non singularity of k we obtain the solution to (52) as follows
z

(59)-(60)

du

(62)

dk0 = d((dk0 kd) 2 )

dk0 .

(63)

Notice that if k is symmetric with positive eigenvalues, then k is the Riccati root of k0 k. Hence
k = k0
d=I

) z = ((k0 k) 2 )

16

1 0

k = (k)

k = I.

(64)