Agnostic Risk Parity:

Taming Known and Unknown-Unknowns

Raphael Benichou, Yves Lempérière, Emmanuel Sérié,

Julien Kockelkoren, Philip Seager,
Jean-Philippe Bouchaud & Marc Potters
Capital Fund Management
arXiv:1610.08818v1 [q-fin.PM] 27 Oct 2016

23 rue de l’Université, 75007 Paris, France

Abstract
Markowitz’ celebrated optimal portfolio theory generally fails to de-
liver out-of-sample diversification. In this note, we propose a new port-
folio construction strategy based on symmetry arguments only, leading
to “Eigenrisk Parity” portfolios that achieve equal realized risk on all
the principal components of the covariance matrix. This holds true
for any other definition of uncorrelated factors. We then specialize our
general formula to the most agnostic case where the indicators of future
returns are assumed to be uncorrelated and of equal variance. This
“Agnostic Risk Parity” (AGP) portfolio minimizes unknown-unknown
risks generated by over-optimistic hedging of the different bets. AGP is
shown to fare quite well when applied to standard technical strategies
such as trend following.

Introduction
Diversification is the mantra of rational investment strategies. Harry Markowitz
proposed a mathematical incarnation of that mantra which is common lore
in the professional world. Unfortunately, the practical implementation of
Markowitz’ ideas is fraught with difficulties and yields very disappointing
results. This has been known for long, with many papers attempting to
identify its flaws and suggesting remedies [1, 2, 3, 4, 5, 6]. The most im-
portant problems are well understood: the optimally diversified Markowitz
portfolio often ends up – somewhat paradoxically – being very concentrated
on a few assets only, which inevitably leads to disastrous out-of-sample risks.
The optimal portfolio is also unstable in time and sensitive to small changes
in parameters and/or expected future gains. In the face of these difficulties,
two distinct branches of research have emerged.
The first one concerns the determination of the covariance matrix of the
N different assets eligible in the portfolio, for example all the stocks belong-
ing to a given index. This covariance matrix is specified by a large number of
entries (N × (N + 1)/2) for which only a limited amount of data is available
(N × T , where T is the length of the time series at one’s disposal). When

1
T is not extremely large compared to N, the empirically determined covari-
ance matrix is highly unreliable and leads to severe instabilities when used in
the Markowitz optimisation program. Recently, some powerful mathematical
tools have been proposed to optimally “clean” the empirical covariance ma-
trix, leading to a very significant improvement in the efficiency of Markowitz
diversification using the so-called “Rotationally Invariant Estimator” (RIE);
for a short review see [7] and refs. therein.
Another crucial step, of course, is to specify a list of expected returns for
each asset. These expected returns result either from quantitative signals
(such as trend following) or from other form of analysis (quantitative or
subjective). These signals are usually extremely noisy and unreliable, so
one should rather speak, as we will do below, of indicators, i.e. possibly
suboptimal and biased predictions of future returns.
Once all this is done, however, a time-worn but fundamental problem
remains [8, 9]. Even when sophisticated statistical tools can adequately deal
with risk, they cannot handle uncertainty, i.e. the intrinsic propensity of
financial markets to behave in a way that is not consistent with prior prob-
abilities. For example, although the future “true” covariance matrix is often
reasonably close to the cleaned (RIE) covariance matrix, correlations can also
suddenly shift to a new regime that was never observed in the past. This is
in fact worse for expected returns that are even more exposed to unknown-
unknowns than volatility or correlations. One therefore needs an extra layer
of control, beyond Markowitz’ optimisation, that acts as a safeguard against
statistically unexpected events.
This is what the second strand of research mentioned above attempts to
address. The idea is to add to the standard risk-return objective function
some extra penalty terms that enforce diversification, typically in the form of
generalized Herfindahl indices or entropy functions [10, 5, 11]. This has led
to important breakthroughs, such as the concept of “Maximally Diversified
Portfolios” (MDP) [3], or more recently, of “Principal Risk Parity Portfolios”
(PRP) (with several variations on this theme, see Refs. [12, 6, 13, 14, 15]).

