• Aucun résultat trouvé

Optimal portfolio liquidation with execution cost and risk

N/A
N/A
Protected

Academic year: 2021

Partager "Optimal portfolio liquidation with execution cost and risk"

Copied!
38
0
0

Texte intégral

(1)

HAL Id: hal-00394997

https://hal.archives-ouvertes.fr/hal-00394997

Submitted on 14 Jun 2009

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires

Optimal portfolio liquidation with execution cost and risk

Idris Kharroubi, Huyen Pham

To cite this version:

Idris Kharroubi, Huyen Pham. Optimal portfolio liquidation with execution cost and risk. SIAM Journal on Financial Mathematics, Society for Industrial and Applied Mathematics 2010, 1, pp.897- 931. �10.1137/09076372X�. �hal-00394997�

(2)

Optimal portfolio liquidation with exeution ost and risk

IdrisKHARROUBI

LaboratoiredeProbabiliteset

ModelesAleatoires

CNRS,UMR7599

UniversiteParis7,

andCREST,

e-mail: kharroubiensae.fr

Huy^en PHAM

LaboratoiredeProbabiliteset

ModelesAleatoires

CNRS,UMR7599

UniversiteParis7,

CREST,and

Institut UniversitairedeFrane

e-mail: phammath.jussieu.fr

June 14,2009

Abstrat

Westudy theoptimalportfolio liquidationproblemoveranite horizonin alimit

orderbook withbid-askspreadandtemporarymarketprieimpatpenalizing speedy

exeutiontrades. Weuseaontinuous-timemodelingframework,butin ontrastwith

previous related papers (see e.g. [24℄ and [25℄), we do not assume ontinuous-time

tradingstrategies. We onsider instead real trading that ourin disrete-time, and

thisisformulatedasanimpulseontrolproblemunderasolvenyonstraint,inluding

the lag variable traking the time interval between trades. A rst important result

of our paper is to showthat nearlyoptimal exeution strategies in this ontext lead

atuallytoanitenumberoftradingtimes,andthisholdstruewithoutassumingadho

anyxedtransationfee. Next,wederivethedynamiprogrammingquasi-variational

inequalitysatisedbythevaluefuntioninthesenseofonstrainedvisositysolutions.

We also introdue a family of value funtions onverging to our value funtion, and

whihisharaterizedastheuniqueonstrainedvisositysolutionsofanapproximation

ofourdynamiprogrammingequation. Thisonvergeneresultisusefulfornumerial

purpose,postponedinafurther study.

Keywords: Optimal portfolio liquidation, exeution trade, liquidity eets, order book,

impulseontrol, visositysolutions.

MSC Classiation (2000) : 93E20, 91B28, 60H30, 49L25.

WewouldliketothankBrunoBouhardforusefulomments.WealsothankpartiipantsattheIstanbul

workshoponMathematialFinaneinmay2009,forrelevantremarks.

(3)

Understanding trade exeution strategies is a key issue for nanial market pratitioners,

and has attrated a growing attention from theaademi researhers. An important pro-

blem faed by stok traders is how to liquidate large blok orders of shares. This is a

hallenge due to the following dilemma. By trading quikly, the investor is subjet to

higher osts due to market impat reeting the depth of the limit order book. Thus,

to minimizeprie impat, it is generally beneial to break up a large order into smaller

bloks. However, more gradual trading over time results in higher risks sine the asset

value an vary more during the investment horizon in an unertain environment. There

has been reentlya onsiderable interest in the literatureon suh liquidityeets, taking

into aount permanent and/or temporary prie impat, and problems of this type were

studiedbyBertsimasand Lo[7 ℄,AlmgrenandCriss[1 ℄,Bank andBaum[5 ℄,Cetin,Jarrow

and Protter [8 ℄, Obizhaeva and Wang [18 ℄, He and Mamayski [13 ℄, Shied an Shoneborn

[25℄,LyVath,Mnifand Pham[17 ℄,Rogers andSingh [24 ℄,andCetin,Sonerand Touzi [9℄,

to mentionsome of them.

