Noname manuscript No.

(will be inserted by the editor)

Diversity-Weighted Portfolios with Negative

Parameter

Alexander Vervuurt · Ioannis Karatzas

Received: date / Accepted: date

Abstract We analyze a negative-parameter variant of the diversity-weighted

portfolio studied by Fernholz, Karatzas, and Kardaras (Finance Stoch 9(1):1–
27, 2005), which invests in each company a fraction of wealth inversely pro-
portional to the company’s market weight (the ratio of its capitalization to
that of the entire market). We show that this strategy outperforms the mar-
ket with probability one, under a non-degeneracy assumption on the volatility
structure and the assumption that the market weights admit a positive lower
bound. Several modifications of this portfolio, which outperform the market
under milder versions of this no-failure condition, are put forward, one of which
is rank-based. An empirical study suggests that such strategies as studied here
have indeed the potential to outperform the market and to be preferable in-
vestment opportunities, even under realistic proportional transaction costs.

Keywords Portfolios · Portfolio generating functions · Relative arbitrage ·

Stochastic Portfolio Theory · Diversity-weighted portfolios

JEL Classification G11

The authors are greatly indebted to Thibaut Lienart, Michael Monoyios, Vassilios Papa-
thanakos, Julen Rotaetxe, and Johannes Ruf for their very careful reading of the manuscript
and their many, insightful, and helpful suggestions. Alexander Vervuurt gratefully acknowl-
edges Ph.D. studentships from the Engineering and Physical Sciences Research Council,
Nomura and the Oxford-Man Institute of Quantitative Finance. Research supported in part
by the National Science Foundation under Grant NSF-DMS-14-05210.

Alexander Vervuurt ( )
Mathematical Institute, University of Oxford, Andrew Wiles Building, Radcliffe Observatory
Quarter, Woodstock Road, Oxford, OX2 6GG, UK
E-mail: [email protected]
Ioannis Karatzas
Department of Mathematics, Columbia University, New York, NY 10027, USA
E-mail: [email protected]
Intech Investment Management, One Palmer Square, Suite 441, Princeton, NJ 08542, USA
E-mail: [email protected]
2 Alexander Vervuurt, Ioannis Karatzas

1 Introduction

Stochastic Portfolio Theory offers a relatively novel approach to portfolio se-

lection in stock markets, aiming – among other things – to construct portfolios
which outperform an index, or benchmark portfolio, over a given time-horizon
with probability one, whenever this might be possible. The reader is referred
to R. Fernholz’s monograph (Fernholz, 2002), and to the more recent overview
by Fernholz and Karatzas (2009). In Fernholz et al. (2005) such outperform-
ing portfolios have been shown to exist over sufficiently long time-horizons,
in market models which satisfy certain assumptions; namely, those of weak
diversity and non-degeneracy.
One such investment strategy is the so-called diversity-weighted portfolio.
This re-calibrates the weights of the market portfolio, by raising them all to
some given power p ∈ (0, 1) and then re-normalizing; see Fernholz et al. (2005).
We study here a variant of this strategy, namely, a diversity-weighted portfolio
with negative parameter p < 0. This strategy invests in each company a pro-
portion of wealth that is inversely proportional to the ratio of the company’s
capitalization to the capitalization of the entire market (this ratio is called the
company’s market weight). As a result, the strategy sells the company’s stock
as its value increases, and buys it as its value decreases.

1.1 Preview

We set up the model and introduce the necessary definitions in Section 2.

Our first main result is that this negative-parameter diversity-weighted port-
folio also outperforms the market over long time-horizons, but now under a
no-failure condition. This postulates that all market weights are uniformly
bounded from below by a positive constant — see Section 3. The no-failure
condition is stronger than diversity.
A rank-based modification of this portfolio, which only invests in small-
capitalization stocks, is then shown in Section 4 to outperform the market,
also under this assumption of “no-failure”. For certain positive parameters
p ∈ (0, 1), this rank-based portfolio beats the market under a milder and
more realistic condition, namely that of limited-failure. This condition posits
a uniform lower bound on market weights only for the m < n largest firms by
capitalization, with n the total number of equities in the market.
We present two additional results in Section 5. The first shows that the
negative-parameter diversity-weighted portfolio outperforms its positive-para-
meter version under an additional boundedness assumption on the covariation
structure of the market; this posits that the eigenvalues of the covariance
matrix in the underlying Itô model are bounded from above uniformly in
time. Our second result studies a composition of these two diversity-weighted
portfolios and shows that, under the condition of diversity, this “mix” almost
surely performs better than the market over sufficiently long time-horizons.
Diversity-Weighted Portfolios with Negative Parameter 3

We carry out an empirical study of the constructed portfolios in Section 6,

demonstrating that our adjustments of the diversity-weighted portfolio would
have quite considerably outperformed the S&P 500 index, if implemented on
the index constituents over the 25 year period between January 1, 1990 and
December 31, 2014. In this study we incorporate realistic proportional trans-
action costs, delistings due to bankruptcies, mergers and acquisitions, and
distributions such as dividends. We compute the relative returns of our port-
folios over this period, as well as the Sharpe ratios relative to the market index.
We discuss the results and possible future work in Section 7.