Diversification and Isotropy

Although interesting, there is a hidden assumption in these penalty terms
that is far from neutral, which is the choice of the assets one considers as
“fundamental”, among which risk should be as diversified as possible in the
portfolio. These assets are chosen to be physical stocks for MDP’s or the
principal components of the correlation matrix in the case of PRP’s. In the
case of long-only portfolios and traditional asset management, the choice
of physical assets as the natural “basis” for portfolio construction might be
reasonable. But for – say – a portfolio of futures contracts with long and short
positions, any linear combination of these assets is a priori feasible (at least
within some overall leverage constraint). In mathematical terms, one can
“rotate” the natural asset basis into any a priori equivalent one. The point,
however, is that a maximally diversified portfolio in one basis can in fact
become maximally concentrated in another! Take for example a portfolio of

2
stocks with equal weightsP wi = 1/N on all N stocks. From the point of Pview
of the (neg)entropy S = i wi ln wi or of the Herfindahl index H = i wi2 ,
this is clearly optimal. But since the leading risk factor associated with the
correlation matrix is itself very close to an equi-weighted allocation on all
stocks, a rotation onto the principal component basis α leads to the worse
possible values for both the entropy and the Herfindahl index. In other
words, the very concept of maximal diversification is not invariant under a
redefinition of the assets considered as “fundamental”. Another vivid example
of the arbitrariness in the definition of fundamental assets is provided by the
interest rate curve or more generally, of contracts with different maturities.
Should one consider the physical contracts, or only one of them and all
associated calendar spreads?
Are there special directions in asset space that play a special role? Can
one unambiguously identify risk factors that are more fundamental than oth-
ers? This is an old problem in quantitative finance, with a long list of papers
attempting to identify these factors, in particular in the equity space. How-
ever, as recently reviewed by Roll [16], there is no consensus on this point.
If risk is associated to volatility (or variance), then the problem is in fact
completely degenerate or, using mathematical parlance, isotropic.
To make this clear, let us consider asset returns ri (i = 1, . . . , N) as
random variables with zero mean1 and (true) covariance matrix C, with
Cij = E[ri rj ]. One can then build N linear combinations of assets such
that their returns rbα are all uncorrelated and of unit variance. But this
choice is not unique: in fact, any further rotation2 in the space of assets
(i.e. an orthogonal combination of the synthetic assets returns rbα ) leads
to another set of uncorrelated, unit variance assets – see below. Among
this infinite choice of potential “factors”, is there any one that stands out,
that would justify applying a maximum diversification criterion among these
special assets? This is the path followed in, e.g. [17], where the further
notion of “Minimum Torsion Bets” was introduced.

Symmetries
We want to propose here a related, but different route based on symmetry
arguments, which fully exploits rotation and dilation invariance at the level
of indicators as well as at the level of returns. First, let us note that one can
rescale the returns of each asset i by an arbitrary factor without changing the
portfolio allocation problem. Investing 1 in a stock is the same as investing 21
on a fictitious “2-stock” contract, with twice the returns as the original stock.
So we can always choose to work with returns with unit variance, a choice
that we will make henceforth. In this case, the covariance matrix C is in fact
1
Here and below, we assume that any non zero average return (coming for example
from predictive signals) is small compared to the volatility, and can be neglected in our
discussion. Still, of course, this non zero average returns is what motivates the portfolio
construction to start with!
2
What we call rotations in this paper in fact includes both proper and improper rota-
tions, i.e. rotations plus inversions.