There are essentially two popular formulation types for the optimal trading problem

in the literature: disrete-time versus ontinuous-time. In the disrete-time formulation,

we maydistinguishpapersonsidering thattradingtake plae atxed deterministitimes

(see [7℄),at exogenous randomdisretetimes given forexamplebythejumpsofa Poisson

proess (see [22 ℄, [6 ℄), or at disrete times deided optimally by the investor through an

impulseontrol formulation (see [13 ℄ and [17 ℄). In this lastase, one usuallyassumes the

existeneof a xedtransationost paidat eah tradinginorder to ensurethatstrategies

donotaumulateintimeandourreallyat disretepointsintime(seee.g. [15 ℄ or[19 ℄).

Theontinuous-timetrading formulation isnotrealistiinpratie,butisommonly used

(asin[8 ℄,[25 ℄or[24 ℄),duetothetratabilityandpowerfultheoryofthestohastialulus

typiallyillustratedbyIt^o'sformula. Inaperfetlyliquidmarketwithouttransationost

and market impat,ontinuous-time trading isoften justied by arguingthat it is a limit

approximationof disrete-timetradingwhenthetimestepgoestozero. However, onemay

questionthevalidityofsuhassertion inthepresene ofliquidityeets.

In this paper, we propose a ontinuous-time framework taking into aount the main

liquidity features and risk/ost tradeo of portfolio exeution: there is a bid-ask spread

in the limit order book, and temporary market prie impat penalizing rapid exeution

trades. However, in ontrast with previous related papers ([25℄ or [24 ℄), we do not as-

sumeontinuous-timetrading strategies. We onsider insteadreal trading thattake plae

in disrete-time, and without assuming ad ho any xed transation ost, in aordane

with the pratitioner literature. Moreover, a key issue in line of the banking regulation

andsolvenyonstraintsistodeneinaneonomially meaningfulwaytheportfolio value

of a position in stok at any time, and this is addressed in our modelling. These issues

areformulatedonvenientlythroughanimpulseontrolprobleminludingthelag variable

trakingthetimeintervalbetweentrades. Thus,weombinetheadvantagesofthestohas-

ti alulustehniques,and therealistimodelingof portfolio liquidation. Inthisontext,

we studytheoptimalportfolio liquidationproblemovera nitehorizon: theinvestorseeks

(4)

minal liquidation wealth, and undera natural eonomi solveny onstraint involving the

liquidationvalueof a portfolio.

A rst important result of our paper is to show that that nearly optimal exeution

strategies inthis modeling lead atually to a nite number of trading times. While most

models dealing with trading strategies via an impulse ontrol formulation assumed xed

transation ost in order to justify a posteriori the disrete-nature of trading times, we

prove here that disrete-time trading appear naturally as a onsequene of liquidity fea-

tures represented by temporary prie impat and bid-ask spread. Next, we derive the

dynamiprogrammingquasi-variationalinequality(QVI)satisedbythevaluefuntionin

thesenseof onstrainedvisosity solutionsinorder to handlestate onstraints. There are

some tehnial diÆulties related to the nonlinearity of the impulse transation funtion

induedbythemarketprieimpat,andthenon smoothnessof thesolvenyboundary. In

partiular,sinewedonotassumea xedtransationfee,whihpreludestheexisteneof

a stritsupersolutionto theQVI, wean notprovediretly aomparison priniple(hene

a uniqueness result) for the QVI. We then onsider two types of approximations by in-

troduing familiesof value funtions onverging to our original value funtion, and whih

are haraterized as uniqueonstrained visositysolutions to their dynamiprogramming

equations. Thisonvergene result isusefulfornumerialpurpose, postponedina further

study.

Theplanofthepaperisorganizedasfollows. Setion2presentsthedetailsofthemodel

andformulatestheliquidationproblem. InSetion3,weshowsomeinterestingeonomial

and mathematial properties of the model, in partiular the niteness of the number of

tradingstrategies underilliquidityosts. Setion4isdevotedto thedynamiprogramming

and visosityproperties ofthe value funtionto ourimpulseontrolproblem. We propose

in Setion 5 an approximation of the original problem by onsidering small xed tran-

sation fee. Finally, Setion 6 desribesanother approximation of the model with utility

penalization by smallost. As a onsequene, we obtain that ourinitial value funtionis

haraterized as the minimal onstrained visosity solution to its dynami programming

QVI.