2 Preliminaries

2.1 The Model

Our market model is the standard one of Stochastic Portfolio Theory (Fern-
holz, 2002), where the stock capitalizations are modeled as Itô processes.
Namely, the dynamics of the n stock capitalization processes Xi (·), i = 1, . . . , n
are given by
 d
X 
dXi (t) = Xi (t) bi (t) dt + σiν (t) dWν (t) , t ≥ 0, i = 1, . . . , n ; (1)
ν=1

here W1 (·), . . . , Wd (·) are independent standard Brownian motions with d ≥ n,

and Xi (0) > 0, i = 1, . . . , n are the initial capitalizations. We assume all
processes to be defined on a probability space (Ω, F, P), and adapted to a
filtration F = {F(t)}0≤t<∞ that satisfies the usual conditions and contains
the filtration generated by the “driving” Brownian motions.
The processes of rates of return bi (·), i = 1, . . . , n and of volatilities σ(·) =
(σiν (·))1≤i≤n,1≤ν≤d , are F-progressively measurable and assumed to satisfy the
integrability condition
n Z
X T  d
X 
|bi (t)| + (σiν (t))2 dt < ∞, P-a.s.; ∀ T ∈ (0, ∞), (2)
i=1 0 ν=1

as well as the non-degeneracy condition

∃ ε > 0 such that: ξ 0 σ(t)σ 0 (t)ξ ≥ ε||ξ||2 , ∀ ξ ∈ Rn , t ≥ 0 ; P-a.s. (ND)

2.2 Relative Arbitrage

We study investments in the equity market described by (1) using portfo-

lios. These are Rn -valued and F-progressively measurable processes π(·) =

0
π1 (·), · · · , πn (·) , where πi (t) stands for the proportion of wealth invested in
stock i at time t.
4 Alexander Vervuurt, Ioannis Karatzas

We restrict ourselves to long-only portfolios. These invest solely in the

stocks, namely
n
X
πi (t) ≥ 0, i = 1, . . . , n , and πi (t) = 1, ∀ t ≥ 0; (3)
i=1

in particular, there is no money market. The corresponding wealth process

V π (·) of an investor implementing π(·) is seen to evolve as follows (we normal-
ize the initial wealth to 1):
n
dV π (t) X dXi (t)
= πi (t) , V π (0) = 1. (4)
V π (t) i=1
Xi (t)

We shall measure performance, for the most part, with respect to the mar-
ket index. This is the wealth process V µ (·) that results from a buy-and-hold

0
portfolio, given by the vector process µ(·) = µ1 (·), · · · , µn (·) of market
weights
n
Xi (t) X
µi (t) := , i = 1, . . . , n , where X(t) := Xi (t). (5)
X(t) i=1

Definition 1 A relative arbitrage with respect to a portfolio ρ(·) over the

time-horizon [0, T ], for a real number T > 0 , is a portfolio π(·) such that

P V π (T ) ≥ V ρ (T ) = 1 P V π (T ) > V ρ (T ) > 0.



and (6)

An equivalent way to express this notion, is to say that the portfolio π(·)
outperforms portfolio ρ(·) over the time-horizon [0, T ].


We call this relative arbitrage strong, if in fact P V π (T ) > V ρ (T ) = 1 ;
and sometimes we express this by saying that the portfolio π(·) outperforms
the portfolio ρ(·) strongly over the time-horizon [0, T ]. t
u

We introduce the Rn×n -valued covariation process a(·) = σ(·)σ 0 (·) and,
writing ei for the ith unit vector in Rn , the relative covariances
µ
0

τij (t) := µ(t) − ei a(t) µ(t) − ej , 1 ≤ i, j ≤ n. (7)

Finally, we define the excess growth rate γπ∗ (·) of a portfolio π(·) as
n
X n 
1 X
γπ∗ (t) := πi (t)aii (t) − πi (t)aij (t)πj (t) . (8)
2 i=1 i,j=1

We shall use the reverse-order-statistics notation, defined recursively by

θ(1) (t) = max {θi (t)} (9)

1≤i≤n


θ(k) (t) = max {θ1 (t), . . . , θn (t)} \ {θ(1) (t), . . . , θ(k−1) (t)} , k = 2, . . . , n
Diversity-Weighted Portfolios with Negative Parameter 5

for any Rn -valued process θ(·). Here ties are resolved lexicographically, always
in favor of the lowest index i . Thus we have

θ(1) (t) ≥ θ(2) (t) ≥ . . . ≥ θ(n) (t). (10)

The non-degeneracy condition (ND) implies, on the strength of Lemma

3.4 in Fernholz and Karatzas (2009) (originally proved in the Appendix of
Fernholz et al. (2005)), that for any long-only portfolio π(·) we have with
probability one:
ε
γπ∗ (t) ≥


1 − π(1) (t) , ∀ t ≥ 0 . (11)
2

2.3 Functionally-Generated Portfolios

A particular class of portfolios, called functionally-generated portfolios, was

introduced and studied by Fernholz (1999).
Consider a function G ∈ C 2 (U, R+ ), where U is an open neighborhood of

∆n+ = x ∈ Rn : x1 + . . . + xn = 1, 0 < xi < 1, i = 1, . . . , n ,


(12)

and such that x 7→ xi Di log G(x) is bounded on ∆n+ for i = 1, . . . , n. Then G

is said to be the generating function of the portfolio π(·) given, for i = 1, . . . , n ,
by

 n
X 
πi (t) := Di log G(µ(t)) + 1 − µj (t)Dj log G(µ(t)) · µi (t). (13)
j=1

Here and throughout the paper, we write Di for the partial derivative with
respect to the ith variable, and D2ij for the second partial derivative with
respect to the ith and j th variables. Theorem 3.1 of Fernholz (1999) asserts
that the performance of the wealth process corresponding to π(·) relative to
the market, satisfies the decomposition (often referred to as “Fernholz’s master
equation”)

T
V π (T )
    Z
G(µ(T ))
log = log + g(t) dt , P-a.s. (14)
V µ (T ) G(µ(0)) 0

The quantity
n
−1 X µ
g(t) := D2 G(µ(t))µi (t)µj (t)τij (t) (15)
2G(µ(t)) i,j=1 ij

is called the drift process of the portfolio π(·).

