Gauge Invariance, Geometry and Arbitrage: Samuel E. Vázquez
Gauge Invariance, Geometry and Arbitrage: Samuel E. Vázquez
Gauge Invariance, Geometry and Arbitrage: Samuel E. Vázquez
Simone Farinelli
Kraus Partner Investment Solutions AG, Militärstrasse 76, 8004 Zurich,
Switzerland; email: [email protected]
We identify the most general measure of arbitrage for any market model governed
by Ito processes, and, on that basis, we develop dynamic arbitrage strategies.
It is shown that our arbitrage measure is invariant under changes of numeraire
and equivalent probability measure. Moreover, such a measure has a geometrical
interpretation as a gauge connection. The connection has zero curvature if and
only if there is no arbitrage. We prove an extension of the martingale pricing
theorem in the case of arbitrage. In our case, the present value of any traded asset
is given by the expectation of future cashflows discounted by a line integral of the
gauge connection. We develop simple dynamic strategies to measure arbitrage
using both simulated and real market data. We find that, within our limited data
sample, the market is efficient at time horizons of one day or longer. However, we
provide strong evidence for nonzero arbitrage in high-frequency intraday data.
Such events seem to have a decay time of the order of one minute.
1 INTRODUCTION
The no-arbitrage principle is the cornerstone of modern financial mathematics. Put
simply, an arbitrage opportunity allows an agent to make a risk-free profit with zero
or negative net investment (see Delbaen and Schachermayer (2008)). Under the no-
arbitrage assumption we can assign, in a complete market, a unique price to the
derivative of any traded assets using the replicating portfolio method (see Musiela
and Rutkowski (2007) or Cvitanic and Zapatero (2004)). The no-arbitrage principle
We would like to extend our gratitude to Eric Weinstein, Pia Malaney, Lee Smolin, Bernd Schroers,
Mike Brown, Simone Severini, Jiri Hoogland, Jim Herriot and Bruce Sawhill for many discussions
and ideas that influenced the results of this paper. The opinions expressed in this paper are the
authors’ own and do not represent those of Capital Fund Management S.A. or of Kraus Partner
Investment Solutions AG. The risk control principles presented are not necessarily used by Capital
Fund Management S.A. or by Kraus Partner Investment Solutions AG. This document does not
provide a comprehensive description of concepts and methods used in risk control at Capital Fund
Management S.A. or at Kraus Partner Investment Solutions AG.
23
24 S. E. Vázquez and S. Farinelli
Now suppose that we make an infinitesimal change in the path of the particle ıx .s/,
keeping the boundary conditions fixed ıx .0/ D ıx .1/ D 0. The interaction
changes by:
Z XZ 1
ı A D ıx .s/xP .s/F
; 0
opportunity. Our main assumption is that the prices of all financial instruments can
be described by Ito processes. Moreover, we ignore transaction costs.2
The organization and main results of the paper are as follows. In Section 2 we
present the class of models that we use in the rest of the paper. They are very gen-
eral and include the case of stocks, bonds and commodities, and more complicated
derivative products. We decompose the dynamics of these models in terms of their
gauge-transformation properties with respect to Equation (1.1). We identify the gauge
invariants and show that they represent an obstruction to the existence of a martingale
probability measure. We conclude Section 2 with an example with three assets, and
we derive a modified nonlinear Black–Scholes equation with arbitrage.
In Section 3 we give a geometrical interpretation to the gauge-invariant quantities
defined in Section 2. Our main goal is to identify the stochastic gauge invariants of
Section 2 with the curvature of a gauge connection. We begin with a review of the
Malaney–Weinstein connection (Malaney (1996) and Weinstein (2006)), which is
done in the context of differentiable economic paths. In Section 3.1 we generalize the
Malaney–Weinstein construction to stochastic processes and prove an asset pricing
theorem. The main result of this section is that the present value of any self-financing
portfolio of traded assets is given by the conditional expectation of future cashflows,
discounted by a line integral of the Malaney–Weinstein gauge connection. We show
how the value functions of different portfolios replicating the same contingent claim
are related to the arbitrage curvature. Readers who are interested only in the arbitrage
measure and the detection techniques can skip Section 3.
In Section 4 we develop a simple algorithm to measure the arbitrage curvature
using financial data. None of the results of Section 4 require an understanding of the
geometry of arbitrage. We explain the main sources of error in such measurement. The
algorithm is applicable to any financial instrument. In Section 5 we provide examples
with financial data of stock indexes and index futures and construct dynamic arbitrage
strategies. We find that, on long timescales, the market is very efficient. However, we
provide strong evidence for nonzero-curvature fluctuations at short timescales of the
order of one minute. We conclude in Section 6.
2 Note that, as pointed out in Shleifer and Vishny (1997), many possible arbitrage opportunities
disappear once we take into account market frictions such as transaction costs. Therefore, it is
important to keep in mind that, even when we measure a nontrivial curvature in the market, it does
not mean that it can always be exploited in a practical trading strategy.
of all such securities is described by Ito processes of the form (see Sondermann (2006)
and Shreve (2000)):
X
dX WD X ˛ dt C a dWa ; 8 2 M (2.1)
a
where the Wa are standard Brownian motions such that the Wa .t / Wa .0/ are inde-
pendently and normally distributed random variables with:
We make no assumptions about the number of Brownian terms, and hence the
completeness of the market. The set of Brownian motions fWa g represents all the
randomness in the market. Therefore, they induce a natural filtration F D .F t / t >0
for our probability space. The coefficients ˛ and a can also be stochastic processes
adapted3 to the filtration F . However, they are assumed to satisfy suitable conditions
to ensure the existence of the price processes X (see Lamberton and Lapeyre (2007)).
This class of models is very general and includes stocks, bonds, options, etc. Moreover,
the case of fat tails in the distribution of returns is also included, since this is known
to be generated by stochastic volatilities a .