2 The model and liquidation problem

We onsider a nanial market where an investor has to liquidate an initial position of

y > 0 shares of risky asset (or stok) by time T. He faes with the following risk/ost

tradeo: if he trades rapidly, this results in higher osts for quikly exeuted orders and

market prie impat; he an then split the order into several smaller bloks, but is then

exposedto theriskofpriedepreiation duringthetradinghorizon. Theseliquidityeets

reeived reently aonsiderableinterest startingwiththepapers byBertsimasand Lo [7℄,

and Almgrenand Criss [1 ℄ in a disrete-time framework, and further investigated among

others inObizhaeva and Wang [18℄, Shied an Shoneborn [25 ℄, or Rogers and Singh [24 ℄

in a ontinuous-time model. These papers assume ontinuous trading with instantaneous

trading rate induing prie impat. In a ontinuous time market framework, we propose

(5)

timethroughanimpulseontrolformulation,andwithatemporaryprieimpatdepending

on thetimeintervalbetweentrades, and inludinga bid-askspread.

Wepresentthedetailsofthemodel. Let(;F;P)beaprobabilityspaeequippedwith

altrationF =(F

t )

0tT

satisfyingtheusualonditions,andsupportingaonedimensional

Brownian motionW ona nitehorizon[0;T℄,T <1. We denotebyP =(P

t

) themarket

prieproess of therisky asset, byX

t

the amount of money(or ash holdings), byY

t the

numberof shares in thestokheld by the investorat time t, and by

t

the time interval

betweentimet andthe lasttrade before t. We setR

+

=(0;1)and R

=( 1;0).

Trading strategies. We assume that the investor an only trade disretely on [0;T℄.

Thisismodelledthroughanimpulseontrolstrategy=(

n

;

n )

n0 :

0

:::

n

:::T

are nondereasing stopping times representing the trading times of the investor and

n ,

n0,are F

n

measurable random variables valued in R and giving the number of stok

purhased if

n

0 or selled if

n

<0 at these times. We denote by A theset of trading

strategies. The sequene (

n

;

n

) may be a priori nite or innite. Notie also that we

do not assume a priorithat the sequene of trading times (

n

) is stritly inreasing. We

introduethelagvariable trakingthetime intervalbetweentrades:

t

= inf

t

n :

n

tg; t2[0;T℄;

whih evolvesaording to

t

= t

n

;

n

t<

n+1

;

n+1

=0; n0: (2.1)

Thedynamis of thenumberof sharesinvested instokisgiven by:

Y

t

= Y

n

;

n

t<

n+1

; Y

n+1

= Y

n+1 +

n+1

; n0: (2.2)

Cost of illiquidity. The market prie of the risky asset proess follows a geometri

Brownian motion:

dP

t

= P

t

(bdt+dW

t

); (2.3)

with onstants b and > 0. We do notonsider a permanent prie impat on the prie,

i.e. the lastingeet of large trader, butfous here on theeet of illiquidity,that is the

prie at whih an investorwill trade the asset. Suppose now that the investor deides at

timet to make an orderin stokshares ofsize e. If theurrent market prieisp, and the

timelag fromthelastorder is ,thenthepriehe atuallyget fortheorder eis:

Q(e;p;) = pf(e;); (2.4)

where f is atemporary prieimpat funtionfrom R [0;Tinto R

+

[f1g. We assume

thattheBorelianfuntionf satisesthefollowingliquidityandtransationostproperties:

(H1f) f(0;) =1,and f(:;) isnondereasing forall 2[0;T℄,

(H2f) (i) f(e;0) =0 fore<0,and (ii)f(e;0) =1 fore>0,

(H3f)

b

:= sup

(e;)2R

[0;T℄

f(e;) <1 and

a := inf

(e;)2R

+ [0;T℄

f(e;) >1.