6 Alexander Vervuurt, Ioannis Karatzas

2.4 Diversity-Weighted Portfolios

We consider now the diversity-weighted portfolio with parameter p ∈ R, defined

as in (4.4) of Fernholz et al. (2005):
(p) (µi (t))p
πi (t) := Pn p
, i = 1, . . . , n. (16)
j=1 (µj (t))

One condition that was shown to be sufficient for the existence of relative
arbitrage in the model (1), under the condition (ND), is that of diversity (see
Corollary 2.3.5 and Example 3.3.3 of Fernholz (2002)). This posits that no
single company’s capitalization can take up more than a certain proportion of
the entire market.
More formally, a market is said to be diverse over the time-horizon [0, T ],
for some real number T > 0, if


∃ δ ∈ (0, 1) such that: P µ(1) (t) < 1 − δ , ∀ t ∈ [0, T ] = 1 . (D)
Models of the form (1) with the property (D) were explicitly shown to exist in
Remark 6.2 of Fernholz et al. (2005). Other constructions have been proposed
by Osterrieder and Rheinländer (2006) and Sarantsev (2014). Fernholz et al.
(2005, see Eq. (4.5)) showed that the portfolio (16) outperforms the market
index µ(·) strongly, in markets satisfying (ND) and (D), for any p ∈ (0, 1), and
over time-horizons [0, T ] with
 
2 log n
T ∈ ,∞ . (17)
εδp
In Theorem 1 below, we show that a similar property holds for the diversity-
weighted portfolio with negative parameter p, in markets with the following
no-failure condition:


∃ ϕ ∈ (0, 1/n) such that P µ(n) (t) > ϕ, ∀ t ∈ [0, T ] = 1 . (NF)
We note that this condition implies diversity with parameter δ = (n − 1)ϕ .

3 Main Result

Here is the main result of this paper.

Theorem 1 In the market model (1), and under the assumptions (ND) and
(NF), the diversity-weighted portfolio π (p) (·) with parameter
 
log n
p ∈ ,0 (18)
log(nϕ)
is a strong arbitrage relative to the market µ(·) over the time-horizon [0, T ],
for any real number
−2n log(nϕ)
T >
. (19)
ε(1 − p) n − (nϕ)p
Diversity-Weighted Portfolios with Negative Parameter 7

Proof We apply the theory of functionally-generated portfolios as in Section

2.3. For any p 6= 0, it is checked that the portfolio (16) is generated by the
function
n
!1/p
X p
Gp : x 7−→ xi . (20)
i=1

We apply the method Pof Lagrange multipliers to maximize this function Gp

n
for p < 0 subject to i=1 xi = 1; this gives xi = 1/n, i = 1, . . . , n. We may
therefore write that, under (NF), the generating function admits the lower and
upper bounds
n  p n n
1−p
X 1 X
p
p X
n = ≤ µi (t) = Gp (µ(t) < ϕp = nϕp . (21)
i=1
n i=1 i=1

Hence, for any T ∈ (0, ∞), we have the lower bound

 
Gp (µ(T ))
log ≥ log(nϕ), (22)
Gp (µ(0))

a negative number. Using assumption (NF) and the lower bound from (21),
we obtain
(p) (µ(n) (t))p ϕp (nϕ)p
π(1) (t) = Pn p
< 1−p = < 1, (23)
i=1 (µi (t)) n n
where the last (strict) inequality follows from (18). In conjunction with (23),
the inequality (11) gives
Z T
ε T (nϕ)p
Z  
∗ (p)
ε
γπ(p) (t) dt ≥ 1 − π(1) (t) dt ≥ T 1− . (24)
0 2 0 2 n

Finally, straightforward computation shows that the drift process of the port-
folio (16), as defined in (15), is equal for all p ∈ R to

gp (·) = (1 − p) γπ∗(p) (·). (25)

We apply now (25), (22) and (24) to Fernholz’s master equation (14), and
conclude that the relative performance of π (p) (·) over [0, T ], with respect to
the market, is given by
(p) T
V π (T )
    Z
Gp (µ(T ))
log = log + (1 − p) γπ∗(p) (t) dt (26)
V µ (T ) Gp (µ(0)) 0
(nϕ)p
 
ε
> log(nϕ) + (1 − p) T 1 −
2 n
> 0, P-a.s.,

provided T satisfies (19). Hence the portfolio π (p) (·) outperforms the market
strongly over sufficiently long time-horizons [0, T ], as indicated in (19). t
u
8 Alexander Vervuurt, Ioannis Karatzas

4 Rank-Based Variants

Real markets typically do not satisfy the (NF) assumption (as stocks can
crash), so it is desirable to weaken this condition of Theorem 1. One possible
modification of (NF), is to posit that “no-failure” only holds for the m stocks
ranked highest by capitalization, for some fixed 2 < m < n, namely


∃ κ ∈ (0, 1/m) such that P µ(m) (t) > κ , ∀ t ∈ [0, T ] = 1 . (LF)

We call this condition (LF) the limited-failure condition, as it postulates

that no more than n − m companies will “go bankrupt” by time T . In this
context, bankruptcy of company i during the time-horizon [0, T ] is defined as
the event {∃ t ∈ [0, T ] such that µi (t) ≤ κ}. We also note that (LF) implies
diversity with parameter (m − 1)κ .

Remark 1 As was noted by V. Papathanakos (private communication), diver-

sity (D) with parameter δ ∈ (0, 1) in turn implies (LF) with parameters
 
1 1 − (m − 1)(1 − δ) 1
m= and κ = < , (27)
1−δ n − (m − 1) m

with bxc the largest integer less than or equal to x ∈ R. To see this, note that,
as long as (k − 1)(1 − δ) < 1, for 1 < k ≤ n, we have the implication

1 − (k − 1)(1 − δ)
µ(k−1) (t) < 1 − δ ⇒ µ(k) (t) > . (28)
n − (k − 1)

We identify the inequality on the right hand side of (28) as a version of (LF).
The largest integer for which this lower bound is positive, is k = b1/(1 − δ)c,
giving (27). t
u

4.1 A Diversity-Weighted Portfolio of Large Stocks

Under the assumption (LF), one can attempt to construct a relative arbitrage
using a variant (introduced in Example 4.2 of Fernholz (2001)) of the diversity-
weighted portfolio with parameter p = r ∈ R which only invests in the m =
m highest-ranked stocks (and thus naturally avoids investing in “crashing”
stocks), namely:
r

 P (µ(k) (t))

, k = 1, . . . , m,
m
πp#t (k) (t) = `=1 (µ(`) (t))
r
(29)

0, k = m + 1, . . . , n.