Looking back at Equation (2.1), we can see that the tangent space dX has a natural
P
decomposition into the directions that contain all the randomness ( a a dWa ) and
those orthogonal to it. Therefore, we will make the following decomposition of the
drift term in Equation (2.1):
X X
˛ D ˛ C ˇ a O a C ˛ A JA (2.3)
a A2N
where N is the space spanned by basis vectors J A WD ŒJ0A ; : : : ; JNA1 such that:
X 9
JA JB D ı AB >
>
>
>
>
>
X >
=
JA D 0 (2.4)
>
>
X >
>
>
JA a D 0; 8a>
>
;
3In simple terms, a process p adapted to F means that it does not depend on future values of the
Brownian motion. In other words, p.t/ can only depend on fWa .s/g up to time s 6 t.
and: 9
1 X >
˛ WD ˛ >
>
N >
>
>
>
X >
>
>
A
˛ WD J ˛ >
A >
=
(2.5)
>
>
O a
WD >
a >
>
a
>
>
1 X >
>
a>
a
WD >
>
;
N
We will refer to N as the null space of the market. Note that this space is orthogonal
to all the randomness in the tangent space dX . However, we need to remember that
N might be trivial. The definition of ˇ a in Equation (2.3) is not unique if the vectors
O a WD ŒO 0a ; : : : ; O N
a
1 are linearly dependent. This is the case of, for example, an
incomplete market with more Brownian motions than traded securities. Moreover,
˛ A is unique up to rotations in the null space. As we will see, the quantities ˛ A are
the unique gauge-invariant measures of arbitrage. The two main goals of this paper
are to give a geometric interpretation to the parameters ˛ A , and to set up a procedure
to measure them using financial data.
Since prices are relative and only reflect an exchange rate between two products,
the units used to measure X are arbitrary. Therefore, the dynamics of the market
must be invariant under a change of measuring units. In mathematical finance, this
is known as a change of numeraire Musiela and Rutkowski (2007), and it can be
interpreted as a gauge transformation:
Our next task is to study the transformation properties of the different terms in
Equations (2.1) and (2.3). The following result follows.
Proposition 2.1 Consider a change of numeraire of the form X ! X , where
is a positive stochastic process adapted to the filtration F and:
X
a
d WD ı˛ dt C ı dWa
a
Then, the coefficients of the Ito processes, Equations (2.1) and (2.3), transform as:
X 9
˛ ! ˛ C ı˛ C a ı a >
>
>
=
a
a a
ˇ ! ˇ C ı a (2.7)
>
>
>
;
a ! a C ı a
and: X
˛!˛C a ıˇ a ; ˇ a ! ˇ a C ıˇ a (2.9)
a
In particular, it follows that O a , ˛ A and JA are invariant under such transformations.
Proof The result in Equation (2.7) above follows from a simple application of Ito’s
rule to the product X0 WD X :
where we defined: Z t
Wa .t/ D Wa .t / C ıˇ a .s/ ds
Note that both ˛ A and J A are unchanged by these gauge transformations. In particular,
suppose that: X
JA a D 0
So far we have taken the existence of the basis vectors J A for granted.A constructive
procedure to find such a basis, if nontrivial, is given by the following proposition.
Proposition 2.2 Let ˝ be the symmetric and real N N matrix with component:
X
˝ D a a ; where N D dim.M/
a
Moreover, define U as the matrix of all ones, eg, U D 1; 8; 2 M. Then, the
matrix G defined as:
1 1
GD˝ .U ˝ C ˝U / C 2 Tr.U ˝/U (2.12)
N N
P
is gauge invariant. Let NG be the null space of matrix G such that J D 0 for any
nontrivial J 2 NG . Then NG D N . In particular, the space NG is spanned by the
orthonormal zero modes of G that are orthogonal to the vector J D .1; 1; : : : ; 1/ .
Proof First we need to prove that the space of vectors J such that J a D 0; 8a,
is in one-to-one correspondence with the zero modes of ˝: ˝J D 0. Obviously, if
J a D 0, it follows that J is also a zero mode of ˝. To prove the converse, suppose
that ˝J D 0, but J a D a , where a ¤ 0 for at least one value of a. Then:
X
0 D J ˝J D .a /2
a
So far we have talked about the full set of securities of the market. However, it
is clear that the decomposition in Equation (2.3) can be done for any subset of the
market. That is, suppose we observe a subset of the prices Xi , i 2 S M. Moreover,
suppose that, within this subset, we can still find some zero modes J A obeying:
X X
JiA ia D 0; 8a; JiA D 0
i i
A
We can then easily lift these vectors to the full set M by taking JM D .J A ; 0/. This
represents a particular choice of basis in the null space N . By observing a subsector of
the market, we will only have access to some of the components of ˛ A . For notational
convenience, in the following we will not distinguish between the full market and a
subset of it.
In a next step we want to link ˛ A with the no-arbitrage condition. But what do
we mean exactly by “no-arbitrage condition”? As a matter of fact there exist two
similar conditions, the no-arbitrage condition and the no-free-lunch-with-vanishing-
risk condition (NFLVR) (see Delbaen and Schachermayer (2008)), which, in discrete
time, are equivalent. In continuous time NFLVR is the stronger condition and is
equivalent to the existence of a martingale measure for the (discounted) asset prices
(see Delbaen and Schachermayer (2008, Chapter 9.4)).
Under the NFLVR assumption (see Delbaen and Schachermayer (2008) and Hunt
and Kennedy (2004)), it is always possible to find a common positive discount factor
and an equivalent probability measure P P such that the discounted prices X
are martingales .t/X .t/ D Et Œ.T /X .T /, where t 6 T . This is known as
the martingale representation theorem (see Sondermann (2006) and Shreve (2000)).