(6)

= p, and a purhase (resp. a sale) of stok shares indues a ost (resp. gain) greater

(resp. smaller)thanthemarketprie,whihinreases(resp. dereases)withthesizeofthe

order. In other words, we have Q(e;p;) (resp. ) pfor e (resp. ) 0,and Q(:;p;)

is nondereasing. Condition (H2f) expresses the higher osts for immediay in trading:

indeed, the immediate market resilieny is limited, and the faster the investor wants to

liquidate(resp. purhase)theasset,thedeeperintothelimitorderbookhewillhaveto go,

andlower(resp. higher)willbetheprieforthesharesoftheassetsold(resp. bought),with

a zero (resp. innite)limitingprie forimmediate bloksale (resp. purhase). Condition

(H2f) also prevents theinvestor to pass orders at onseutive immediate times, whih is

thease inpratie. Instead ofimposinga xedarbitrary lagbetweenorders, we shallsee

that ondition (H2) implies that trading times are stritly inreasing. Condition (H3f)

aptures atransation ost eet: at time t, P

t

is themarket ormid-prie,

b P

t

is thebid

prie,

a P

t

is the ask prie, and (

a

b )P

t

is the bid-ask spead. We also assume some

regularity onditionson thetemporaryprieimpatfuntion:

(Hf) (i) f isontinuouson R

(0;T℄,

(ii)f isC 1

on R

[0;Tand x 7!

f

is boundedon R

[0;T℄.

Ausualform(seee.g. [16 ℄,[23 ℄,[2 ℄)oftemporaryprieimpatandtransationostfuntion

f,suggestedbyempirialstudiesis

f(e;) = e j

e

j

sgn(e)

a 1

e>0 +1

e=0 +

b 1

e<0

; (2.5)

with the onvention f(0;0) = 1. Here 0 <

b

< 1 <

a ,

a

b

is the bid-ask spread

parameter, > 0 is the temporary prie impat fator, and > 0 is the prie impat

exponent. In our illiquidity modelling, we fous on the ost of trading fast (that is the

temporary prie impat), and ignore as in Cetin, Jarrow and Protter [8℄ and Rogers and

Singh [24 ℄ thepermanent prieimpat of a largetrade. This lasteet ould be inluded

inour model, byassuming a jump of theprie proess at the trading date, dependingon

theorder size,see e.g. He and Mamayski [13 ℄ and LyVath,Mnif andPham [17℄.

Cashholdings. Weassumeazerorisk-freereturn,sothatthebankaountisonstant

betweentwo tradingtimes:

X

t

= X

n

;

n

t<

n+1

; n0: (2.6)

When adisretetrading Y

t

=

n+1

oursattime t=

n+1

,thisresultsina variationof

theashamount givenbyX

t := X

t X

t

= Y

t :Q(Y

t

;P

t

;

t

) dueto theilliquidity

eets. Inother words,we have

X

n+1

= X

n+1

n+1 Q(

n+1

;P

n+1

;

n+1 )

= X

n+1

n+1 P

n+1 f(

n+1

;

n+1

n

); n0: (2.7)

Notie that similarly as in the above ited papers dealing with ontinuous-time trading,

we do not assume xed transation fees to be paid at eah trading. They are pratially

insigniantwithrespetto theprieimpatandbid-askspread. We anthennotexlude

(7)

n+1

n+1 n

some n. However,notie thatunderondition(H2f), animmediatesaledoesnotinrease

the ash holdings, i.e. X

n+1

= X

n+1

= X

n

, while an immediate purhase leads to a

bankrupty,i.e. X

n+1

= 1.

Liquidation value and solveny onstraint. A key issue in portfolio liquidation is

to dene in an eonomially meaningful way what is the portfolio value of a positionon

ash and stoks. In our framework, we impose a no-short sale onstraint on the trading

strategies,i.e.