Here, pt (k) is the index of the stock ranked k th at time t (with ties re-
solved “lexicographically” once again, by choosing the lowest index), so that
µpt (k) (t) = µ(k) (t).
Diversity-Weighted Portfolios with Negative Parameter 9

We shall denote by Lk,k+1 (t) ≡ ΛΞk (t) the semimartingale local time ac-
cumulated at the origin, over the time-interval [0, t], by the continuous, non-
negative semimartingale


Ξk (·) := log µ(k) (·)/µ(k+1) (·) , k = 1, . . . , n − 1. (30)

Definition 2 Let us consider the local times at the origin of all continuous,
nonnegative semimartingales of the form


log µ(k) (·)/µ(k+r) (·) , k = 1, . . . , n − r , r ≥ 2 ,

and call them “higher-order collision local times”. t

u

We shall assume throughout this section the following:

Assumption 1 All higher-order collision local times vanish. t

u

The reader should consult Banner and Ghomrasni (2008) for general theory
on ranked semimartingales, as well as Ichiba et al. (2011) for sufficient condi-
tions that ensure the evanescence of such higher-order collision local times.
When attempting to construct a relative arbitrage with the portfolio (29),
one encounters a problem. To wit: an application of Theorem 3.1 of Fernholz
(2001) asserts that the following master equation holds for this rank-based
portfolio:
 π# ! Z T
Gr# (µ(T ))

V (T )
log = log + (1 − r) γπ∗# (t) dt (31)
V µ (T ) Gr# (µ(0)) 0
Z T π # (t)
pt (m)
− dLm,m+1 (t),
0 2

with Gr# the generating function of π # (·); compare with (26). Due to the
unbounded nature of semimartingale local time, the final term in (31) (referred
to as “leakage” by Fernholz (2001)) admits no obvious almost sure bound. Thus
there is no lower bound on the market-relative performance of π # (·) that holds
under reasonable conditions.
Since in real markets this local time term is typically small, we do still
expect this portfolio to have a good performance, and therefore include it in
our empirical study — see Section 6.

4.2 A Diversity-Weighted Portfolio of Small Stocks

Let us then fix 1 ≤ m < n, and define a small-stock diversity-weighted portfolio

by 
0,
 k = 1, . . . , m,
[ r
πpt (k) (t) = (µ(k) (t)) (32)
 Pn

r
, k = m + 1, . . . , n.
`=m+1 (µ(`) (t))
10 Alexander Vervuurt, Ioannis Karatzas

With this new dispensation, the local time term in (31) changes sign for π [ (·)
(see equation (38) below) and the problems mentioned in the previous subsec-
tion disappear. In fact, we can show that the new portfolio in (32) outperforms
the market under the assumptions (ND), (LF), for certain positive values of
the parameter r ; as well as under the assumptions (ND), (NF), when r is
within an appropriate range of negative values.

Proposition 1 We place ourselves in the context of the market model (1),

and under Assumption 1 and the condition (ND).
(i) If the condition (LF) also holds, the small-stock diversity-weighted portfolio
of (32) with parameter
 
log(2)
r∈ − ,1 and m=m−2 (33)
log((m + 1)κ)
is a strong relative arbitrage with respect to the market µ(·) over the time-
horizon [0, T ], provided that

4 log n−m



1 − r log (m + 1)κ
κ< and T > T+[ := 2
. (34)
2(m + 1) εr(1 − r) 2 − (m + 1)−r κ−r

(ii) If (NF) also holds, the small-stock diversity-weighted portfolio of (32) with
parameter  
log(n − m)
r∈ ,0 (35)
log((m + 1)ϕ)
is a strong arbitrage relative to the market µ(·) over the time-horizon [0, T ],
provided
−2(n − m) log((m + 1)ϕ)
T > T−[ :=
. (36)
ε(1 − r) n − m − (m + 1)r ϕr

Proof The portfolio (32) is generated by the function

n
!1/r
X
r
Gr (x) := x(`) , (37)
`=m+1

a rank-based variant of (20). In a manner analogous to (31), Theorem 3.1

of Fernholz (2001) implies the following master equation for our small-stock
portfolio:
 π[    Z T
V (T ) Gr (µ(T ))
log = log + (1 − r) γπ∗[ (t) dt (38)
V µ (T ) Gr (µ(0)) 0
Z T π[
pt (m) (t)
+ dLm,m+1 (t).
0 2
We note the change of sign for the last term, when juxtaposed with the ana-
logue (31) of this equation for the large-stock portfolio in (29).
Diversity-Weighted Portfolios with Negative Parameter 11

Case (i): With r ∈ − log(2)/ log((m + 1)κ), 1 as in (33), we have
1 log(2)
κ< =⇒ r>− > 0. (39)
2(m + 1) log((m + 1)κ)
Recalling (LF) and m = m − 2 , we have the following bounds
n
X
2κr < (µ(m−1) (t))r + (µ(m) (t))r + (µ(`) (t))r
`=m+1
n
X
r
= (µ(`) (t))r = Gr (µ(t)) (40)
`=m+1
n  r
X 1
−r
≤ = (n − m) m + 1
m+1
`=m+1

for the generating function Gr , and lead to the lower bound

   
Gr (µ(T ))
1 n−m
log > log (m + 1)κ − log . (41)
Gr (µ(0)) r 2
Note that, using the lower bound in (40) and µ(k) (t) ≤ 1/k , k = 1, . . . , n , we
obtain
[ (µ(m+1) (t))r (m + 1)−r
π(1) (t) = Pn < < 1, (42)
`=m+1 (µ(`) (t))
r 2κr
where the last bound follows from the inequalities in (39).
The non-degeneracy condition (ND) now gives
Z T
ε T
Z  
ε 1
γπ∗[ (t) dt ≥ [


1 − π(1) (t) dt > T 1− (43)
0 2 0 2 2(m + 1)r κr

in conjunction with (11) and (42). Whereas the nonnegativity of πp[ t (m) (t) ,
coupled with the nondecrease of the local time Lm,m+1 (·), shows that
Z T πp[ t (m) (t)
dLm,m+1 (t) ≥ 0 . (44)
0 2
We use this and apply (41) and (43) to the rank-based master equation (38),
to obtain
 π[   
V (T )
1 n−m
log > log (m + 1)κ − log (45)
V µ (T ) r 2
 
