Thesis
Reference
Computational methods for selecting optimal financial investment strategies
LULA, Jonela
Abstract
This thesis focuses on empirical asset allocations problems. The nonconvex optimization problem arising from our models specification is solved by means of heuristic optimization methods. Three empirical applications of a particular heuristic, the Threshold Accepting method, are proposed. The first problem that we consider is the replication of the Credit Suisse/Tremont (CST) Hedge Fund Index using liquid instruments such as equities, commodities and bonds. Our specification yields portfolios appearing to be an attractive substitute to the tracked index. In the second application we explore whether an asset allocation approach to Foreign Exchange Market is profitable. Our approach dominates the benchmark portfolio and technical trading model, as well as being less volatile. In the last application, we incorporate Asset-Liability Management in the asset allocation decision. The model used is the multistage programming, which outperforms the other approaches such as mean-variance or minimum downside risk.
LULA, Jonela. Computational methods for selecting optimal financial investment strategies . Thèse de doctorat : Univ. Genève, 2013, no. SES 796
URN : urn:nbn:ch:unige-266898
DOI : 10.13097/archive-ouverte/unige:26689
Available at:
http://archive-ouverte.unige.ch/unige:26689
Disclaimer: layout of this document may differ from the published version.
Computational methods for
seleting optimal nanial
investment strategies
Thèse
présentée à laFaulté des sienes éonomiques etsoiales de
l'Université de Genève
par
Jonela Lula Chamay
pour l'obtention du grade de
Doteur ès sienes éonomiques et soiales
mention éonométrie
Membres du jury de thèse :
Prof. Manfred Gilli, direteur de thèse,Université de Genève
Prof. JayaKrishnakumar,présidentedu jury, Université deGenève
Prof. Dietmar Maringer, Université de Bâle
Prof. DanielRoyer, Université de Genève
Thèse n o
796
Genève,le26février2013
l'impressiondelaprésentethèse, sansentendre,parlà,n'émettre auuneopinion
surlespropositionsquis'ytrouventénonéesetquin'engagentquelaresponsabilité
deleurauteur.
Genève,le26février2013
Ledoyen
BernardMORARD
Impressiond'aprèslemanusritdel'auteur
Résumé iii
Abstrat v
Aknowledgements vii
Introdution 1
1 Repliatinghedgefund indieswith optimizationheuristis 7
1.1 Hedgefunds . . . 7
1.2 Themodels: Indextraking . . . 8
1.3 Appliation: Trakinghedgefundindies . . . 11
1.4 Conludingremarks . . . 18
2 FXtrading: An empirial study 21 2.1 Introdutiontoforeignexhange . . . 21
2.2 Tehnialanalysis inurrenymarkets . . . 22
2.3 Themodels . . . 24
2.3.1 Assetalloationapproah . . . 24
2.3.2 Tehnialtradingapproah . . . 26
2.4 Empirialresults . . . 27
2.4.1 Dataandbaktestingsheme . . . 27
2.4.2 Simulationresultforassetalloationapproah . . . 29
2.4.3 Simulationresultsfortehnialtrading. . . 33
2.5 Robustnesshek . . . 34
2.5.1 Resultsfor2010 . . . 34
2.5.2 Comparingresultstothebenhmark . . . 35
2.6 Conludingremarks . . . 36
3 Asset-Liability Management for Individual Investors in a Multi- Senario and Multi-PeriodSetting 39 3.1 Introdution. . . 39
3.2 MultistageModel . . . 43
3.3 TheMultivariateSwithingRegimeModel. . . 44
3.4 OptimizationModel . . . 47
3.5 Comparisontootherinvestmentstrategies . . . 49
3.6 Empirialresults . . . 50
3.6.1 Dataandbaktestingsheme . . . 50
3.6.2 Example: Firstportfolioat31-De-2004 . . . 51
3.6.3 Results . . . 53
3.7 Conludingremarks . . . 57
Conlusion 59
Cettethèsetraitedesproblèmesempiriques d'alloationdesatifs.
Nous utilisons des modèlessans ontraintes et, par onséquent et étant donnéla
non-onvexitédesproblèmesprésentés,lesméthodesd'optimisationquis'imposent
sontlesméthodesditesheuristiques.
Sontproposéesainsitroisappliationsd'uneméthodeheuristiquepartiulière,soit
laméthodeThresholdAepting,
Lepremiersujettraité,dansleadredeetravail,esttel: répliquerlaperformane
de l'indie CreditSuisse/Tremont(CST) HedgeFunden utilisantdesinstruments
liquides,tels queles ations,lesmatièrespremièreset lesobligations. Aussi nous
onstatonsquenotremodèleproduitdesportefeuillesquisuiventdeprèsl'évolution
de l'indie. Cesportefeuilles sontplusliquides, plustransparentset les fraissont
moins élevés.
Dansladeuxièmeappliation,noustentonsdesavoirsiuneapprohed'alloation
d'atifspourlemarhédesdevises(FOREX)peutêtrerentable. Etils'avèreque
notreapprohesemblepluseaequeleportefeuillederéférene,meilleureaussi
que eluiqu'uneapprohed'analyse tehniquepeutproduire. Le portefeuille que
nousobtenonsparl'approhed'alloationd'atifsest égalementmoinsvolatile.
Dansladernièreappliation,nousonsidéronslagestionatif-passif(Asset-Liability
Management,ALM)dans ladéisiond'alloationd'atifs. Lemodèlepourlequel
nousavonsoptéestlaprogrammationmultistage. Etlàenore,lesrésultatsnous
amènent à la onlusion que que notre modèle semble plus judiieux que elui
produit par d'autres approhes, telles notamment les models mean-variane ou
minimumdownsiderisk.
This thesis fouses on empirial asset alloations problems. We speify models
without any restritions, as a onsequene the nononvex optimization problem
arisingfrom ourmodelsspeiationissolvedbymeansof heuristioptimization
methods.
Three empirial appliations of a partiular heuristi, the Threshold Aepting
method, areproposed.
TherstproblemthatweonsideristherepliationoftheCreditSuisse/Tremont
(CST) Hedge Fund Indexusing liquid instruments suh as equities,ommodities
and bonds. Our speiation yieldsportfolios appearing tobean attrative sub-
stitute to thetrakedindex. These portfoliosare moreliquid, theomposition is
transparentand theyhavelowerfees.
IntheseondappliationweexplorewhetheranassetalloationapproahtoFor-
eignExhangeMarketisprotable. Ourapproahdominatesthebenhmarkport-
folioandtehnialtradingmodel, aswellas beinglessvolatile.
In the last appliation, we inorporate Asset-Liability Management in the asset
alloationdeision. The model used isthe multistageprogramming. This model
outperforms the other approahes suh as mean-variane or minimum downside
risk.
First of all I would liketo thank my supervisor, Professor Manfred Gilli, for his
supportandguidane during alltheseyears. It was anopportunityworkingwith
him,from whomIwas taughttoworkwithpreisionandexatitude.