In our language, this means that there is a gauge transformation mapping X to
P -martingales. In other words, if there is no arbitrage, price processes are gauge
equivalent to P -martingales for some probability measure P . The result of the
RT
for some adapted process a . The reason is that the stochastic integral t a dWa is
a martingale:
Z T
Et a
dWa D 0
t
2.1 An example
Consider the case of three assets X , D 0; 1; 2, where X0 is a savings account and
X1 , X2 are some other risky assets. All prices are measured in the same common
units. We will assume only one Brownian motion. Therefore, the dynamics of the
prices are described by:
)
dX0 D rX0 dt
(2.14)
dXi D Xi Œ˛i dt C i dW ; i D 1; 2
For later convenience, we assume that the interest rate r is deterministic. In order
to carry out the decomposition in Equation (2.3) we need to find a basis for the null
space N . In this case, since there is only one Brownian motion and two risky assets,
there will be only one null direction. To calculate it, we start by identifying the ˝
matrix:
0 1
0 0 0
B C
˝ D @0 12 1 2 A (2.15)
0 1 2 22
We can now construct the G matrix using Equation (2.12). The explicit form of G is
not very illuminating. The unnormalized eigenvectors of G are found to be:
0 C 1
0 1 0 1 1 2
31 31
2 1 C B 1 22 C
B 1 C 2 C B 1 C 2 C B C
B C B C B C
V1 D B C; V2 D B C; V3 D B 2 21 C
@ 0 A @ 1 A B C
@ 1 22 A
0 0
1
(2.16)
where: 9
GV1 D 0 >
=
GV2 D 0 (2.17)
>
;
2 2 2
GV3 D 3 .1 C 2 1 2 /V3
In order to find a basis for the null space N defined in Equation (2.4), we need to
project V1 or V2 into the space orthogonal to the vector J 0 D .1; 1; : : : ; 1/ . To do
this, we define the projection matrix:
PU WD 13 U; PU2 D PU (2.18)
where U is the 33 all-ones matrix. Note that 1PU projects into the space orthogonal
to J 0 . Our choice for the normalized null vector is then:
0 1
1 2
.1 PU /V1 1 B C
J Dp Dp p @ 2 A (2.19)
Œ.1 PU /V1 .1 PU /V1 2 2
2 1 C 2 1 2 1
where we identify:
9
@t V @X V @2 V >
>
˛2 D C ˛1 X C 12 12 X 2 X =
V V V (2.25)
@X V >
>
2 D 1 X ;
V
Comparing Equation (2.25) with Equation (2.23), we find that:
2 21
˛2 D r C ˇ2 C ˛Q p p
2 12 C 22 1 2
@t V @X V @2 V
D C ˛1 X C 12 12 X 2 X (2.26)
V V V
where, from Equation (2.22):
22 1
˛1 D r C ˇ1 C ˛Q p p (2.27)
2 12 C 22 1 2
Therefore, after some algebra, Equation (2.26) becomes a modified nonlinear Black–
Scholes partial differential equation:
@ t V C rX@X V C 12 12 X 2 @X
2
V
1=2
p X @X V X @X V
C 2˛Q 1 C 1 r V D 0 (2.28)
V V
Note that, for ˛Q D 0, this reduces to the familiar Black–Scholes equation. The non-
linear Black–Scholes equation is a special case of the more general pricing theorem
presented in Section 3.
It is important to remember that the arbitrage parameter ˛Q in Equation (2.28) can
in general depend on time and the stock price. Therefore, in principle, almost any
deformation of the option price is possible. It follows that Equation (2.28) can be
solved only if the arbitrage dynamics are known. For example, consider the case
where we set:
2
1 X 2 @X V
˛Q WD 3=2
.Q 12 12 / (2.29)
2 V Œ1 C .X @X V =V /..X @X V =V / 1/1=2
for some constant Q 12 . Then, the option price obeys the usual Black–Scholes equation
but with the “wrong” volatility:
@ t V C r.X@X V V / C 12 Q 12 X 2 @X
2
V D0 (2.30)
A ! A C d (3.2)
This is the analog of the transformation rule of the vector potential in electrodynamics.
A self-financing portfolio can be seen as being parallel transported with the con-
nection A as:
rP V D .d A/V jP D 0 (3.3)
where rP is the covariant derivative along the trajectory . The solution to this equa-
tion is simply: Z
V .T /
D exp A WD D (3.4)
V .t /
Note that the curvature is invariant under a gauge transformation, as d.ACd/ D dA.
In the approximation where economic agents are price takers, the price trajectory
X.t / is given exogenously, and we are only allowed to make changes in the portfolio
nominals . In other words, we can write dX D XP dt in Equation (3.5). We can
then restrict the curvature to the submanifold corresponding to the .t; / coordinates.
The induced curvature in this submanifold is given by:
X P
1 X XP
R D P X X
d ^ dt
. X /2 ; X X
X
WD R;t d ^ dt (3.6)
In this case, the path dependency of the Divisa index, Equation (3.4), can be written
as:
XZ T
ı log D D ds R;t .s/ı .s/ (3.7)
t
The zero-curvature condition implies that the prices of all securities evolve by the
same common inflation factor.
The relation between curvature and arbitrage goes as follows. Suppose that the
prices obey the zero-curvature condition given above. It follows that, for any self-
financing portfolio, we have:
Z
V .T / D V .t / exp A
Z T
D V .t / exp ˛.s/ ds (3.8)
t
while V .t / D 0. In other words, we have made wealth out of nothing. In the next
section we show how this construction carries over to the stochastic case.
Then, if there exists a change of measure satisfying the Novikov condition, then there
exists a (nonunique) equivalent probability measure P P under which the price
processes obey:
X X
dX D X ˛ C ˛ A JA dt C a dWa (3.11)
A a
Moreover, the present value of V .t / given some final payoff V .T /, T > t , is given by:
Z
V .t/ D Et V .T / exp (3.12)
Finally, the path dependency of the present value of the portfolio, with fixed final
payoff, is parameterized by:
XZ T Z
ıV .t/ D ds E t V .T / exp ı .s/R;t .s/
(3.14)
t
where R;t are the components of the curvature two-form defined in the reduced base
space .t; /:
R D d
1 X
D P 2
˛ A X X .JA JA / d ^ dt
. X /
;;A2N
X
WD R;t d ^ dt (3.15)
where: P P
˛ X a a X
aD P ; b D P (3.17)
X X
Now consider the combination V 0 WD V , where we take (see Equation (2.1)):
X X
a a a
d D a C b ˇ dt ˇ dWa (3.18)
a a
1X a 2 X
d log D .ˇ / dt ˇ a dWa (3.21)
2
a a
P
where is defined in Equation (3.13), with ˛ WD ˛ a ˇ a a .