3
the correlation matrix between stocks. Now, the linear transformation
X

rbi = C−1/2 ij rj (1)
j

ri rbj ] = δij , i.e., to a set of uncorrelated assets. Here C−1/2 is

is such that E[b
defined as the positive-definite square root of C, namely:
X 1
C−1/2 = √ ua uTa , (2)
a
λa
where λa and ua are the eigenvalues and eigenvectors of C. This is the
meaning we will give throughout this paper to the square-root of a symmetric
matrix. As noted above, there is a large degeneracy in the construction
of the set of uncorrelated assets: any rotation of b r would do. A natural
choice at this point is to insist that the r̂i ’s are “as close as possible” to the
original normalized returns, so that the financial intuition about the resulting
synthetic assets is preserved (to wit, SPX d ≈ SPX). This is the case for the
r̂i ’s defined in Eq. (1) (see Appendix for a proof of this statement)3 .
The same construction can be applied for statistical indicators of future
returns that we call pi , i = 1, . . . , N. We insist that pi is not necessarily the
“true” expectation value of the future ri , but simply the best guess of the
investor based on his information/skill set/biases, etc. A standard example
considered below is a trend indicator based on a moving average of past
returns, but any quantitative indicator based on information or intuition
would do. These indicators fluctuate in time and are also characterized by
some covariance matrix Qij = E[pi pj ].4 This matrix is in general non trivial,
as one may systematically predict similar returns for two different assets i
and j, leading to Qij > 0. In any case, one can as above build N uncorrelated
linear combinations of indicators, given by:
X

pbi = Q−1/2 ij pj , (3)
j

with the above interpretation for Q−1/2 . The p̂i ’s are then all uncorrelated
and of unit variance, i.e. with the same scale of predictability in all direc-
tions, and “as close as possible” to the original pi ’s, which is again financially
meaningful. At this stage, any rotation in the space of (synthetic) assets also
rotates the new indicators pbi while keeping them all uncorrelated and of unit
variance. The portfolio construction problem has thus become completely
isotropic.

Rotationally Invariant Portfolios

How does all this help us to construct a truly agnostic Risk Parity portfolio,
with no reference to a specific set of assets deemed fundamental? A simple
3
The choice of normalization for the returns r is important here. Indeed working with
non-normalized returns would lead to a different result for b r. The choice we made is in
line with our isotropy assumption.
4
The usual case of static “long-only” indicators is special since the corresponding cor-
relation matrix is ill-defined. This will be the subject of a forthcoming work.

4
observation is that the realized gain G of a portfolio invested in the synthetic
asset α proportionally to pbα is given by:5
N
X N
X
G= pbα · rbα := Gα . (4)
α=1 α=1

This portfolio has several very desirable properties:

• The risk associated with each synthetic asset is the same: E[Gα2 ] =
p2α ]E[b
E[b rα2 ] = 1, provided one neglects E[Gα ] – see footnote 1.

• The gains associated with different synthetic assets are uncorrelated:

E[Gα Gβ ] = δα,β – see previous footnote 4.

• Most importantly, the total gain G is invariant under any further si-
multaneous rotation R of the assets and the indicators, as should be
for a scalar product:

X N
N X N
X
GR = Rα,β pbβ · Rα,γ rbγ
α=1 β=1 γ=1
N X
X N N
X
= pbβ · rbγ Rα,β Rα,γ
β=1 γ=1 α=1
N
X
= pbβ · rbβ ≡ G (5)
β=1

where we have used the fundamental property of rotation matrices

RRT = I.

The last property means that any arbitrary choice of uncorrelated, unit vari-
ance synthetic assets with its corresponding set of indicators leads to the very
same gain, so one does not need to decide on supposedly more fundamental
investment factors.
Why should one invest in the synthetic asset α proportionally to pbα ? On
the basis of symmetry arguments, this is the only rational choice. All invest-
ments directions are made statistically equivalent, any other choice would cor-
respond to an arbitrary breaking of isotropy. In the language of Markowitz
optimisation, this corresponds to the optimal portfolio of synthetic assets
when the expected future return of α is S pbα , where the expected Sharpe
ratio S is independent of α. Note that this in fact relies on the assumption
that E[bpα rbβ ] = Sδα,β , i.e. that at the level of uncorrelated factors, there
is no significant cross-prediction left. This is, we believe, a very plausible
assumption in practice – see below.
5
This implicitly assumes that the cross-correlations between the pbα and the rbβ6=α are
small, which is in fact an important hypothesis underlying our rotational symmetry prin-
ciple.