I am grateful to Professor Daniel Royer, with whom I had the hane to teah
Mathematis in thisFaulty. Throughouttheyears,he hasbeenarolemodelfor
his generosityandpassioninteahing.
I am also partiularly thankful to Professor Jaya Krishnakumar and Professor
DietmarMaringerfor aeptingtotakepartin mythesisommitteeandfortheir
onstrutivesuggestions.
I would also like to thank Gerda Cabej. It was a pleasure working and sharing
withhertheseintense moments.
I would also liketo express my gratitude to IlirRoko andEnrio Shumman for
theirollaborationandtheirvaluablesuggestions.
ThisworkwouldnotbethesamewithoutthehelpofEvisKëllezi,withoutwhom
I wouldhavenotstartedthisprojet.
IamsinerelythankfultoallprofessorsoftheformerDepartmentofEonometris,
for theirpassionofteahing andpartiularly, ProfessorGabrielleAntilleGaillard
forherhelpandsupport throughoutmystudiesinthisFaulty.
ThanksmustgotoMaroRigoforkindlyreadingthisthesisandprovidingvaluable
suggestions.
Andagain,manythankstoallmyolleaguesinthisdepartment,formero-workers
ofAvendis Capitalandtoallmyfriendsfortheirenouragement.
Andlastbutnotleast,themostspeialgratitudegoestomyfamily,fortheontin-
uoussupportandmotivation, in partiularthanksto InaFurerajforallher help,
Charles-Antoineforhislove,andtomyfatherwhohasneverstoppedghting.
Theoptimizationmodelsthatwillbepresentedinthisthesis,aresuhthatannot
be solved by means of lassial optimization methods. As a matter of fat, the
lassialoptimization methods (suh as linear andquadrati programming)relay
onananalytialmodeloftheobjetivefuntion. Thesetehniques areeientfor
problemswithsmoothobjetivefuntion.
Unfortunately,inrealitywehavetofaeomplexobjetivefuntionsthatannotbe
solvedbyusinglassialoptimizationmethods. Thisiswhereheuristitehniques
taketheleadastheyoeranaturalwaytooveromethedeieniesofthelassial
methods. Heuristisdonotrequirepartiularpropertiesfortheobjetivefuntion,
in partiular derivatives are not needed (only funtion evaluations). The main
purpose of this thesis is to illustrate how, these omplex optimization problems
an beeientlysolvedbyresortingto heuristioptimization tehniques.
Heuristis have been put into pratie relatively reently on a larger sale, as a
result of the impressive development of omputing power at a low ost. These
methods adapt easily to any optimization, are robust to hanges, give 'optimal
enough'solution(GilliandShumann,2011a)totheproblemathandandareeasy
to implement.
Heuristis may bedivided in two ategories: trajetory methods and population
based methods. Trajetory methods work with a single solution, whereas the
population-based methods onsider a population of solution simultaneously. In
thisresearh,onlyatrajetorymethodbasedonaloalsearhisused.
AmongthetrajetorymethodswehaveSimulatedAnnealing,ThresholdAepting
andTabuSearh. Thesemethodsdierfromeahotherinthewaytheaeptane
riterionishosen. ApartfromTabuSearh,allthesemethodsaeptuphillmoves
whihallowthemtoesapeloalminima.
Inourresearh,weonlyuseThresholdAepting(TA).Thismethodwasproposed
simultaneouslyby Duek andSheuer(1990)and Mosatoand Fontanari(1990).
Throughitsaeptaneriterionwhihinludes adereasingthreshold(henethe
name), the algorithm aepts not only improvements, but also impairments of
the objetive funtion in order to esapeloal minima. It begins with an initial
solutionandperformsanumberofiterations. Ateahiteration,anewandidateis
generated in theneighbourhoodoftheurrentsolution. Basedontheaeptane
riterion, thenewsolutioniskept. If thenewsolutionisrejeted, TA retainsthe
urrentsolutionandthealgorithmsearhesforanewsolutionintheneighbourhood
ofthisurrentsolution.
Inwhat followstheThresholdAeptingisusedfortheseletionofoptimalasset
alloationstrategies.
There already is a signiant body of researh that deals with the problem of
portfolio optimization. Dierent diretions have been pursued. Rokafellar and
Uryasev(2000,2002)werethersttolinearizetheportfoliooptimizationproblem
in ordertoeientlysolveitwithstandardlinearprogrammingtehniques. Mor-
ton,Popova,Popova,andZhong(2006)showthat whenreturnsarenotnormally
distributed,theportfoliooptimizationproblemannotbesolvedviastandardnon-
linearprogrammingtools. Theyuseanapproximationapproahrooted inMonte
Carlosimulation,usingabranh-and-boundsolutionproedureforsolvingadi-
ultmixed-integerprogram.
As mentionedabove,wetakeadierentresearh diretion basedonanumberof
ontributions where heuristis, Threshold Aepting in partiular, is applied for
theportfoliooptimization problem. DuekandWinker(1992)are therst touse
heuristisoptimization,speiallytheThresholdAepting,inaportfolioseletion
problem. LaterGilliandKëllezi(2002a,b)useThresholdAeptingforminimizing
Value-at-Risk and expeted shortfall. Maringer (2005)applies heuristi methods
tosolveavarietyofportfoliomanagementproblems,dealingamongotherwithin-
tegervariablesand ardinalityonstraints. Chang,Meade, Beasley,and Sharaiha
(2000)disussardinalityonstrainedportfoliosandMansiniandSperanza(1999)
usealinearprogrammingapproah tosolvetheportfolioproblem withminimum
tradingsizeonstraints.
Meanwhile aninreasing numberof papershave been written providing evidene
of thegoodperformane ofheuristioptimizationtehniques appliedto problems
in nane 1
.
Ourstudyfollowsthesameapproah(inthatweuseheuristisinnane),applied
to aseriesofempirialappliations. Theseonsistin:
•
atrakingproblemforaHedgeFundIndex•
anassetalloationproblemintheframeworkof ForeignExhangeMarket•
anassetalloationproblemwithlongtermliabilityThe traking problem 2
for Hedge Fund Index is presented in Chapter 1. The
problem isapplied toCreditSuisse/TremontHedgeFundIndex(CST).
Consideringbenhmarkreturnsintheobjetivefuntionreetsinamorerealisti
way the motivation of today's investmentmanagers. In pratie, besides aiming
atthebestpossiblerisk-adjustedreturnontheirinvestments,amajoronernfor
portfoliomanagersisndingawayto performbetterthanabenhmarkspeied
in themandateguidelines andquite oftenhowtoinrease thehanes ofoutper-
1
fori.e. Gilli,Maringer,andShumann(2011a)andthereferenestherein.
2
ExploredinGilliandKëllezi(2002b).
formingtheirpeers. Browne(2000)is,toourknowledge,thersttoformalizethe
problemofportfoliooptimizationrelativetoabenhmarkinontinuoustime. The
benhmarkperformaneanbedeterministi,i.e.: axedrateofreturnperyear,
in asewhen an absolute return is the investmentobjetive,or random, suh as
thereturn onapredened market index (likeS&P 500). Thelateraseonerns
themajorityoflong-onlymandates inbondorequityinvestments.