Now consider making a change of probability measure such that:
Z t
Wa WD Wa ˇ a .s/ ds
It is easy to see that, under P , the price processes will obey Equation (3.11) of the
theorem. Moreover, the Radon–Nykodým derivative is given by:
Z T XZ T
dP 1X a 2 a
D exp .ˇ .s// ds C ˇ dWa (3.22)
dP 2
a t a t
The last result of the theorem, Equation (3.14), follows simply by making a small
change in the portfolio nominals, and keeping the boundary conditions on V fixed.
Note that the curvature of is zero if and only if ˛ A D 0, which implies, together
with the Novikov condition, the no-arbitrage condition. Moreover, the probability
measure P might not be unique, as the choice of ˇ a in general is not. This also
implies that ˛ is not unique in general.
A special case of a self-financing portfolio is a portfolio containing just one base
asset.
Corollary 3.2 Under the same assumptions as Theorem 3.1, for all assets in the
market model 2 M:
Z T X
X .t/ D E t X .T / exp
˛ C ˛ J dt 0
A A
(3.26)
t A
In Section 2.1 we derived a modified Black–Scholes equation for the case of three
assets. Now we can use the result of Corollary 3.2 to prove a generalization of such
an equation. Consider the following vector of assets:
n
X X X
A
@ t ˚i C ˛ C ˛ JjA Xj @j ˚i ˛ C
˛ A A
JnCi ˚i
j D1 A A
n
1 X
C ˝j k Xj Xk @j @k ˚i D 0 (3.29)
2
j;kD1
and: X
˛ D r ˛ A J0A (3.31)
A
0 D @ t V C .˛ C ˛J
Q 1 /X@X V .˛ C ˛J
Q 2 /V C 12 12 X 2 @X
2
V
D @ t V C .r C ˛.J Q 2 J0 //V C 12 12 X 2 @X
Q 1 J0 //X @X V .r C ˛.J 2
V
D @ t V C r.X@XV V / C 12 12 X 2 @X
2
V
p p
C 2˛V Q 1 C X @X log V .X @X log V 1/ (3.36)
4For notational simplicity, we will still use N for the null space of the particular market subsector
under study. However, it is important to keep in mind that this is not the null space of the full market.
X JA dX
˛A D (4.2)
X dt
The vectors J A must be calculated using an estimate for the quadratic variation:
dhlog X ; log X i
˝ WD
dt
and the results of Proposition 2.2. We introduce the notation:
X
A2 WD .˛ A /2
A
A2 > 0 (4.3)
N Z
X t X
0 .t/ WD i .s/ dXi .s/ C J0A .t /˛ A .t / (4.4)
iD1 0 A
X J A .t /˛ A .t /
i
i .t/ WD ; i D 1; : : : ; N (4.5)
Xi .t /
A
N
X
.d X C dh ; X i/ D 0 (4.7)
D0
where “d” denotes the Ito differential (see Chapter 4.1.2 in Lamberton and Lapeyre
(2007)). This is proved by the following computation:
N
X
.d X C dh ; X i/
D0
N
X
D d0 X0 C dh0 ; X0 i C .di Xi C dhi ; Xi i/
iD1
N
XX X N
X X
˛ A
JiA dXi ˛ A JiA
D C d.˛ A J0A / C Xi d
Xi Xi
A iD1 A iD1 A
XN X A A
˛ Ji
C d ; Xi
Xi
iD1 A
X X N
D d ˛A JA
A D0
D0 (4.8)
Since the self-financing condition is fulfilled, the portfolio value can be computed as:
N
X
V .t / D .t /X .t /
D0
Z t X N
X
A
D i .s/ dXi .s/ C ˛ JA
0 A D0
„ ƒ‚ …
D0
Z tX N
X JiA .s/
D ˛ A .s/ dXi .s/
0 Xi .s/
A iD1
Z tX N
X
A
JA .s/
D ˛ .s/ dX .s/
0 D0
X .s/
A
Z tX
D ˛ A .s/2 ds
0 A
Z t
D A2 .s/ ds (4.9)
0
Since the arbitrage curvatures are positive, we see that V .0/ D 0 and V .t / > 0 for
all times t 2 Œ0; T . The proof is complete.
There is no continuous time trading in the markets, and we can only do measure-
ments in discrete time. Moreover, our estimate of ˝ will always include errors. This
means that we will always have a noise term on the right-hand side of Equation (4.1).
The goal of this section is to explain the basic steps used to measure arbitrage cur-
vature, and to understand the major sources of error in such measurements. A key
aspect of our algorithm is that we test directly for the gauge invariance of the arbi-
trage signal. This allows us to check the robustness of our estimators. We find that
the gauge invariance of the arbitrage signal, as predicted by the stochastic models, is
indeed obeyed with good accuracy in the real market.
GO JO A D A JO A ; .JO A / JO B D ı AB ; A; B D 0; 1; : : : ; N 1 (4.11)
0 D 0 6 1 6 2 6 6 N 1 (4.12)
Our estimate for the basis of N will be to chose the first k eigenvectors with the
smallest eigenvalues: JO A , A D 1; : : : ; k < N 1. In doing this, we are assuming
that dim.N / D k.
Once we have calculated JO A , we can compute our estimate of ˛ A in discrete time:
X JOA .t /
˛O A .t C ıt / D ŒX .t C ıt / X .t / (4.14)
ıtX .t /
Note that JO A .t/ is constructed with information up to time t only. This estimate is
consistent with the nonanticipating nature of Ito integrals. The time step ıt is, of
course, arbitrary. Our estimate for A2 now becomes:
k
X k
X
AO 2 .t C ıt / D A
Œ˛O .t / C 2
˛O A .t /Œ˛O A .t C ıt / ˛O A .t / (4.15)
AD1 AD1
In the limit of short timescales, and if there is nontrivial arbitrage, we expect that this
estimator will converge to the true signal:
X
AO 2 .t C ıt / D A2 .t/ C ˛ A d˛ A D A2 .t / C O.ıt /; ıt ! 0 (4.16)
A
The convergence in Equation (4.16) is only valid if, in the limit ıt ! 0, we have:
E t Œ˛O A .t Cıt / ˛O A .t/ D O.ıt /; cov t Œ˛O A .t Cıt /; ˛O B .t Cıt / D O.ıt / (4.17)
Therefore, we expect that, if there is nontrivial arbitrage in the market, the estimator
(4.15) will give us a positive signal on average. Since the timescale is arbitrary, it is
convenient to set ıt D 1 henceforth.