5
 Now, we need to convert the above isotropic risk portfolio invested in syn-
thetic assets into tradeable contracts. This simply follows from the definition
of rbα and pbα :
N X
X N



G = Q−1/2 αj
pj C−1/2 r
αi i
α=1 i,j=1
N N X
N
!
X X


= C −1/2
αi
Q−1/2 αj pj ri
i=1 α=1 j=1
N
X
:= πi ri , (6)
i=1

where the last equation defines the physical position πi in a asset i, which is
thus found to be:
N X
X N



πi = ω C−1/2 αi
Q−1/2 αj
pj , (7)
α=1 j=1

where ω is a constant that sets the overall risk of the portfolio, or, in vectorial
form (using the symmetry of C):

π = ω C−1/2 Q−1/2 p (8)

This is the central result of this paper, that we now comment and spe-
cialize to several situations. First, let us notice that the above portfolio
construction is such that the expected risk along any eigen-direction of C is
the same, hence the name “Eigenrisk Parity Portfolio” (ERP) – on this topic,
see also [12, 13]. Indeed, the expected risk along the ath principal component
is given by:

Ra = E[(π · va )2 ]λa , (9)

where λa is the ath eigenvalue of C and va the corresponding eigenvector.

Simple algebra then leads to:
 2
2 1
Ra = ω √ E[(va · Q−1/2 p)2 ]λa = ω 2 ∀a, (10)
λa
where we have used the fact that the expected covariance of the indicator
is Q. Note that although for any given day the allocation π points in a
specific direction and is thus “fully concentrated” in that sense, this direction
is expected to change over time – provided the indicators themselves are not
static. Isotropy is thus statistically restored on long enough time scales.

Agnostic Risk Parity

Now, the naive choice for the indicator covariance matrix Q should be pro-
portional to the return covariance matrix itself, i.e. Q ∝ C. In a stationary

6
world where the indicators would really statistically predict future returns,
i.e. pi = E[rifut. ], this assumption would be natural, at least when C is com-
puted on the time scale of the predicted returns, which is usually much longer
than a day. Interestingly, plugging Q ∝ C in Eq. (8) above precisely leads
to the standard Markowitz optimal portfolio: π = ωC−1 E[r fut. ]. However,
this is a highly over-optimistic view of the world that only deals with “known
unknowns”. Directional predictions are extremely uncertain, much more so
than risk predictions. In fact, directional predictions should not even be pos-
sible in an efficient market. If one insists that some signals may (weakly)
predict future returns, it is wiser not to assume any particular structure
on the correlation matrix of these indicators that any optimizer would use
to hedge some bets with other bets. The most agnostic choice, less prone
to unknown unknowns, is to choose Q = σp I, i.e. no reliable correlations
between the realized predictions, and the same amount of predictability (or
expected Sharpe ratio) on all assets. This leads to a very interesting portfolio
construction:

−1/2
π ∗ = ωCRIE p (11)

coined henceforth as “Agnostic Risk Parity” (ARP) because this specific

asset allocation allows one to precisely balance the risk between all the prin-
cipal components of the (cleaned) covariance matrix CRIE , in the worst-case
scenario where the realized correlations between indicators would completely
break down.
Note that there is no explicit optimisation used in this argument – rather,
we look for a rotationally invariant portfolio construction with the minimal
amount of information on the correlation structure of the indicators. The
risk distribution per eigen-mode for various portfolio allocations is drawn in
Fig. 1, when the realized covariance of the indicators is Q = σp I. Note
that, as is well known, the Markowitz optimisation scheme tends to over-
allocate on small eigen-modes, which can lead to significant out-of-sample
(bad) surprises [2], a bias that is corrected within the ARP framework.
Finally, one might believe that although uncertain, part of the return
correlations could be inherited by the indicators. A simple way to encode
this is to use for Q a shrinkage estimator, i.e. Q ∝ ϕCRIE + (1 − ϕ)I, where
ϕ ∈ [0, 1] allows one to smoothly interpolate between complete uncertainly
(ϕ = 0), corresponding to ARP, and the standard Markowitz prescription
(ϕ = 1).

Agnostic Trend Following

The previous discussion was rather formal. As an example, we consider
here the universal “Trend” indicator, based on a 1-year flat moving average
of past returns of a collection of 110 futures contracts (commodities, FX,
indices, bonds and interest rates) – see the discussion in [18]. We normalize
the returns of all futures and all the predictors to have unit variance. We then
use three different portfolio constructions: equal 1/N risk on each physical

7
Figure 1: Realized risk carried by different eigen-modes resulting from three
portfolio constructions: 1/N on futures contracts, Markowitz, and Agnostic
Risk Parity, all in the case where indicators are such that their realized
covariance is Q = σp I.

8
Figure 2: Profit & Loss (P&L) curves for universal trend following for four
portfolio constructions: 1/N on futures contracts, Markowitz with or without
a cleaned RIE correlation matrix, and Agnostic Risk Parity, again with RIE.
The universe here is composed of 110 contracts (commodities, FX, indices,
bonds and interest rates). The trend indicator is a 1-year flat moving average
of past returns. All P&L’s are rescaled such that their realized volatility is
the same.

asset, Markowitz optimal portfolio with either the raw empirical correlation
matrix or a cleaned version CRIE (using the RIE estimator detailed in [7],
and no future information) and the Agnostic Risk Parity, again using the
RIE estimator for CRIE . The P&L’s of the different portfolio since 1998 are
shown in Figure 2. While part of the improvement comes – as expected – from
using a cleaned correlation matrix, we see that Agnostic Risk Parity yields
the best result. Clearly the true correlation of predicted yearly returns Q is
nearly impossible to measure without centuries of data, hence motivating the
choice Q = σp I. We have observed similar results for other standard CTA
strategies.

9
Perspectives
In summary, we have offered a new perspective on portfolio allocation, which
avoids any explicit optimisation but rather takes the point of view of symme-
try. In a context where linear combinations of assets can easily be synthesized
in a portfolio whose risk is measured through volatility, the asset space can be
made fully “isotropic”, in the sense that no preferred directions (correspond-
ing to specific risk factors) can be identified. Therefore, in the absence of
extra information, portfolio construction should respect this symmetry. This
only requirement leads to a precise allocation formula, Eq. (8), that gen-
eralizes Markowitz’ prescription such as to take into account the expected
correlation between the predicted returns of each asset in the portfolio. We
have argued that the most agnostic choice, which is probably the most ro-
bust one out-of-sample, is to assume that these correlations are zero, i.e. that
one should refrain from trying to hedge different bets if there is no certainty
about the correlations between these bets. This leads to an Agnostic Parity
Portfolio that realizes an equal risk over all principal components of the co-
variance matrix. We found that such an allocation over-performs Markowitz’
portfolios when applied to classic technical (CTA) strategies, such as (uni-
versal) trend following. There are several routes that should be explored
further. For example, non-quadratic measures of risk, such as skewness or
kurtosis, would break rotational symmetry and possibly lead to meaningful
fundamental risk factors that should be maximally diversified (see e.g. [19]).
We leave this for future work.