There arevarious waysofintroduing benhmark relatedreturnsinthe objetive
funtion. One an try to maximize the probability of ahieving the benhmark
return, ignoring the magnitude by whih the target is missed or exeeded. An-
other possibility would be to minimize the downside or themagnitude by whih
thebenhmark ismissed,dened by Demboand King(1992)as theexpeted re-
gret. Mortonet al.(2006)proposeafamily ofobjetivefuntions onstruted as
a weighted sumof theprobabilityof ahievingabenhmark and expeted regret
relativetoanother(lessaggressive)benhmark,whihanbeinterpretedasaom-
binationofmeasuresofrewardandrisk. Thesemeasuresarerelatedtothealready
widelyknownriskmeasures,Value-at-RiskandExpetedShortfall.
In ourstudy, we rstminimize themaximumdrawdownand then wetest anew
objetive funtion inluding also a measure of orrelation between the portfolio
and theindex. These objetivefuntions aresubjetto aset ofonstraints,more
speiallybudget onstraints,minimum andmaximumholding sizesfor thesets
ofassets,totaltransationostsandardinality. Theoptimizationproblemweare
faingisanon-onvexone,giventheobjetivefuntionandsomeoftheonstraints.
Inordertosolveit,theThresholdAeptingmethodis used.
Heuristi optimization tehniques are also used for the appliation presented in
Chapter 2. Weapply anassetalloationapproah toaset oftik-by-tikdataof
veurrenypairsandoptimizeamulti-objetivefuntion (maximumdrawdown,
Omega,momentum andombinationof them). Wealsoapplyatehnialtrading
rulebasedmodel. Inpartiular,weexploreappropriatelevelsoftimeaggregation
andrebalaningfrequenies,andweretainhourlyaggregationanddailyrebalane-
ment. Results are afterwards ompared to a benhmark portfolio (buy-and-hold
strategy).
In the last hapter, we introdue long term liabilities in the asset management
problem of an investor. This involves long planning horizons and multi-period
realloationsinordertoadapttotheevolutionofthemarket.Thisorretsforone
ofthelimitationsoftheMeanVarianeapproahwhihisitssingleperiodhorizon
(not favorable forlongterminvestors). TheAsset-LiabilityManagementaims to
ontroltheriskassoiatedtofutureliabilities,aswellastondtheoptimalportfolio
as a trade-o betweenthehighest returnof theportfolio and thelowest possible
assoiatedrisk. Dynamis ofthe assetsin thesuessive periods are represented
in a senario tree and a multistage programming approah is used to ompute
realloations. The optimization problem is one again solved by means of the
ThresholdAeptingmethod.
It wasourintentioninthisthesistomakeeahhapter astand-alonereading. As
aonsequene,Chapters13mayontainsomeredundantinformation.
Repliating hedge fund indies
with optimization heuristis
∗†
1.1 Hedge funds
Startingfromtheearly90s,hedgefundsisonsideredtobeoneofthemostpopular
setorof thenanialindustryandthis ismainly due to theperformanesofthe
hedge funds. Amongthedesirablefeatures ofhedge funds wehavealpha genera-
tion,loworrelationwiththemarket,attrativerisk-returnproleandalternative
riskpremia 1
. Themaindrawbaksarehighmanagementfees,lakoftranspareny
and illiquidity. This leadsto theinentivetorepliatetheattrativefeaturesofa
hedge fund with liquid assets, thus providing afully transparent, liquid and low
ost alternativeto a hedge fund. There exists a largeliterature giving morede-
tailedaboutthemotivationforhedgefundrepliation(egTill(2004)Shneeweiss,
Kazemi,andMartin (2002),Shneeweiss,Kazemi,andMartin(2003)).
∗
ThishapterisbasedonGilli,Shumann,Cabej,andLula(2010).
†
AllauthorsgratefullyaknowledgenanialsupportfromtheeuCommissionthroughmrtn-
t-2006-034270omisef. EvisKëlleziprovidedhelpfulomments.
1
seeLhabitant(2006),Lo(2008)andCoggan(2011)foroverviews
Several tehniques are used to repliate the properties of hedge fund returns.
Among themain approaheswehavefatormodels (Sharpe(1992), Hasanhodzi
and Lo (2007)), trading rules (Kat and Palaro (2005)) and traking tehniques
(Roll (1992), Rudolf, Wolter, and Zimmermann (1999), Alexanderand Dimitriu
(2005)). Kat(2007)proposetradingstrategiestogeneratereturnswithstatistial
properties,similartothoseofhedgefunds.
Our approah diers in that we proeed in speifying the models without any
restritionswithrespettotheirtratabilitywithlassialoptimizationtehniques.
Thisexibilityispossiblebeauseweuseheuristioptimizationmethods allowing
forthesolutionofvirtuallyanyoptimizationproblemgivenofourseitissound.
We test a varietyof models of inreasing omplexity going from a pure traking
objetive to drawdown. In between, we explore ombinations of traking error
withexessreturn,orrelationoftheportfoliowiththeindexandthemarket. The
omputed portfolios satisfyrealistionstraints,suh asminimum andmaximum
holding size, maximum ardinality and also take into onsideration transation
osts. Thenononvexoptimization problemarisingfrom thesespeiationsan-
not besolvedwith lassial gradientbased methods. Our heuristi optimization
tehniqueprovesveryeientforthiskindofproblems(Maringer(2008),Maringer
andOyewumi (2007),Maringeranddi Tollo(2009),Krink,Mittnik,andPaterlini
(2009)).
The hapter is organizedas follows: in Setion 2 dierent models are presented
andtheoptimizationproblemisformulated. Setion3givesinformationaboutthe
analyzeddataanddetailstheresultsforthedierentmodels. Setion4onludes.
1.2 The models : Index traking
A straightforwardapproah is to onstrut a trakingportfolio whih minimizes
the distane between historial portfolio and index returns. We denote
r P
andr I
thehistorialreturnvetorof thetrakingportfoliorespetivelytheindex andr E = r P − r I
the exessreturn of the portfolio. In order not to penalize upsidedeviationsforaportfoliowealsoonsiderthemeanexessreturn
r E = 1 n P
(r P − r I )
leadingtothefollowingobjetivefuntion
λ k r E k α − (1 − λ) r E
(1.1)where
α
speies apartiular distaneandλ ∈ [0, 1]
denesalinearombinationbetweentrakingerrorandexessreturn. ThisapproahhasbeenexploredinGilli
andKëllezi(2002b)usingartiialdata.
One of the main goalsis to onstrut a portfolio following the index as lose as
possiblebut beinglittlesensitivetoadversemarketmovements. This suggeststo
inlude the orrelation between traking portfolio and market into the objetive
funtion
λ k r E k α − (1 − λ) r E + ρ r P ,r M
(1.2)where
ρ r P ,r M
denotesthisorrelationomputedfromthehistorialdata. Minimiz-ingtheobjetivefuntion minimizesthisorrelation.