ε 1
+ (1 − r) T 1 −
2 2(m + 1)r κr
> 0, P-a.s.,

if and only if T > T+[ as defined in (34). We conclude that, under these
conditions, π [ (·) strongly outperforms the market over the horizon [0, T ].
12 Alexander Vervuurt, Ioannis Karatzas

Case (ii): With r ∈ log(n − m)/ log(nκ), 0 and under the condition (NF),
we have the following bounds

n
−r
X
r
(n − m) (m + 1) ≤ (µ(`) (t))r = Gr (µ(t)) < (n − m) ϕr . (46)
`=m+1

For the first inequality, we have used the simple fact that µ(`) (t) ≤ µ(m+1) (t) ≤
1/(m + 1) holds for ` = m + 1, · · · , n . Whereas, from (35), we have

[ (µ(n) (t))r ϕr
π(1) (t) = Pn < <1 (47)
`=m+1 (µ(`) (t))
r (n − m) (m + 1)−r

by analogy with (23). The inequalities (46) lead to the lower bound
 
Gr (µ(T ))

log > log (m + 1)ϕ ; (48)
Gr (µ(0))

and the non-degeneracy condition (ND), in conjunction with (11) and the
[
upper bound (47) on the largest portfolio weight π(1) (·), give

T T
(m + 1)r ϕr
Z Z  
ε εT
γπ∗[ (t) dt [


≥ 1− π(1) (t) dt > 1− . (49)
0 2 0 2 n−m

Again, we use (44) and apply (48) and (49) to the master equation (38), to
obtain
[
V π (T ) (m + 1)r ϕr
   

εT
log > log (m + 1)ϕ + (1 − r) 1 − (50)
V µ (T ) 2 n−m
> 0, P-a.s.,

provided T > T−[ as defined in (36).

Hence this small-stock, negative-parameter portfolio π [ (·), outperforms the
market over the horizon [0, T ]. t
u

5 Further Considerations

We present now some other results, the first of which shows that the diversity-
weighted portfolio (16) with p ∈ (0, 1) is outperformed by its negative-para-
meter counterpart under sufficient conditions. We also study a particular com-
bination of these two types of diversity-weighted portfolios, showing that it
outperforms a non-degenerate diverse market — see Proposition 3 further be-
low.
Diversity-Weighted Portfolios with Negative Parameter 13

5.1 Negative vs. Positive Parameter

In the following Proposition 2, besides (ND) we will impose also an upper

bound on the eigenvalues of the covariance matrix a(·), in the form of the
bounded variance assumption:

∃ K > 0 such that: ξ 0 σ(t)σ 0 (t)ξ ≤ K||ξ||2 , ∀ ξ ∈ Rn , t ≥ 0 ; P-a.s. (BV)

By Lemma 3.5 of Fernholz and Karatzas (2009), the bounded variance condi-
tion (BV) implies that for long-only portfolios π(·) we have the almost sure
inequality
γπ∗ (t) ≤ 2K 1 − π(1) (t) , ∀ t ≥ 0.


(51)

Proposition 2 Let us place ourselves in the market model (1), under the
conditions (ND), (BV) and (NF).
−
The diversity-weighted portfolio π (p ) (·) with negative parameter
 
log n
p− ∈ ,0
log(nϕ)
+
is then a strong arbitrage relative to the diversity-weighted portfolio π (p ) (·)
with positive parameter
−
ε(n − (nϕ)p )(1 − p− ) o
 n 
+
p ∈ max 0, 1 − ,1 , (52)
4K(n − 1)

over any horizon [0, T ] of length

−2 log(nϕ)
T > . (53)
C
Here the positive constant C is defined as
−
!
ε (nϕ)p 2K
1 − p− − (n − 1)(1 − p+ ) .


C := 1− (54)
2 n n

±
Proof To simplify notation a bit, we write π ± (·) and G± for π (p ) (·) and Gp± ,
respectively. Note that for p+ > 0 the inequalities in (21) reverse, i.e.,
+
p+ +
nϕp ≤ G+ (µ(t)) ≤ n1−p , (55)

which again gives the lower bound of (22) for p = p+ . Using (51) with the
+
observation π(1) (t) ≥ 1/n, we get that
Z T Z T
γπ∗+ (t) dt ≤ 2K +



1 − π(1) (t) dt ≤ 2KT 1 − (1/n) . (56)
0 0
14 Alexander Vervuurt, Ioannis Karatzas

Hence, recalling (25), we see that by virtue of (55) and (56) the master equa-
tion (14) for π(·) = π + (·) leads to the upper bound
+
V π (T )
 
≤ − log(nϕ) + 2(1 − p+ )KT 1 − (1/n) ,


log P-a.s. (57)
V µ (T )

Combining (26) and (57), we find that

− − +
V π (T ) V π (T ) V π (T )
     
log = log − log (58)
V π+ (T ) V µ (T ) V µ (T )
> 2 log(nϕ) + CT
> 0, P-a.s.,

provided that
CT > −2 log(nϕ) > 0 . (59)
Here, the constant C is given by (54). An easy calculation shows that (52)
implies C > 0, whereas the last inequality in (59) comes from ϕ < 1/n. It
follows that the first inequality in (59) is equivalent to the condition posited
in (53), and that in this case π − (·) outperforms π + (·) strongly over the time-
horizon [0, T ]. t
u
+
Proposition 2 shows that, as long as the diversity-weighted portfolio π (p ) (·)
is “sufficiently similar” to the market portfolio µ(·) (and thus “far enough”
−
from π (p ) (·)), it is outperformed strongly, over sufficiently long time-horizons,
−
by the diversity-weighted portfolio π (p ) (·) with negative parameter – pro-
vided, of course, that the aforementioned conditions on the volatility structure
and non-failure of stocks hold.

5.2 Mixing

Since the no-failure assumption (NF) does not hold in real markets, it is of
interest to find variants of (16) which exhibit similar performance, but require
weaker assumptions. The rank-based variant in Proposition 1 is one attempt
at this; Proposition 3 right below is another.