There are several candidates for an estimator for ˝. The “right” choice of ˝O should
reflect our beliefs about the true dynamics of the asset values. Here we will simply
take the empirical estimator for covariance of the time series of log returns for a
window of length L. More precisely, our data consists of a number of time series for
the prices X , D 0; : : : ; N 1, in certain units, say US dollars.5 Our estimator
reads:
L1
O 1 X X .t i / X .t i/
˝ .t/ D log log
L X .t i 1/ X .t i 1/
iD0
L1
1 X X .t i/ X .t j /
2 log log (4.18)
L X .t i 1/ X .t j 1/
i;j D0
For more sophisticated estimators, see Hardle et al (2008) and Zhang (2006). We are
now in a position to summarize the most basic algorithm to detect arbitrage.
4.1.1 Algorithm
(1) Starting with the time series for X , D 0; : : : ; N 1, in an interval Œt; t L,
we estimate ˝O .t/ using Equation (4.18).
(2) We then calculate the GO matrix using Equation (4.10), and its orthonormal
eigenspace. The eigenvectors will be labeled as JO A , A D 0; 1; : : : ; N 1,
in order of increasing eigenvalues 0 D 0 6 1 6 6 N 1 . Moreover,
JO 0 / .1; 1; : : : ; 1/ .
(3) Given a guess for the dimension of the null space k D dim.N /, we take as its
O JO A , A D 1; : : : ; k.
basis the following eigenvectors of G:
(5) Roll the time window by one step, and repeat steps (1)–(4). Once we have
more than one estimate for ˛O A , we can calculate our final arbitrage estimator
AO 2 from Equation (4.15).
(6) In order to explicitly check for gauge invariance, we repeat steps (1)–(5), using
each asset X as a numeraire. For example, if we want to use X1 as a numeraire,
we divide all elements of the time series by the corresponding element of X1 , eg,
X .s/ ! X .s/=X1 .s/; 8 and s 2 Œt; t L. Then we repeat steps (1)–(5)
with the new time series. Note that this is a nontrivial transformation in the data
and, in practice, we will get different estimates for ˛O A .
Before discussing the results of Algorithm 4.1.1, we need to understand what the
main sources of error are in our signal. This is done in the next subsection.
to test for it is to make our calculations in different gauges and see how different the
answers are. We show examples of this in the following sections.
The second source of error in our signals comes from a gauge-invariant noise term.
In fact, we will see that this is the dominant noise contribution. In order to understand
this noise, it is convenient to discretize the Ito integral and write our estimate for ˛ A
as:
X JOA .t /
˛O A .t C 1/ D ŒX .t C 1/ X .t /
X .t /
X X
D JOA .t /JB .t /˛ B .t / C JOA .t /a ˇ a .t /
;B ;a
X
C JOA .t /a ŒWa .t C 1/ Wa .t /
;a
A
WD ˛trend .t / C "A .t C 1/ (4.19)
with E t Œ"A .t C 1/ D 0. Since JO A is only an estimate for the real J A , we have that
P A a
O A
J ¤ 0 in general. Therefore, our error in the estimate of J will induce an
extra noise term in the signal. Moreover, it will also induce some gauge dependency.
To see this, note that, under a change of numeraire, we have a ! a C ı a and
ˇ a ! ˇ a C ı a . It is then easy to check that the trend will transform according to:
X
A A
˛trend ! ˛trend C JOA a ı a
;a
However, note that the noise term is gauge invariant. In fact, one expects the term
P A a
O
J to be quite small. Moreover, since, in Algorithm 4.1.1, the gauge transfor-
mation is of the order ı a D O.a /, we expect the gauge dependence coming from
the trend to be negligible. We will see that, in real financial data, most of the signal
can be accounted for by the gauge-invariant noise term.
We are interested in estimating the size of the noise contribution. For that, we
compute the variance of the noise using information up to time t :
X
var t ˛O A .t/˛O A .t C 1/
A
X 2
D Et ˛O A .t/.˛O A .t C 1/ E t Œ˛O A .t C 1//
A
X
D ˛O .t/˛O B .t/.ŒJO A .t / G.t /JO B .t //
A
A;B
X X
D Œ˛O A .t/2 A .t / C ˛O A .t /˛O B .t /.ŒJO A .t / ıG.t /JO B .t // (4.22)
A A;B
where we remind the reader that k is our estimate for the dimension of N , and the
eigenvalues of GO have been ordered so that 1 6 2 6 6 k .
An interesting consequence of Equation (4.23) is that we can place a fundamental
bound on the size of the arbitrage curvature in order that it is detectable. We have:
p
A2 > var t ŒAO 2 H) A2 > k (4.24)
This means that, in order to have a chance to detect arbitrage, one needs to find
financial products whose time series are as correlated as possible, which implies a
very small value of k .
The third source of error is market microstructure noise. This effect is relevant in
high-frequency data, when the size of the price movements is comparable with the
bid–ask spread. In order to model this noise, it is convenient to set X0 WD 1 as our
numeraire. The standard way of simulating this noise is to introduce an additional
jump term
i .t/ to the log prices Xi , i D 1; : : : ; N . More precisely, the observed
price is XQ i and it is given by:
log XQ i .t / D log Xi .t / C
i .t / (4.25)
where Xi is the “true” Ito process, and, for simplicity, we assume that:
9
EŒXi
j D 0 >
=
EŒ
i D 0 (4.26)
>
;
EŒ
i
j WD
2 ıij
Moreover, the noise terms are uncorrelated between different times. We can then show
that our estimator will be contaminated by an amount:
4
2
EŒA D EŒA 2 C O
O 2 2
; ıt ! 0 (4.27)
ıt ıt 2
where: X 2
WD dim.N / E J0A
A
It can be shown that > 0. Therefore, we see that the microstructure noise leads a neg-
ative contribution to our estimation of A2 . The absolute value of such a contribution
diverges as we move toward higher frequencies (ıt ! 0).