Model(1.1)fousingonthetrakingerror,alreadyleadstoaportfoliohighlyor-
relatedwiththeindex. Neverthelessoneanthinktoontrolthisorrelationmore
speiallybyintroduingitintotheobjetivefuntion. Denoting
ρ r P ,r I
thisorre-lationbetweenportfolioandtheindex,itanbemaximizedwiththenewobjetive
funtion
λ k r E k α − (1 − λ) r E + ρ r P ,r M − ρ r P ,r I .
(1.3)Afurtherdesirablefeatureofthetrakingportfoliowouldbetohavelowdrawdown.
Foraseriesofportfoliovalues
v t , t = 0, 1, 2 . . . T
thedrawdownisdenedasD t = v t max − v t
where
v max t
istherunningmaximum,iev t max = max { v s | s ∈ [0, t] }
. Followingthisdenition
D
isa vetor forwhih wean ompute themean, standarddeviationor the maximum element
D max = max( D )
whih is the one we use in our nextobjetivefuntion. Inotherwords
D max
measuresthelargestdropoftheportfoliovalue overthe time horizon. Inarst stepweonsider onlythe minimization of
themaximumdrawdownforouranalysis. Inaseondstepweombinemaximum
drawdownminimizationwiththeobjetivefuntion denedin (1.3)yielding
λ k r E k α − (1 − λ) r E + ρ r P ,r M − ρ r P ,r I + D max .
(1.4)The optimization problem
Thegeneraloptimization problemanbestatedas follows
min x Φ(x)
(1.5)X
j ∈J
x j p 0j = v 0
(1.5′
)
x inf j ≤ x j ≤ x sup j j ∈ J
(1.5′′
)K inf ≤ # {J } ≤ K sup
(1.5′′′
)
.
.
.
where
Φ
isoneoftheobjetivefuntionspresentedpreviouslyinthissetionandx
isavetorwith
x j
thequantityofassetj
intheportfolio. Thisoptimizationissubjet to aset ofonstraintswhih areinpartiularthebudgetonstraint(1.5′
)with
v 0
theinvestablewealthand
p 0j
theprieofassetj
atthebeginningoftheinvestmentperiod. Constraint(1.5
′′
)speiesminimumandmaximumholdingsizefortheset
ofasset(
J
)intheportfolio. Nextwehavetheardinalityonstraint(1.5′′′
)whihallowsfortratabilityoftheresultingportfolios. Furtheronstraintsmightinlude
totaltransationost,setoronstraintsandotherliquidityissues.
2
Theoptimization problems(1.5)arenononvexdueto theformulationoftheob-
jetivefuntionandpartoftheonstraints. Theyannotbesolvedwithalassial
approah relyingon informationfrom gradientvalues. We thereforeuse Thresh-
oldAeptingwhihisaheuristioptimizationmethodfromthelassoftrajetory
methods. DetaileddesriptionofthetehniqueanbefoundinWinker(2001),Gilli
and Winker(2009) and for appliations see Gilli and Shumann(2010). Matlab
odeisavailableathttp://omisef.eu.
Heuristimethodsdonotproeedrandomlybutexplorethesearhspaeaording
to ertainrules. Neverthelesstheresultingnumberof funtionevaluationsisrela-
tivelyimportant. Intheaseoftheoptimizationofasingleportfoliotheomputing
timestaysintherangeofafewseonds. Howeverthebaktestingweondutedfor
ourappliationneededtheevaluationof approximativelyhalfamillionportfolios.
InordertokeepouromputationsmanageabletheyweredistributedwithMatlab's
ParallelComputing Toolboxonthe MyrinetClusterof theUniversity ofGeneva.
Myrinet is a Linux Cluster with 32 nodes, eah on a Sun V60x dual IntelXeon
2.8GHzwith2GB ofRAM.For moredetailsseehttp://sp.unige.h/.
1.3 Appliation : Traking hedge fund indies
Data and baktesting sheme
The index to repliate is the Credit Suisse/Tremont Hedge Fund Index (CST)
availableatwww.hedgeindex.om. Aordingtotheinformationonsite thisindex
isasset-weighted,inludesmorethan5000fundswithaminimumofUS$50million
under management. The observations are monthly and over the period from
January1999toOtober2009.
The instruments used for the repliation omprise equity, ommodity and bond
indies. Inthesetofequityindieswehaveabout54seriesinludingbroadmarket,
bluehips,setoraswellassizeandstyleindies. Thereare12ommodityindies
and 12 bond indies from government,orporate andemerging markets. The set
ofdatahasbeenolletedfromBloomberg.
In order to analyze the performane of the suggested portfolios we baktest the
strategiesovertenyearswithrebalaningwhereweaountfor10basispointsof
2
OptimizationwithonstraintsontotaltransationosthavebeenusedinGilliand Këllezi
(2002b),setorandliquidityonstraintsinGilli,Shumann,diTollo,andCabej(2011b).
transation osts. Therolling window hasahistoriallengthof
H
and aholdingperiodof
F
. Inthis appliationH
is oneyearandF
isthree months. This leadsto trajetoriesofportfoliosvalueswheretheportfolioshavebeenrebalanedforty
times. Theshemebelowsummarizesthetehnique.
Period
1
t 1 − H t 1 t 1 + F
F
H
invest
Period
2
t 2 − H t 2 t 2 + F
rebalane
Forallportfoliostheminimumholdingandmaximumholdingforanassetis
x inf j = 1%
respetivelyx sup j = 20%
. Maximumardinality ofthe portfolios islimited toK sup = 10
.Togaininsight into thestohastis of the simulatedportfolioswejakknife from
the historial observations so as to ompute a set of results 3
for whih we then
onsider empirialdistributions fordierentfeaturesoftheportfolio.
Results
In the following we present for eah objetive funtion the median path of the
portfolio value for varying parameter settings, aplot of the kernel estimation of
the density the empirial distribution of the mean yearly return of the traking
portfolioandaplotoftheempirialdistributionoftheorrelationoftheoptimized
portfolios withthe market. Themedian path isdened withrespetto the nal
wealthoftheportfoliosgeneratedwiththejakkning.Forallmodelsweset
α = 2
forthenormmeasuringthetrakingerror.
Performanewillalsobesummarizedandomparedwiththeindexintablesshow-
ingthetrakingerror
TE
denedasthestandarddeviationofr E
,theSharperatio(
S
),theannualizedreturnandvolatility(r
,vol
)andtheorrelation(ρ r P ,r M
)withthe market of the portfolios minimizing the objetive funtion, realized overthe
simulationperiod.
3
Inthisaseweomputed100trajetoriesforeahspeiationoftheobjetivefuntion.