Proposition 3 Define the portfolio

b(t) = p(t)π + (t) + (1 − p(t))π − (t),

π (60)
±
with π ± (·) = π (p ) (·) diversity-weighted portfolios as defined in (16) with p+ ∈
(0, 1) and p− < 0, and the mixing proportion p(·) given by

Gp+ (µ(t))
p(t) = ∈ (0, 1). (61)
Gp+ (µ(t)) + Gp− (µ(t))
Diversity-Weighted Portfolios with Negative Parameter 15

In the market model (1), with assumptions (ND) and (D), the portfolio π b(·)
is a strong arbitrage relative to the market µ(·), over time-horizons [0, T ] with
− + −

2(1 + n(1/p )−1 ) log n(1/p )−1 + n(1/p )−1
T > T := . (62)
εδ(1 − p+ )

Proof The portfolio (60) is generated by the function G b = G+ + G− , with

± (p± )
G± = Gp± the generating functions of π (·) = π (·) as in (20); let g± (·) =
gp± (·) be their drift processes as in (25).
The drift process of the composite portfolio π
b(·) is then
n
−1 X µ
g(t) =
b D2ij G(µ(t))µ
b i (t)µj (t)τij (t)
2G(µ(t))
b
i,j=1
n
−1 X µ
= p(t) D2 G+ (µ(t))µi (t)µj (t)τij (t)
2G+ (µ(t)) i,j=1 ij
n
−1 X µ
+ (1 − p(t)) D2 G− (µ(t))µi (t)µj (t)τij (t)
2G− (µ(t)) i,j=1 ij
= p(t)g+ (t) + (1 − p(t))g− (t)
= (1 − p+ )p(t)γπ∗+ (t) + (1 − p− )(1 − p(t))γπ∗− (t) ; (63)

the final step follows from (25). We note that

γπ∗− (t) ≥ 0, (64)

as this holds for any long-only portfolio by Lemma 3.3 of Fernholz and Karatzas
(2009), which together with the observation that p(t) < 1 allows us to obtain
the bound
g(t) ≥ (1 − p+ )p(t)γπ∗+ (t).
b (65)
We note the simple bounds
n n
X X
p+
p+ +
1= µi (t) ≤ µi (t) = G+ (µ(t) ≤ n1−p , (66)
i=1 i=1

and that the lower bound in (21) holds even without (NF), so now
n
X 1/p−

p− −
0 ≤ µi (t) = G− (µ(t)) ≤ n(1/p )−1
. (67)
i=1

Using the lower bound from (66), and the upper bound from (67), we assert
that  −1
G− (µ(t)) 1
p(t) = 1 + ≥ > 0. (68)
G+ (µ(t)) 1 + n(1/p− )−1
One can easily see that in the positive-parameter diversity-weighted portfo-
lio, one’s proportion of wealth invested relative to the market portfolio is
16 Alexander Vervuurt, Ioannis Karatzas

diminished for large-cap stocks, and increased for small-capitalization stocks;

therefore
+ (µ(1) (t))p
π(1) (t) = Pn p
≤ µ(1) (t). (69)
i=1 (µi (t))

The non-degeneracy condition (ND) implies (11), which by (69) and the di-
versity assumption (D) leads to
ε
ε
ε
γπ∗+ (t) ≥ +
1 − π(1) (t) ≥ 1 − µ(1) (t) > δ . (70)
2 2 2

Finally, applying (65) to the master equation (14), and then using the
bounds (66), (67) and (68), (70), we conclude that
Z T
V πb (T )
   
G+ (µ(T )) + G− (µ(T ))
log ≥ log +
+ (1 − p ) p(t)γπ∗+ (t) dt
V µ (T ) G+ (µ(0)) + G− (µ(0)) 0
+ − ε δT
≥ − log n(1/p )−1 + n(1/p )−1 + 1 − p+



2 1 + n(1/p− )−1
> 0 , P-a.s. (71)

under the condition that T satisfies (62). t

u

We have shown that the long-only portfolio (60) outperforms strongly the
market over sufficiently long time-horizons, under the assumptions of non-
degeneracy and diversity. This is a property that the portfolio π + (·) also has
on its own, as proved in the Appendix of Fernholz et al. (2005).
We remark also that, since the “threshold” T of (62) is strictly decreasing
in p− , and the lower bound (71) is strictly increasing in p− for fixed T , our
result becomes stronger the closer the negative parameter p− gets to the
+
origin. As we take p− ↑ 0, we recover the well-known result that π
b(·) = π (p ) (·)
is a strong arbitrage relative to the market.

Remark 2 The statement of Proposition 3 can be strengthened, by weakening

the diversity condition (D) to that of weak diversity over the horizon [0, T ].
This notion is defined in equation (4.2) of Fernholz et al. (2005) as
!
Z T
1
∃ δ ∈ (0, 1) such that: P µ(1) (t) dt < 1 − δ = 1. (WD)
T 0

Under this weaker assumption, it is straightforward to see that the following

modification of (70) will hold, thus maintaining the validity of the proof:
Z T Z T Z T
ε ε ε
γπ∗+ (t) dt ≥ (1 − +
π(1) (t)) dt ≥ (1 − µ(1) (t)) dt > δ T. (72)
0 2 0 2 0 2

t
u
Diversity-Weighted Portfolios with Negative Parameter 17

5.3 Arbitrary Time Horizons

We should note here that, since both the (NF) and (LF) conditions imply
diversity, by Lemma 8.1 and Example 8.3 of Fernholz et al. (2005) it follows
that short-term relative arbitrage exists. That is, using the “mirror portfo-
lios” of Fernholz, Karatzas and Kardaras, one can construct a portfolio which
outperforms the market over any time-horizon [0, T ].