We thank N. Bercot, J. Bun, R. Chicheportiche, S. Ciliberti, C. Deremble,

L. Duchayne, L. Laloux, A. Rej for many useful discussions on these issues.

Appendix
We consider random variables r of mean zero and unit variance, with a
correlation matrix given by C. We are looking for b r, a linear transformation
of r such that E[bri rbj ] = δij . Clearly b
r=C −1/2
r satisfy this property, and
any solution is of the form b r = RC −1/2
r where R is a rotation matrix. We
further demand that the following Mahalanobis distance d(R) is minimized:
 
d(R) = E (b r − r) · C−1 (b r − r) . (12)

Expanding the square, one readily sees that the only term that depends on
R is −2Tr[RC−1/2 ], which must be maximized. Since C−1/2 is a positive
definite matrix, it is immediate to show that the optimal solution is R = I,
i.e. C−1/2 is diagonal in the basis of C. Although natural in the present con-
text, we note that changing the Mahalanobis distance to any other positive
definite quadratic form where C−1 is replaced by any (matrix) function of C
– including the identity matrix – leads to the same result.

10
References
[1] F Black, R Litterman, Global portfolio optimization, Financial Analysts
Journal, 1992

[2] R Michaud, Efficient Asset Management: A Practical Guide to Stock

Portfolio Optimization and Asset Allocation. Cambridge, MA: Harvard
Business School Press, 1998.

[3] Y Choueifaty, Y Coignard, Towards Maximal Diversification, Journal

Portfolio Management, Fall, 40-51 (2008).

[4] T Roncalli, Introduction to Risk Parity and Budgeting, Chapman and

Hall, (2013)

[5] A Meucci, Managing diversification, Risk, 22, 74-79 (2009)

[6] R Deguest, L Martellini, and A Meucci, Risk parity and

beyond - from asset allocation to risk allocation decisions
http://ssrn.com/abstract=2355778. EDHEC-Risk Working Paper
(2013).

[7] J Bun, J P Bouchaud, M Potters, Cleaning correlation matrices,

Risk magazine (March 2016) & J Bun, J P Bouchaud, M Potters,
arXiv:1610.08104, to appear in Physics Reports, 2016.

[8] J M Keynes, The General Theory of Employment, Interest and Money,

1934

[9] N N Taleb, The Black Swan, Random House Trade (2010)

[10] J P Bouchaud, M Potters, J P Aguilar, Missing information and asset

allocation, arXiv preprint cond-mat/9707042 (1997)

[11] G Frahm, C Wiechers, On the diversification of portfolios of risky assets,

University of Cologne (2011)

[12] M H Partovi, M Caputo, Principal Portfolios: Recasting the Efficient

Frontier, Economics Bulletin, 7, 1-10 (2004)

[13] C Kind, Risk-Based Allocation of Principal Portfolios,

http://ssrn.com/abstract=2240842 (2012)

[14] H Lohre, U Neugebauer, C Zimmer, Diversified Risk Parity Strategies

for Equity Portfolio Selection, Journal of Investing, 21, 111-128, (2012)

[15] D H Bailey, M Lopez de Prado, Balanced Baskets: A New Approach

to Trading and Hedging Risks Journal of Investment Strategies, 1, 21-62
(2012)

[16] K Pukthuanthong, R Roll, A Protocol for Factor Identification,

http://ssrn.com/abstract=2517944 (2014)

11
[17] A Meucci, A Santangelo, R Deguest, Risk Budgeting and
Diversication Based on Optimized Uncorrelated Factors,
http://ssrn.com/abstract=2276632, Risk Magazine (November 2015).

[18] Y Lempérière, C Deremble, P Seager, M Potters and J P Bouchaud,

Two Centuries of Trend Following, Journal of Investing Strategies 3, 41-
61, (2014), and refs. therein.

[19] E Baitinger, A Dragoschy, A Topalovaz, Extending the Risk Par-

ity Approach to Higher Moments: Is There any Value-Added?
http://ssrn.com/abstract=2682630 (2015)

12