Optimization oftraking error and exess return
Weomputedresultsfortheobjetivefuntiondesribedin(1.1)forvaryingvalues
of the parameter
λ
ontrollingthe weightingbetween traking errorand reward.Figure1.1showsthemedianpathsfor100simulationsfor
λ = 1
(darkline)whihorrespondstominimizingonlytrakingerror. Weantradetrakingerroragainst
nalwealthbydereasing
λ
. Agoodompromiseisobtainedforλ = 0.75
yieldingportfolios loseto theindex whereashigher weights for thereward (lowervalues
for
λ
)resultin highernal wealth but alsohigher volatility. The dotted vertial linesindiates therebalaningdates.Jan00 May01 Oct02 Feb04 Jul05 Nov06 Apr08 Aug09
0.5 1 1.5 2 2.5 3
λ=1.00 λ=0.75 CST S&P500
Figure 1.1: Medianpaths for portfolios minimizingthe objetivefuntion (1.1) for
λ = 1
andλ = 0 . 75
.CSTstandsforCSTremont.Figure1.2illustratesanotherfeatureofthesimulatedportfolios. Itshowstheplot
of the kernel estimation of the densityof the empirial distribution of the mean
yearlyreturnof thetrakingportfolio(left panel). Thevertialline indiatesthe
meanyearlyreturnoftheindex. TherightpanelinFigure1.2showstheempirial
distribution of the orrelation of the optimized portfolios with the market. The
dotted line orrespondsto the orrelation of theindex with the market. For the
trakingportfoliosweobservehighervalues. Table1.1summarizestheperformane
fortheportfoliosobtainedwithobjetivefuntion (1.1).
Table1.1: Resultsforthemedianpathofthesimulatedportfoliosforobjetivefuntion(1.1).
TE S r vol ρ r P ,r M
λ = 1
2% 0.45 4% 9% 0.62λ = 0.75
2% 0.62 6% 10% 0.55CST 1.04 7% 7% 0.54
The medianpath of theportfolio for
λ = 1
in Figure 1.1 doesnotlook favorabledespitethefatthatithasalowtrakingerror. Thisresultmighthurtintuitionbut
2 4 6 8 10 12 14 16 18 0
0.2 0.4 0.6 0.8 1
λ=1.00 λ=0.75 CST
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0 2 4 6 8 10 12 14
λ=1.00 λ=0.75 CST
Figure1.2:Densityofmeanyearlyreturn
r
(leftpanel)andorrelationρ r P ,r M
(rightpanel).the trakingerrorompares the returnsand not portfolio values. Twoportfolios
with a dierent rst return but idential returns for the remaining periods will
translateinto o-evolvingtrajetories. The portfolio for
λ = 0.75
hasanaveragereturninreased by50%omparedwith
λ = 1
in exhangeof insigniantlossintrakingperformaneandsmallinreaseofvolatility. Furthermoretheorrelation
oftheportfoliowiththemarketdereases.
Optimization oftraking error, exess return and
ρ r P ,r M
Thehigherorrelationofthetrakingportfolioswiththemarketisnotadesirable
feature. To improve on this we now turn to the objetive funtion (1.2) whih
inludestheorrelationofthetrakingportfoliowiththemarket. Figure1.3gives
the orresponding median pathsfor valuesof
λ = 1, 0.75, 0.7
. Forλ = 1
, wherenoexessreturnenterstheoptimization,weobserveapartiularlysmoothmedian
pathalmostnotaetedbythedropin theS&P500attheendof2008.
Results in the right panel of Figure 1.4 are remarkablewhen ompared with the
ones in Figure1.2 as thedistributions indiate that, while maintainingthe same
levelofreturns,orrelationisnowwellontrolled.
Table1.2: Resultsforthemedianpathofthesimulatedportfoliosforobjetivefuntion(1.2).
TE S r vol ρ r P ,r M
λ = 1
2% 0.60 4% 7% 0.35λ = 0.75
2% 0.68 6% 9% 0.09λ = 0.70
3% 0.63 7% 10% 0.07CST 1.04 7% 7% 0.54
Jan00 May01 Oct02 Feb04 Jul05 Nov06 Apr08 Aug09 0.5
1 1.5 2 2.5 3
λ=1.00 λ=0.75 λ =0.70 CST S&P500
Figure1.3:Medianpathsforportfoliosminimizingtheobjetivefuntion(1.2)ontrollingorre-
lation
ρ r P ,r M
withthemarket.2 4 6 8 10 12 14 16 18
0 0.2 0.4 0.6 0.8 1
λ=1.00 λ=0.75 λ=0.70 CST
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0 2 4 6 8 10 12 14
λ=1.00 λ=0.75 λ=0.70 CST
Figure1.4: Densityofmeanyearlyreturn
r
(leftpanel)andorrelationρ r P ,r M
(rightpanel)fortheobjetivefuntion(1.2)ontrollingorrelationwiththemarket.
Optimization oftraking error, exess return,
ρ r P ,r M
andρ r P ,r I
Wenowhekwhetherobjetivefuntion (1.3)whihonsiderstrakingerrorand
orrelationbetweentheportfolioandthetrakedindexenhanesperformane. No-
tie thattheeet of
λ
sare notomparablebetweenthedierent objetivefun-tionsduetotheimpatoftheadditionalterms. Theresultsindiatenosigniant
hangeinoverallperformane,weratherobserveashifttoimprovedSharperatios
for allvaluesof
λ
. Again lowervaluesforλ
,iehigherweightingforexessreturnleads to portfolios with highernal wealth but of ourseat the ostof inreased
volatility.
As visible in Figure 1.6, from the point of view of returns
λ = 0.6
produes at-trativeportfolioswhihmoreovershowquiteloworrelationwiththemarketand
Jan00 May01 Oct02 Feb04 Jul05 Nov06 Apr08 Aug09 0.5
1 1.5 2 2.5 3
λ=0.68 λ=0.65 λ=0.60 CST S&P500
Figure1.5:Medianpathsforportfoliosminimizingtheobjetivefuntion(1.3)ontrollingorre-
lation
ρ r P ,r M
withthemarketandρ r P ,r I
withtheCST.highSharperatios.
2 4 6 8 10 12 14 16 18
0 0.2 0.4 0.6 0.8 1
λ=0.68 λ=0.65 λ=0.60 CST
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0 2 4 6 8 10 12 14
λ=0.68 λ=0.65 λ=0.60 CST
Figure1.6: Densityofmeanyearlyreturn
r
(leftpanel)andorrelationρ r P ,r M
(rightpanel)fortheobjetivefuntion(1.3)ontrollingorrelationswiththemarketandtheCST.
Optimization oftraking error, exess return,
ρ r P ,r M
,ρ r P ,r I
andD max
Wenowexploretheimpatoftheintrodutionofdrawdowninourobjetivefun-
tionontheoverallperformaneoftheportfolios. Figure1.7 belowreports results
for pure maximum drawdown minimization (dark line) as well as results for ob-
jetivefuntion(1.4) withdierentweightingoftrakingerrorandexessreturn.