One way of detecting the presence of microstructure noise is to note that:
Q
Xi .t C ıt / XQ i .t /
lim E log log D
2 < 0 (4.28)
ı t!0 XQ i .t / XQ i .t ıt /
In other words, the microstructure noise induces a negative correlation between sub-
sequent log returns. We find that this effect is quite pronounced for equity and futures
data. However, for stock indexes, the effect seems to be negligible. This is mainly due
to the fact that the microstructure noise “averages out” between all the stocks in the
index.
There is an extra source of error, which is intrinsic to Algorithm 4.1.1, but only if we
use a rolling window in our estimation of JO A . For example, suppose that we estimate
JO A .t/ and then roll the window and estimate JO A .t C 1/. Even if the matrices G.t
O / and
O C 1/ are near, JO A .t/ and JO A .t C 1/ can differ by a large orthogonal transform. It
G.t
can be just a sign flip, for example, since the eigenvalue equations are invariant under
JO A ! JO A . However, suppose two eigenvalues are near to each other, ie, 1 2 .
Then, any linear combination of JO 1 and JO 2 is also approximatively an eigenvector of
O In physics, this is known as the problem of degenerate perturbation theory (see,
G.
for example, Sakurai (1994)). More generally, we have that:
X
lim JO A .t C 1/ D C AB JO B .t / (4.29)
O
kG.t/ O C1/k!0
G.t B
C C D 1 (4.30)
The problem can be solved if we can determine C . If so, we can construct the “correct”
eigenvectors: X
JQ A .t C 1/ WD .C /AB JO B .t C 1/
B
so that:
lim JQ A .t C 1/ D JO A .t /
O
kı Gk!0
0.06
0.04
Log prices
0.02
0.02
0.04
0 50 100 150 200
Time
X0 W 1; 0a D 0; ˛0 D 0 (5.4)
which implies:
N 1 d N 1 2
1 XX a a XX A A
˛D ˇ i ˛ Ji (5.5)
N aD1
iD1 iD1 AD1
FIGURE 2 Result for the arbitrage detection algorithm applied to the simulated data in
Figure 1 on the facing page.
2×10−8
1.5×10−8
1×10−8
A2 5×10−9
−5×10−9
−1×10−8
100 120 140 160 180 200
Time
Here we assume (correctly) k D 2 null directions. The solid horizontal line at A2 108 is the correct value of A2 .
The gray solid line is the US dollar value of the signal.
In the previous section we discussed how the main source of error in our detection
technique can be related to the biggest eigenvalue k of the set f1 ; : : : ; k g. This
led to a gauge-invariant noise term. In this simulated sample data, we find that k
1021 , and so, using Equation (4.23), we find:
p p
varŒAO 2 108 1021 1015
Therefore, the noise term is very small in this case. The fluctuations seen in Figure 2 are
an artifact of this particular model. To understand them, we can expand Equation (5.1)
as:
X .t C 1/ X .t / X
D ˛ C a Ba .1/ C " C (5.6)
X .t/ a
where:
1 X
" D ˝ C a b Ba .1/Bb .1/ (5.7)
2
a;b
This extra noise term, " , is the reason for the gauge-invariant fluctuations in Figure 2.
The noise term vanishes if we integrate dX using an infinite partition of the time
interval, as it is assumed in Ito integrals. Of course, this is never possible in practice.
FIGURE 3 Result for the arbitrage detection algorithm applied to the simulated data in
Figure 1 on page 52.
2×10−8
1.5×10−8
1×10−8
A2 5×10−9
−5×10−9
−1×10−8
100 120 140 160 180 200
Time
Here we assume (incorrectly) k D 1 null directions and use a fixed window of 100 time steps. The error bars give
the range of values obtained using the different gauges. The solid dots are the mean of all results. The gray line near
to the time axis is the US dollar value of the signal.
Nevertheless, we see that, in this example, the extra noise is very small compared with
the arbitrage parameter A2 . In fact, we expect this noise to be very small in general
since it is of order varŒ" D O..a /4 /.
It is interesting to see what happens if we assume the wrong number of zero modes.
For example, in Figure 3 we show what happens if we take k D 1. We see that we
get a gauge dependent signal. Finally, in Figure 4 on the facing page we show what
happens if we assume k D 3. In this case, the biggest eigenvalue is k 108 .
As the figure shows, most of the fluctuations are coming from the gauge-invariant
noise described in the previous section. To see this we have plotted the expected noise
according to Equation (4.23):
p
noise˙ .t C 1/ D .A2 ˙ var t ŒAO 2 .t C 1//
v
u k
uX
D A ˙t
2
.˛O A .t //2 A .t / (5.9)
AD1
where A2 108 is the true value of the arbitrage (which is also the mean of the
signal). We see that this noise accounts for most of the fluctuations and it makes the
true arbitrage signal almost undetectable. The main point we would like to make here
is that the correct value of k can be estimated from the quality of the signal.
FIGURE 4 Result for the arbitrage detection algorithm applied to the simulated data in
Figure 1 on page 52.
2×10−7
0
A2
−2×10−7
−4×10−7
Time
Assuming k D 3 null directions (incorrectly), and using a fixed window of 100 time steps. The gray lines are the
noise terms noise˙ estimated according to Equation (5.9). The error bars give the range of values obtained using
the different gauges. The solid dots are the mean of all results. The dashed line is the US dollar value of the signal.
FIGURE 5 Result for the arbitrage detection algorithm applied to the simulated data in
Figure 1 on page 52, now including microstructure noise with variance
2 D 105 .
0.00002
−0.00002
A2
−0.00004
−0.00006
−0.00008
100 120 140 160 180 200
Time
We can now investigate the effect of the market microstructure noise discussed
in the previous section. In order to do this, we include additional white-noise terms
in the price processes of Equation (5.1) as described in the previous subsection (see
0.00003
0.00001
−0.00001
0 50 100 150 200
Time
The average of this signal is equal to 2 D 105 (dashed horizontal line), in agreement with the theoretical prediction.
(see also Equation (4.28)), where XQ i is the contaminated price. According to Equa-
tion (4.28), we should have EŒ
O 2 D
2 . This is precisely what we observe in Figure 6.