5.4 Two Portfolios with a Threshold

We mention briefly two other variants of the portfolio (16) with p < 0, which
exhibit similar behavior for mid- and upper-range market weights (in the sense
that they invest in stocks which decrease in value relative to the market, and
sell stock when a company’s market weight increases), but start selling stock
once a firm’s market weight falls below a certain threshold. The idea behind
this threshold is that it represents a value below which the investor fears
bankruptcy of the firm, and wishes to liquidate the position in the firm so as
to minimize losses. The two portfolios we propose have weights

µi (t)k e−µi (t)/θ

Γi (t) = Pn k −µj (t)/θ
, i = 1, . . . , n (73)
j=1 µj (t) e

and
µi (t)α (1 − µi (t))β
Bi (t) = Pn α β
, i = 1, . . . , n . (74)
j=1 µj (t) (1 − µj (t))

Here, k, θ and α, β are all positive constants; the threshold values below which
the portfolio weights become increasing functions of market weights are θk and
α/(α + β), respectively. So far, our results regarding these portfolios remain
restricted to empirical ones (see Section 6).

6 Empirical Results

We test the validity and applicability of our theoretical results by conducting

an empirical study of the performance of our portfolios using historical market
data. More precisely, we back-test by investing according to the portfolios
studied here in the n = 500 daily constituents of the S&P 500 index for
T = 6301 consecutive trading days between 1 January 1990 and 31 December
2014. We obtain these data from the Compustat and CRSP data sets (these
data sets are obtainable from http://wrds-web.wharton.upenn.edu/wrds/).
We have incorporated dividends and delistings such as mergers, acquisitions
and liquidations.
We aim to simulate as realistically as possible the wealth evolution of an
investor implementing our portfolios without any additional information, and
as such impose on each trade proportional transaction costs equal to 0.5%
18 Alexander Vervuurt, Ioannis Karatzas

of the total absolute value traded. We rebalance the portfolio using a simple
Total Variance criterion, that is, we only rebalance when the total variance
“distance”
n
X
TV(eπ (t), π(t)) = π ei (t) − πi (t)
ei (t) π (75)
i=1

between the target portfolio π(·) and the portfolio π

e(·) becomes larger than a
certain threshold, which we determine empirically. Here, π e(t) is the vector of
proportions of wealth invested in stocks obtained when the investor does not
rebalance at time t, that is
V π̄ (t − 1) Xi (t)
ei (t) = π̄i (t − 1)
π , i = 1, . . . , n, t = 2, . . . , T,
V π̄ (t) Xi (t − 1)

where π̄(t − 1) is the portfolio that was actually implemented in the previous
time step.
We use R to program our simulations — the code is available upon request.
We summarize our findings in Table 1, which displays the average annual rel-
ative returns in excess of the market (denoted “Market-RR”), and the Sharpe
ratios of our portfolios over the entire 25-year period that we use; the latter is
computed as
Rπ T
SharpeRatio(π(·)) = π
· . (76)
StdDev(R ) 25
Here, Rπ = {Rπ (t), t = 1, . . . , T } are the daily returns of the portfolio π, with
mean Rπ , namely
V π (t)
Rπ (t) = π (t
− 1, t = 2, . . . , T, (77)
V − 1)
and the sample standard deviation StdDev is defined as follows for any se-
quence of numbers x1 , . . . , xk ∈ R with average x:
k

2 1 X
StdDev(x1 , . . . , xk ) := (xi − x)2 . (78)
k − 1 i=1

Moreover, we compute the following measure of performance for each portfolio:

γπ 1
γ
eπ := · , (79)
StdDev(Rπ ) 25
that is, the total growth rate divided by the sample standard deviation. The
former is estimated as
 π 
V (T )
γπ = log . (80)
V π (1)
Figure 1 showcases the wealth processes corresponding to our portfolios.
Diversity-Weighted Portfolios with Negative Parameter 19

Table 1 Some measures of performance for the studied portfolios when traded between 1
January 1990 and 31 December 2014 on the constituents of the S&P 500 index, over which
the market had an average annual return of 9.7190%. We write π E (·) for the equally-weighted
portfolio πiE (·) = 1/n, i = 1, . . . , n. The parameter “TV threshold” determines over which
value of TV as in (75) we rebalance, which we determine in-sample by trial-and-error.

Portfolio TV Threshold Market-RR Sharpe Ratio γ

eπ
µ(·) N/A 0% 9.6012 8.1678
π E (·) 0.0005 2.2331% 11.181 9.7127
π (p) (·), p = 0.5 0.0022 1.6125% 12.187 10.949
π (p) (·), p = −0.5 0.0015 4.9655% 12.472 10.910
π # (·), r = −0.5, m = 470 0.0025 2.5778% 12.221 10.873
π [ (·), r = 0.5, m = 30 0.0001 0.9578% 10.274 8.8215
π [ (·), r = −0.5, m = 30 0.0100 1.7645% 10.702 9.2114
b(·), p+ = 0.5, p− = −0.5
π 0.0022 1.6125% 12.187 10.949
Γ (·), k = 0.65, θ = 1e-4 0.0020 3.6336% 10.821 9.0796
B(·), α = 1e-4, β = 2 0.0002 1.8701% 10.783 9.2920

From Table 1 and Figure 1, we can see that all portfolios outperform the
market by quite a margin (however, only after the year 2000), both in terms
of the market-relative return, as well as the Sharpe Ratio. We also note that
the simulated realizations of wealth processes are much more volatile than
that of the market; the portfolios seem to exploit market growth much better,
but also lose value more quickly when the market does poorly. Moreover, the
negative-parameter portfolios we study appear to perform better than their
positive-parameter versions over the period studied. The reader will notice
that the positive-parameter portfolio π (p) (·) with p = 0.5, and the mixing
portfolio π
b(·), give identical results, which is because the weight (61) of the
positive-parameter diversity-weighted portfolio is always very near 1 (namely
1 − p(t) ∼ 1e-11). Finally, we wish to stress that the above portfolio and
threshold parameters were optimized only heuristically, and therefore there
is considerable room for improvement (that is, through more systematic opti-
mization) — our empirical study is merely a demonstration of outperformance
of the market with the portfolios studied.