As previously,higherweightsforexessreturn(low
λ
s)produehigh nalwealthportfolios.
Lookingat thenalwealth distributiongivenin Figure1.8 wesee thatfor
λ = 1
the expetation of the mean yearly returns is approximatively that of the index
Table1.3: Resultsforthemedianpathofthesimulatedportfoliosforobjetivefuntion(1.3).
TE S r vol ρ r P ,r M
λ = 0.68
3% 0.74 8% 10% 0.17λ = 0.65
3% 0.79 8% 10% 0.15λ = 0.60
3% 0.88 10% 11% 0.09CST 1.04 7% 7% 0.54
Jan00 May01 Oct02 Feb04 Jul05 Nov06 Apr08 Aug09
0.5 1 1.5 2 2.5 3
D max λ =1.00 λ =0.70 λ =0.00 CST S&P500
Figure1.7:Medianpathsforportfoliosminimizingtheobjetivefuntion(1.4)ontrollingorre-
lations
ρ r P ,r M
,ρ r P ,r I
andthemaximumdrawdown.howeverfor
λ = 0
weobserveanimpressiveshiftofthedistribution totheright.Table1.4: Resultsforthemedianpathofthesimulatedportfoliosforobjetivefuntion(1.4).
TE S r vol ρ r P ,r M
DD max
2% 1.01 9% 9% 0.37λ = 1
2% 0.81 6% 8% 0.24λ = 0.7
2% 0.97 9% 9% 0.17λ = 0
4% 0.94 16% 17% 0.23CST
1.04 7% 7% 0.54Inthelightoftheseresultsportfoliosobtainedfromthislastmodeloerproperties
suitable tosubstitute theCreditSuisse/Tremonthedge fund index. Inpartiular
for
λ = 0.7
wehaveomparableSharperatiobuthigheraveragereturnandlowerorrelation with the market. However the portfolio an be modulated hoosing
dierentvaluesfor
λ
in orderto meet dierent preferenesor risk aversionof aninvestor.
Itmightbeinterestingtoshowhowthedierentassetlassesarerepresentedinthe
2 4 6 8 10 12 14 16 18 0
0.2 0.4 0.6 0.8 1
D max λ=1.00 λ=0.70 λ=0.00 CST
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0 2 4 6 8 10 12 14
D max λ=1.00 λ=0.70 λ=0.00 CST
Figure1.8:Densityofmeanyearlyreturn(leftpanel)anddensityoforrelationwiththemarket
(right panel) for the objetive funtion (1.4) ontrolling orrelations
ρ r P ,r M
,ρ r P ,r I
and themaximumdrawdown.
median portfolios. This isplotted in Figure1.9 andwenotie howtheoptimizer
reats to market onditions. For instane, in periods of distress, the weight of
xed inome instrumentsgains importane at the expenseof equities. Also, the
portfoliowhereexessreturnisfavoredhasrelativelyhigherweightedommodities
andequities.
1.4 Conluding remarks
Thegoalofthisresearhistoinvestigatetheuseofasetofmodelsforthelassial
problem of index traking. Thespeiationof themodels is madewithout on-
sideringtheonstraintsimposedbythelassialoptimizationframework. Instead,
weresorttoheuristioptimizationmethodsapableofeientlyhandlingomplex
nononvexproblems. Westartfromasimplespeiationombiningtrakingerror
andexessreturn,whihgivensomeoftheonstraints,isalreadynottratablewith
anylassialmethod. Thismodelisthensuessivelyenrihedbyaddingelements
to ontrol for the orrelationof the portfolio with themarket and laterwith the
index. Ourresults onrmthat weare able to derease signiantlythe orrela-
tionwiththemarket. Finally,weexploretheimpatofminimizing themaximum
drawdown as well as a ombination of it with traking error, exess return and
orrelations. Thislast speiationyieldsportfoliosappearingto beanattrative
substitute to the traked index. Indeed, theirSharperatios ompare to those of
the index, while their average mean return is higher and their orrelation with
the market is lower. These results are notpoint estimates but are supported by
0 0.5 1
0 0.5 1
0 0.5 1
Jan00 May01 Oct02 Feb04 Jul05 Nov06 Apr08 Aug09
0 0.5 1
bonds
equities
commodities
Figure 1.9: Medianportfolioompositionintermsofassetlasses(lowerband equities,middle
ommoditiesandupperbandbonds)fortheobjetivefuntion(1.4). Fromtoptobottom
D max
,λ = 1, 0.7
and0
.empirialdistribution obtainedfromintensivebaktesting.
FX trading: An empirial
study
∗†
2.1 Introdution to foreign exhange
The foreign exhange (FOREX) market was reatedin 1971, year in whih the
oating exhangerates began to appear andeversineitis onsidered to be one
of the largest and most important markets. The high liquidity, the guaranteed
quantity, theround-the lok business hoursworldwidetradingativity, two-way
market,aessibilitytoalltradersandlowtransationostsarefeaturesthathar-
aterizetheFOREXmarket. Partiipantsinthismarketareentralbanks,banks,
brokers,ommerialompanies,hedge funds,investorandspeulators.
Historially,theurrenymarketoriginatesin theearly20th entury. Inthe'30s,
LondonwastheenterofnaneandtheBritishpoundthemainurrenyusedin
tradingas wellas,thereserveurreny. AfterWorldWarIItheUnitedStatesbe-
ame theleadingenter andtheU.S.dollarthe"vehile"urreny. Allurrenies
arequoted againsttheU.S.dollarandthemajoronesaretheeuro,Japaneseyen,
BritishpoundandSwissfran.
The foreignexhange hasexperiened spetaular growthin its volume. In1988
the turnoverreah the value of $600 billion, rises up to 1 trillion in September
1992and inApril2010thedailyaverageforeignexhangemarketturnoverhit$4
trillion. Currenyandinterestratevolatility,advanesintehnologyandteleom-
†
ThishapterisbasedonGilli,Cabej,Lula,andShumann(2012).
†
WethankEvisKëllezifordisussionsandomments.
muniations had a inuene in the growth of the ativity in this market. This
is also due to theintrodution, in the early 1990s, of a tehnologial innovation:
eletroni trading. Nowadays,eletroni tradingrepresentssome50%ofthetotal
volume (BIS, 2010) and it has ompletely hangedthe struture of the markets
failitatingtheinterationamongators andtheintegrationofnewinformation.
The foreignexhange market ishighly eient ompared to other nanialmar-
kets,oeringsigniantlyreduedprotopportunities. However,urrenymarkets
are strongly aetedby monetary poliies and entral bank interventions. These
are thesoure ofinformationasymmetrywhih leadsto, at leasttemporary,inef-
ienies.
2.2 Tehnial analysis in urreny markets
Tehnialanalysis(TA)isbasedonDow'stheory 1
,thatpriesmoveintrends,the
trendishistoriallyrepetitiveandthevolumeonrmsthetrend.