Y

t

0; 0tT;

whih isinlinewiththebankregulationfollowingthe nanialrisis,and we onsiderthe

liquidation funtionL(x;y;p;) representing the netwealth value thatan investor witha

ash amount x, would obtained byliquidating his stok position y 0 by a single blok

trade,whenthemarketprieispandgiventhetimelag fromthelasttrade. Itisdened

on RR

+ R

+

[0;Tby

L(x;y;p;) = x+ypf( y;);

and we imposetheliquidationonstrainton tradingstrategies:

L(X

t

;Y

t

;P

t

;

t

) 0; 0tT:

We have L(x;0;p;) = x, and under ondition(H2f)(ii), we notiethat L(x;y;p;0) = x

fory 0. Wenaturally introduetheliquidationsolveny region:

S = n

(z;)=(x;y;p;)2RR

+ R

+

[0;T: y>0 and L(z;)>0 o

:

We denote its boundaryand its losure by

S =

y S[

L

S and

S = S[S;

where

y S =

n

(z;)=(x;y;p;)2RR

+ R

+

[0;T: y=0 and x=L(z;)0 o

;

L S =

n

(z;)=(x;y;p;)2RR

+ R

+

[0;T: L(z;)=0 o

:

We also denote byD

0

theornerlineinS:

D

0

= f0gf0gR

+

[0;T =

y S\

L S:

Admissible trading strategies. Given (t;z;) 2 [0;T

S, we say that the impulse

ontrolstrategy =(

n

;

n )

n0

is admissible, denoted by 2 A(t;z;), if

0

=t ,

n

t, n 1,and theproessf(Z

s

;

s )=(X

s;

Y

s

;P

s

;

s

);tsTg solutionto (2.1 )-(2.2)-

(2.3)-(2.6)-(2.7), withan initialstate (Z

t

;

t

) =(z;) (andtheonventionthat (Z

t

;

t )

= (z;) if

1

> t), satises (Z

s

;

s

) 2 [0;T

S for all s 2 [t;T℄. As usual, to alleviate

notations,weomitted the dependeneof (Z ;)in (t;z;;), when there isno ambiguity.

(8)

th eta = 0 .1

th eta = 0 .5 y : s to c k s h a re s

y : s to c k s h a re s

y : s to c k s h a re s

y : s to c k s h a re s x:

cash

x:

cash x:

cash x:

cash D D

D D

Figure1: DomainS inthenonhathedzoneforxedp=1and evolving from1:5to 0:1.

Here

b

=0:9and f(e;)=

b exp(

e

) fore<0. Notie thatwhen goesto 0, thedomain

onverges to theopen orthant R

+ R

+ .

(9)

theta=1

x:cash amount

y: stock amount p: mid−price

Figure 2: Lower bound of the domain S for xed = 1. Here

b

= 0:9 and f(e;) =

b exp(

e

) fore<0. Notiethat whenp isxed, we obtaintheFigure1.

p=1

x:cash amount

y: stock amount theta: time−lag order

Figure 3: Lowerboundof thedomain S for xedp =1 with f(e;)=

b exp(

e

) fore<0

and

b

=0:9. Notiethat when is xed,we obtaintheFigure1.

Références

Documents relatifs

L’objectif est maintenant de mettre en place des écrans interactifs dans la station, pour que les gens puissent accéder à tous types d’informations et qui nous

Clearly, the induction of Gls24 by copper and the physical interaction of CopZ and Gls24 in vivo and in vitro strongly suggest a role of Gls24 in the defense against copper stress

It would indeed seem that enhanced information work in the refugee law area serves the dual purpose of improving the legal protection of refugees in the present situation at

By comparing the optimal execution strategy of a non-informed agent with the one of an informed agent, one can derive the additional gain due to the signal.. In order to obtain a

To integrate the fourth moment into the optimisation problem, two approaches were considered: a mean-kurtosis formulation equivalent to mean-variance but using kurtosis as the

In the present, we obtain and exercise optimal control formulations of conditioned problems involving higher moments of returns to evaluate the impact of conditioning information

We show how the same problem, in the presence of a risk- free asset and given a single conditioning information time series, may be expressed as a general constrained

information in such a way that more general variations can be tackled (using numerical algorithms if necessary) Want to integrate this type of optimisation problem into an