7 Discussion and Suggestions for Future Research

We wish to point out that all of our results require certain assumptions on
the volatility structure of the market, as well as on the behavior of the market
weights. The no-failure condition (NF) definitely does not hold in real mar-
kets, since typically some companies do crash. The weaker (LF) assumption
is an improvement on this, and can be argued to hold for m not too close
20 Alexander Vervuurt, Ioannis Karatzas

Fig. 1 The wealth processes corresponding to the portfolios in Table 1, namely: the market
portfolio µ(·); the diversity-weighted portfolios π (p) (·) with p = 0 (i.e. the equally-weighted
portfolio), p = 0.5, and p = −0.5; the rank-based diversity-weighted portfolios π # (·) with
r = −0.5, m = 470 and π [ (·) with r = 0.5, m = 30, and r = −0.5, m = 30; the mixed
b(·) from (60), with p+ = 0.5, p− = −0.5; and the portfolios (73) and (74), with
portfolio π
k = 0.65, θ = 1e-4 and α = 1e-4, β = 2, respectively.

to n — although, to the authors’ knowledge, there is no mechanism holding

this in place, as there is for the diversity assumption (D) which can be im-
posed by anti-trust regulation. In this regard, see Strong and Fouque (2011),
as well as the recent paper by Karatzas and Sarantsev (2014) in which the
number of companies is allowed to fluctuate – due to both splits and mergers
of companies.
It has been raised by V. Papathanakos (private communication) that from
a practitioner’s point of view, there are several restrictions on the implementa-
tion of negative-parameter diversity-weighted portfolios on a large scale. One
of these is that one typically demands that no position owns more than, say,
1% of the outstanding shares of a security; our portfolios strongly invest in
small stocks. Another constraint is related to liquidity: regular rebalancing in a
predictable way (which is implied by all functionally-generated portfolios) in-
curs very high transaction costs. It would be useful to model these phenomena
and their influence on portfolio performance.
More generally, it would be of great interest to develop a theory of transac-
tion costs in the framework of Stochastic Portfolio Theory, which would allow
one to improve upon, or even optimize over, rebalancing rules given a certain
Diversity-Weighted Portfolios with Negative Parameter 21

portfolio. A first attempt at this was made by (Fernholz, 2002, Section 6.3),
where R. Fernholz estimates the turnover in a diversity-weighted portfolio.
Another idea is to replace the almost sure assumptions (D), (LF), and
(NF), by probabilistic versions of these assumptions, where the corresponding
bounds on market weights hold with a probability close to but smaller than
1. It would be interesting to see whether probabilistic relative arbitrages could
be constructed, which have a certain “likelihood of outperforming” the market
— see Bayraktar et al. (2012) for a first study in this direction.
One could also incorporate additional information on expected drifts and
bankruptcies, to improve the simple portfolios set forth in this paper. The
approach by Pal and Wong (2013) might be applied to achieve this. Moreover,
it would be of interest to develop methods for finding the optimal relative
arbitrage within the class of functionally-generated portfolios; an attempt at
this was made in Pal and Wong (2014), and for more general strategies in
Fernholz and Karatzas (2010) and Fernholz and Karatzas (2011). Limitations
on the existence of relative arbitrages with respect to certain portfolios have
been established in the recent paper Wong (2015).

References

Adrian D. Banner and Raouf Ghomrasni. Local times of ranked continuous

semimartingales. Stochastic Processes and their Applications, 118(7):1244–
1253, 2008.
Erhan Bayraktar, Yu-Jui Huang, and Qingshuo Song. Outperforming the
market portfolio with a given probability. Ann. Appl. Probab., 22(4):1465–
1494, 2012.
Daniel Fernholz and Ioannis Karatzas. On optimal arbitrage. Ann. Appl.
Probab., 20(4):1179–1204, 2010.
Daniel Fernholz and Ioannis Karatzas. Optimal arbitrage under model uncer-
tainty. Ann. Appl. Probab., 21(6):2191–2225, 2011.
Robert Fernholz. Portfolio generating functions. Quantitative Analysis in
Financial Markets, River Edge, NJ. World Scientific, 1999.
Robert Fernholz. Equity portfolios generated by functions of ranked market
weights. Finance Stoch., 5:469–486, 2001.
Robert Fernholz. Stochastic Portfolio Theory. Springer, 2002.
Robert Fernholz and Ioannis Karatzas. Stochastic portfolio theory: A sur-
vey. In Alain Bensoussan and Qiang Zhang, editors, Handbook of Nu-
merical Analysis. Vol. XV. Special volume: mathematical modeling and nu-
merical methods in finance, volume 15 of Handbook of Numerical Analysis.
Elsevier/North-Holland, Amsterdam, 2009.
Robert Fernholz, Ioannis Karatzas, and Constantinos Kardaras. Diversity and
relative arbitrage in equity markets. Finance Stoch., 9(1):1–27, 2005.
Tomoyuki Ichiba, Vassilios Papathanakos, Adrian Banner, Ioannis Karatzas,
and Robert Fernholz. Hybrid Atlas models. Ann. Appl. Probab., 21(2):
609–644, 2011.
22 Alexander Vervuurt, Ioannis Karatzas

Ioannis Karatzas and Andrey Sarantsev. Diverse market models of com-

peting Brownian particles with splits and mergers. Preprint. Available at
arXiv:1404.0748, 2014.
Jörg R. Osterrieder and Thorsten Rheinländer. Arbitrage opportunities in
diverse markets via a non-equivalent measure change. Ann. Finance, 2:
287–301, 2006.
Soumik Pal and Ting-Kam Leonard Wong. Energy, entropy, and arbitrage.
Preprint. Available at arXiv:1308.5376, 2013.
Soumik Pal and Ting-Kam Leonard Wong. The geometry of relative arbitrage.
Preprint. Available at arXiv:1402.3720, 2014.
Andrey Sarantsev. On a class of diverse market models. Ann. Finance, 10(2):
291–314, 2014.
Winslow Strong and Jean-Pierre Fouque. Diversity and arbitrage in a regula-
tory breakup model. Ann. Finance, 7:349–374, 2011.
Ting-Kam Leonard Wong. Optimization of relative arbitrage. Ann. Finance,
page “to appear”, 2015.