Tehnialanalysisusesinformation regardingthehistorial prieof aurreny or
seurityand triesto foreastfuture priebehavior. Kaufman(2005)denesteh-
nial analysis as the systemati evaluation of prie, volume, breadth and open
interest, for the purpose of prie foreasting. Aording to Lo, Mamaysky, and
Wang (2000),tehnialanalysisdiersfrom quantitativenane. Andthisdier-
eneliesinthefatthattherstismainlyvisual,resortingtopatternreognition,
while the latter is mainly algebrai and numerial, using tools from probability
andstatistis. Osler(2000)examinesfromaneonomiperspetivehowtehnial
analysis anpredittheexhangerates.
Tehnial analysis is one of the rst attempts to take advantageof the market's
ineienies. For this reason many eonomists dene tehnial analysis as irra-
tional and despite the undeniable widespreadof TA among investors, aademis
havebeenquite septialon thesubjet as TA protability goesagainst thee-
ientmarkethypothesis(EMH).Indeed,literatureoftenreferstoTAasatestfor
EMH.
Fromthebeginnings,tehnialanalysisplayedadominantroleinurrenytrading
and still remains an important tool in the investors' deision making proess.
2
Aording to Henderson (2006) the tehnial analysis has an important role in
prediting exhange rates beause the urreny market follows a trend in short
term and there are speulators in this market. The rst examples of tehnial
1
writtenbyCharlesH.DowontheWallStreetJournal
2
AordingtoTaylorandAllen(1992)about90%ofurrenydealersusesomehartistinput.
trading rules (TTR) an be traed bak to Alexander (1961, 1964) who applied
lterrules 3
totestfortrendsin thestokmarkets. FamaandBlume(1966)show
that,inthisframework,ifdividendsandtradingostsaretakenintoaount,lter
rules do not perform well. However historially, in the urreny markets, lter
rulesyieldedenouragingresults. DooleyandShafer(1983)provideevideneofthe
outperformaneofsmall lters,i.e. hanges inthe orderof1to 5%,with respet
to buy-and-holdstrategies. Further, Sweeney (1986)shows that, for lter below
2%, strategies remain protable after adjusting for transation osts. Nowadays,
thereexistsalargesetofmoresophistiatedtehniques,inludingstrategiesbased
onstop-lossorders,movingaverages,hannels,momentumosillatorsandrelative
strengthindies. Toidentifytradingrules,algorithmiapproahes,suhasGeneti
Programminghavebeenlargelyexplored(seeDaorogna(1993)forearlyworkand
Neely,Weller,andDittmar(1997)foramoreompletesurvey).
A quite thorough literaturereview an be foundin Menkho and Taylor(2007),
ParkandIrwin(2007). Inanattempt toexploretheprotabilityofthestrategies
mentionedabove,ParkandIrwin(2007)lassifytheempirialevideneinto'early'
and'modern'studies. Therstonestendto showthattehnialtradingrulesare
protable in urreny markets alone, whilethe'modern' onesrevealprotsin all
markets,atleastuntilthebeginningofthe1990s. Apopularpaperinthisategory
isBrok,Lakonishok,andLeBaron(1992)whouseabootstrapproedureandtwo
tradingsystems(one basedonmovingaverageosillators andanotheron trading
range break)in the stokmarket.
4
Marsh (2000)and Olson(2004)highlightthe
delinein protabilityforpreviouslysuessfulstrategies. Thisiswhyweexplore
if anasset alloation approah to the FXmarket mightproveprotable. Toour
bestknowledgelittleliteraturerelatedtosuh anapproahexists.
Thehapterisorganizedasfollows: Setion2.3desribestheassetalloationmod-
elsandapartiularlterruletestedforthetehnialtradingapproah. Empirial
resultsaregiveninSetion2.4wherewealsopresentthedataaswellasthebak-
testingsheme. ArobustnesstestispresentedinSetion2.5. Setion2.6onludes.
3
Alterruleonsistsinabuyorsellsignalresultingfromagivenperentage hangeinthe
prieproess.
4
For an appliation of bootstrap in the foreignexhange markets see Levih and Thomas
(1993).
2.3 The models
2.3.1 Asset alloation approah
Inthis asethebuying andsellingdeisionsare taken asin ativeportfolio man-
agement. Thelassialassetalloationapproahonsistsinmaximizingameasure
of reward for agiven risk level. In the early work of Markowitz (1952)portfolio
rewardis measuredbyexpeted returnand portfoliorisk byitsvariane, this, in
partiular, foromputationalonveniene. Theaveatsof suh portfoliosare well
knownand havebeen extensivelydisussed in theliterature(Mihaud,1989; Co-
hen and Pogue,1967; Jobson and Korkie, 1980;Best and Grauer,1991). Having
explored,withsomesuess,avarietyofalternativeobjetivefuntions (Gilliand
Shumann, 2011b;Gilli et al.,2011b, 2010)we usefuntions suh as theOmega
ratio, drawdown,momentum and ombinations ofthese. The objetivefuntions
are omputed from portfolio returns, either historial or simulated. The models
are speied in inreasing order of omplexity by adding objetives whih might
beonurringorompeting. Thisorrespondstowhat isknownasmultiobjetive
optimization.
Theassetalloationproblemisformalizedintheusualway,i.e.
min x Φ(x)
(2.1)X
j ∈J
x j p 0j = v 0
(2.1′
)
x inf j ≤ x j ≤ x sup j j ∈ J
(2.1′′
)where
Φ(x)
istheobjetivefuntionwithx
thevetorofinvestedquantities.More- overwehavethelassialbudgetonstraint(2.1′
)andonstraint(2.1
′′
)limitsmin-
imumandmaximumholdingsize. Notethat
v 0
istheinitialapitalinUS dollars.A rstobjetivefuntion usedfor oururrenyportfoliosismaximumdrawdown
dened as
max(D t )
whereD t = v t max − v t
with
v t
,t = 0, 1, 2, . . .
the wealth path of the urreny portfolio andv max t
therunningmaximum,i.e.
v max t = max { v s | s ∈ [0, t] }
.Aseond objetivefuntion weinvestigatedarepartial momentsdened as
M lo = Z r d
−∞
(r d − r) m lo f (r)dr
andM up = Z ∞
r d
(r − r d ) m up f (r)dr
with
r d
areturntarget andm lo
,m up
theorder ofthe moment. Returndensitiesf ( · )
arenotknownandinordertoevaluatetheintegralsweresorttohistorialorsimulatedportfolioreturns
r
. WethenhaveM lo = 1 n S
n S
X
s=1
(r d − r s ) m lo 1 { r s <r d }
andM up = 1 n S
n S
X
s=1
(r s − r d ) m up 1 { r s >r d }
where
1
is the indiator funtion. In our appliationr d
is set to zero and themomentsareoforderone,thustheobjetivefuntionbeomes
Omega = M lo
M up
= − P
r s 1 { r s <0 }
P r s 1 { r s >0 }
.