In the next subsection, we will see that such signals are typical of high-frequency
security prices.
FIGURE 7 Arbitrage detection algorithm applied to daily closing prices of three major US
indexes: DJA, IXIC and NYA.
0.00002
0.00001
A2 0
−0.00001
−0.00002
Here we show a sample of 500 data points. The gray lines are an estimate of the variance of the gauge-invariant
noise using Equation (5.11). The error bars give the range of values obtained using the different gauges. The solid
bars are the mean of all results. The black line is the US dollar value of the signal.
to July 16, 2009: a total of 1227 data points. The gauge-invariant matrix GO has been
estimated using a moving window of 500 days. We have found the following values
for the eigenvalues:
Therefore, it is reasonable to assume that the null space has only one dimension,
k D 1. The result of the arbitrage detection algorithm is shown in Figure 7. We
can see that the signal is indeed gauge invariant to a very high level of accuracy.
In Figure 7 we have also included an estimate for the gauge-invariant noise term
described in Section 4.2. In this case we have assumed that the average of the signal
is zero (ie, no-arbitrage), and so our estimate for the expected noise is:
p
noise˙ .t C 1/ D ˙ var t ŒAO 2 .t C 1/
v
u k
uX
D ˙t .˛O A .t //2 .t / A (5.11)
AD1
Looking at Figure 7, we see that the noise can explain most of the signal. Therefore,
we find that our results are consistent with A2 D 0, and hence no arbitrage. In Figure 8
on the next page we show a histogram of the different values of AO 2 . As pointed out
FIGURE 8 Histogram of different values of AO 2 obtained using daily market data for DJA,
IXIC and NYA.
300
250
200
150
100
50
0
−0.0001 −0.00005 0 0.00005 0.0001
A2
p
The signal-to-noise ratio is EŒAO 2 = varŒAO 2 0.0709.
above, the signal is consistent with A2 D 0 since the signal-to-noise ratio is very low:
EŒAO 2
p 0:0709
varŒAO 2
It is very instructive to look at the trading strategy exploiting the arbitrage discussed
in Proposition 4.1. In discrete time, the initial value of this portfolio is:
X
V .0/ D .0/X .0/ D 0
In Figure 9 on the facing page we show the value of this portfolio for the daily data
of the three US indexes. We include the integrated profit and loss of the indexes
themselves for comparison. We have multiplied the index signals by a numerical
FIGURE 9 Integrated profit and loss of the arbitrage portfolio (dashed black line) for the
daily data of the US stock indexes DJA, IXIC and NYA.
0.0005
∫ 0 A2 ds and indexes
0.0005
t
0.0010
500 600 700 800 900 1000 1100 1200
Time
We also show the integrated profit and loss of the indexes themselves.
FIGURE 10 Arbitrage detection algorithm applied to high-frequency data for three major
US indexes: DJA, IXIC and NYA.
3×10−8
2×10−8
1×10−8
A2 0
−1×10−8
−2×10−8
−3×10−8
Here we show a sample of 100 data points. The gray lines are an estimate of the variance of the gauge-invariant
noise using Equation (5.11). The error bars give the range of values obtained using the different gauges. The solid
dots are the mean of all results. The black line is the US dollar value of the signal.
factor so that it fits in the same picture. Therefore, the overall scale on the vertical
axis is irrelevant. We can see that, as expected, the performance of this portfolio is
very poor for such low-frequency data.
FIGURE 11 Histogram of AO 2 obtained using high-frequency data for DJA, IXIC and NYA.
200
150
100
50
0
−1×10−8 0 1×10−8 2×10−8
A2
p
The signal-to-noise ratio is EŒAO 2 = varŒAO 2 0.32.
Next we look at the same index set (DJA, IXIC, NYA), but now at short timescales.
As an example, we study high-frequency data obtained on July 28, 2009. The data
points are separated by 7–10 seconds. The data was collected using the “Financial-
Data” package of Mathematica. The gauge-invariant matrix GO has been estimated
using a moving window of 500 data points. We have also assumed one null direc-
tion (k D 1). A sample of the arbitrage detection algorithm is shown in Figure 10
on the preceding page. It is quite obvious from this figure that the signal has a very
significant positive skewness. In fact, a prominent feature of the signal is a series
of positive peaks. These transient events have a duration of the order of five to ten
time steps, which, for this data, is about one minute. The amplitude of the peaks is
quite significant compared with the noise. We argue that these peaks are precisely
temporary fluctuations with A2 ¤ 0, that is, nonzero-curvature events in the market.
To show that these are not isolated events, Figure 11 shows the histogram for the full
data sample. We can see significant positive skewness in the signal, compared with
the daily data (see Figure 7 on page 57). In fact, we find a significant signal-to-noise
ratio:
EŒAO 2
p 0:32
varŒAO 2
The integrated profit and loss of the arbitrage portfolio of Equation (5.12) are shown
in Figure 12 on the facing page. We can see a very good performance in comparison
with the daily data (see Figure 9 on the preceding page). Because of model risk, such
FIGURE 12 Integrated profit and loss of the arbitrage portfolio (dashed black line) for the
high-frequency data of the US stock indexes DJA, IXIC and NYA.
8×10−7
∫ 0 A2 ds and indexes 6×10−7
4×10−7
2×10−7
−2×10−7
t
−4×10−7
500 600 700 800 900 1000 1100 1200
Time
We also show the integrated profit and loss of the indexes themselves.
a portfolio can indeed have a finite probability of a loss on short timescales. However,
we see that, on longer timescales (integrated signal), the probability of a loss goes
to zero asymptotically as t ! 1. This is an example of a statistical arbitrage as
discussed in Pole (2007) and Bondarenko (2003).
We have also studied the effect of the microstructure noise on the high-frequency
signal. In particular, we have computed the estimate of the noise
O 2 defined in Equa-
tion (5.10). We have found that, for this particular data sample, the contribution from
such noise is very low:
EŒ
O 2
p 0:04
varŒ
O 2
However, if we look at traded assets such as stocks and futures, the effect becomes
quite significant.
FIGURE 13 Histogram of AO 2 obtained using high-frequency data for the index futures:
ESU09.CME, YMU09.CBT, NQU09.CME, SPU09.CME.
1000
800
600
400
200
0
−1×10−7 −5×10−8 0 5×10−8 1×10−7 1.5×10−7
2
A
p
The signal-to-noise ratio is EŒAO 2 = varŒAO 2 0.061. This figure illustrates the negative effects of the market
microstructure noise for the simplest trading strategy.
EŒAO 2
p 0:061
varŒAO 2
The integrated profit and loss of the simple portfolio of Equation (5.12) are shown in
Figure 14 on the facing page.
We have also calculated the effect of the microstructure noise, by computing the
estimate
O 2 defined in Equation (5.10). The effect for this data sample is about an
order of magnitude bigger than the previous example:
EŒ
O 2
p 0:15
varŒ
O 2
This noise is the main obstacle to a detection of A2 . Nevertheless, one can devise
more complicated detection methods that filter out the microstructure noise, whose
effect is minimized by trading at lower frequency but using higher-frequency data
to calculate the null space. As a matter of fact, when looking at the expression in
Equation (4.27) describing the contamination of the arbitrage estimator by the market
FIGURE 14 Integrated profit and loss of the arbitrage portfolio (dashed black line) of Equa-
tion (5.12) for the high-frequency data of the US index futures: ESU09.CME, YMU09.CBT,
NQU09.CME, SPU09.CME.
−5×10−7
−1×10−6
t
−1.5×10−6
500 1000 1500 2000 2500
Time
We also show the integrated profit and loss of the futures themselves. This figure illustrates the negative effects of
the market microstructure noise for the simplest trading strategy.
microstructure noise, we note that, for small values of the trading period ıt , the
longer the period, the smaller the contamination. So, if we keep high-frequency data
to compute the (approximate instantaneous) covariance matrix ˝, O while increasing
the trading period in such a way that the approximation relation (4.27) still holds, we
can filter out the market microstructure noise and obtain a clear arbitrage signal. In
Figure 15 on the next page we show the integrated profit and loss of this particular
strategy.
A detailed study of similar strategies requires a precise calibration of the trading
period: it must be sufficiently long that it clearly reduces the contamination but,
at the same time, sufficiently short that the arbitrage contamination relation is still
valid. Every asset universe requires its own calibration, which needs, as all statistical
parameters utilized here do, a regular update.
6 CONCLUSIONS
In this paper we have defined a general measure of arbitrage that is invariant under
changes of numeraire and equivalent probability measure. Our main assumption is
that all financial instruments can be described by Ito processes. This is not a very
strong assumption, as many complex financial models, including those reflecting the
non-Gaussian nature of stock returns, can be modeled this way. We showed that
the gauge-invariant arbitrage measure can be interpreted in terms of the curvature
FIGURE 15 Integrated profit and loss of an arbitrage portfolio (dashed black line) for
low-frequency trading and high-frequency data (for null space computational purposes) of
the US index futures: ESU09.CME, YMU09.CBT, NQU09.CME, SPU09.CME.
3×10−6
∫ 0 A2 ds and futures
2×10−6
1×10−6
0
t
1×10−6
500 1000 1500 2000 2500
Time
We also show the integrated profit and loss of the futures themselves.This particular strategy is designed to minimize
the effects of the microstructure noise.
period to reduce the market microstructure noise and by regularly recalibrating the
parameters.
From a scientific perspective, we believe that our findings represent a modest step
toward an understanding of nonequilibrium market dynamics. Gauge theories provide
the natural mathematical language to that aim, and arbitrage opportunities can be
interpreted as a nonzero-curvature fluctuation in an economy out of equilibrium. It is
interesting that most of our current economic and financial thinking relies so heavily
on the assumptions of general equilibrium theory.
There has been a growing consensus that we need a better understanding of the
nonequilibrium dynamics of the economy (see, for example, Farmer and Geanakoplos
(2008)). In particular, we would like to understand what the relaxation timescale is for
nonequilibrium fluctuations to disappear (if they do). Within the limited data sample
that we have shown in this paper, the relaxation time seems to be of the order of one
minute. However, this can be very different in other sectors of the market.
REFERENCES
Ackert, L. F., and Tian, Y. S. (1999). Efficiency in index options markets and trading in stock
baskets. Working Paper, Federal Reserve Bank of Atlanta.
Bleecker, D. (1981). Gauge Theory and Variational Principles. Addison-Wesley, Boston,
MA.
Bondarenko, O. (2003). Statistical arbitrage and securities prices. Review of Financial Stud-
ies 16(3), 875–919.
Cvitanic, J., and Zapatero, F. (2004). Introduction to the Economics and Mathematics of
Financial Markets. MIT Press, Cambridge, MA.
Delbaen, F., and Schachermayer, W. (2008). The Mathematics of Arbitrage. Springer.
Fama, E. F. (1998). Market efficiency, long-term returns and behavioral finance. Journal of
Financial Economics 49(3), 283–306.
Farinelli, S. (2011). Geometric arbitrage theory and market dynamics. Working Paper. URL:
http://ssrn.com/abstract=1113292.
Farmer, J. D., and Geanakoplos, J. (2008). The virtues and vices of equilibrium and the
future of financial economics. Working Paper. URL: http://arxiv.org/abs/0803.2996.
Gatev, E., Goetzmann, W. N., and Rouwenhorst, K. G. (2006). Pairs trading: performance
of a relative-value arbitrage rule. Review of Financial Studies 19(3), 797–827.
Hardle, W., Hautsch, N., and Pigorsch, U. (2008). Measuring and modeling risk using
high-frequency data. Discussion Paper, Humboldt University, Berlin. URL: http://sfb649.
wiwi.hu-berlin.de.
Hoogland, J., and Neumann, D. (1999a). Scale-invariance and contingent claim pricing.
Working Paper. URL: http://arXiv.org:cond-mat/9906048.
Hoogland, J., and Neumann, D. (1999b). Scaling invariance in finance II: path-dependent
contingent claims. Working Paper. URL: http://arXiv.org:cond-mat/9907185.
Hoogland, J., and Neumann, D. (2000). Asians and cash dividends: exploiting symmetries
in pricing theory. Working Paper. URL: http://arXiv.org:cond-mat/0006133.