Athirdobjetivefuntiononsidersmomentumdierentialsforthehistorialport-
folios. Thefuntion tobemaximizedis
Mom t i = (v t i − v t i − s )
| {z }
shortmomentum
− (v t i − v t i − ℓ )
| {z }
longmomentum
where
v t i
is the value of the portfolio at the rebalanement datet i
. The timeintervalsmeasuringthelongandshortmomentsare
ℓ
,respetivelys
.The kind of problems presented above generally annot be solved with lassial
methods, beause of nononvexities in the objetive funtion and/or the on-
straints. Therefore we resort to heuristi tehniques, in partiular threshold a-
epting, whih is a loal searh algorithm that only performs objetive funtion
evaluationsand doesnotrequiretheevaluation ofderivatives. Ageneraldesrip-
tionofthresholdaeptinganbefoundin Winker(2001),implementationdetails
arepresentedinGilliandWinker(2009);Gillietal.(2011a)andadisussionabout
thestohastisofthesolutionispresentedin GilliandShumann(2011a).
Heuristis are omputationally intensive and in order to redue exeution times
omputationshavebeendistributedoveralusterof32mahines 5
usingMatlab's
ParallelComputingToolbox.
5
MyrinetClusteroftheUniversityofGeneva(http://sp.unige.h/m yrin et) .
2.3.2 Tehnial trading approah
As mentionedin Setion 2.2 there exist avariety oftehnial tradingrules. This
appliationislimitedtotheuseoflterrulesgeneratingbuyandsellsignalsbased
ontrendidentiation. Thealibrationof themodelisnotbasedonoptimization
but the parameters are rather seleted through intuition, observation and expe-
riene. Indeed in most ases optimization of tehnial trading models leads to
overtting,i.e. goodin-sampleperformanebutpoorout-of-samplepredition. As
a onsequene of the absene of optimization, these tehniques neessitate little
omputationaleort.
Inthefollowingwewillmodelthetrendfortheomingholdingperiod. Todothis
we investigate the empirial distribution of the urreny pairs identifying three
states, i.e. positive,negativeand neutralor zero return. Theneutralstate fora
urrenyisdenedbyaurrentreturnlyingbetweenagivenupperandlowerem-
pirialquantile. Returnbelowthelowerquantileharaterizeanegativestateand
thoseaboveapositiveone. Inourpatternreognitionexperimentweusesequenes
Q_L Q_U
0.25 0.75
neutral
negative positive
Figure2.1:Empirialdistributionofpooledurrenypairsreturns.
of3statestodenepartiulartrends. Aneutralstateinthesequenewillbeinter-
pretedasaontinuationofthepreviousstate. Thisyieldsthefollowing8possible
sequenes. Thebuy signalsforeah urrenypair translateinto atransation of
axedamountin theorder ofseveralperentoftheinitialwealth. Thereforethe
portfoliowillbeonlygraduallyinvested. Similarlyasellsignalwillleadto axed
redution of the invested position. This allowsthe portfolio to gradually swith
betweensituationsoffullandpartial investment.
Table2.1:Setofpossiblesequenes.
t 0 − 2 t 0 − 1 t 0 t 0 + 1
Signalր ր ր
uptrend buyր ր ց
trendreversal holdր ց ր
trendreversal holdր ց ց
downtrend sellց ր ր
uptrend buyց ր ց
trendreversal holdց ց ր
trendreversal holdց ց ց
downtrend sell2.4 Empirial results
2.4.1 Data and baktesting sheme
The data have been downloadedfrom www.oanda.omand omprisetik-by-tik
time series of two sets of ve urreny pairs, whih are EUR/USD, GBP/USD,
CAD/USD, CHF/USD and JPY/USD. One set overs the period from 31-De-
2007to 31-De-2008andtheseondtheperiodfrom01-Jan-2010to31-De-2010.
Thenumberofdatapointsexeedssome10millionobservationsforeahurreny
pair. Inoursimulationthemidquoteshavebeenonsidered.Thedataarenotused
tik-by-tikbutathourlyordailyfrequeny. Forthemodellingproessweonsider
themedianoftheintrahour,respetivelyintradaytikswhileforrebalaningthe
last tik is used. Figure 2.2 shows the plot of the mid prie time series for all
urreny pairs of the rst set (year 2008). For better readability alltime series
startatone. Figure 2.3showstheintra hourvolatilityforthesameseries.
Dec07 Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08 Jan09 EUR 1
GBP 1 CAD 1 CHF 1 YEN 1
Figure2.2: Saledmidquotesofurrenyrates.
Dec07 Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08 Jan09 EUR
GBP CAD CHF YEN
Figure2.3: Intrahourlyvolatilityofurrenyrates.
Theperformaneofthesuggestedportfoliosisbaktestedoverthewholedata set
using a rolling window tehnique. The window is partitioned into an historial
segment of length
H
, used for the alloation deisions, and a holding period oflength
F
wheretheobservationsareusedtoassesstheout-of-sampleperformane.Theshemebelowsummarizesthetehnique.
Period
1
t 1 − H t 1 t 1 + F
F
H
invest
Period
2
t 2 − H t 2 t 2 + F
rebalane
In order not to rely on simple point estimates of the simulated performane we
alwaysompute a set of 100 simulationsfrom jakknifedhistorial observations.
The jakkning onsist in removing randomlysome 10%of datapointsfrom the
historialwindow. Theresultsobtainedfromthe100simulationsaresummarized
withthemedianpathwhihorrespondstothewealthpathoftheportfoliohaving
themediannalwealth. Wealsogivetheempirialdistributionsofthenalwealth.
Simulation results are ompared with a benhmark portfolio whih is the result
of a buy-and-hold strategy. The initial alloation of the benhmark portfolio is
proportional to the turnover of the urrenies in 2007 (BIS (2007)) and results
into the following weight vetor
[0.44 0.20 0.07 0.08 0.21]
. We also omputed a1 / N
benhmark portfolio. The wealth path of this portfolio is very lose to thebuy-and-holdbenhmarkandthereforeitisnotpresentedin theresults.
2.4.2 Simulation result for asset alloation approah
This approah lends itself to a large spetrum of speiations onsidering the
many possibilities of hoies for the aggregation level of the data, the length of
thehistorial window,rebalaningfrequenyand soon. An optimization overall
theseriteriaisnotfeasible(andnotwise)andoneisforedtoproeedwithsome
pragmatism.
Havingtested afewoptionsforthefrequenyofrebalanementsandtheaggrega-
tion levelof the prieswe retaineddaily rebalanementsand hourlyaggregation.
Inidentally,longerholdingperiods(foreasthorizons)implyfewerrebalanements
forthebaktesting. Thistranslatesintolowertransation ostsand,from aom-
putational pointof view,in asmallernumberofoptimizations.
To begin weompare the median wealth path and the nal wealth densities for
the models presented above, whih are maximum drawdown (
DD
), Omega andmomentum driven portfolios (
Mom
). The historial window has a length of 80hoursand rebalanementsourdailyat 9am. All positionsin theportfoliosare
onstrainedtoabuy-inthresholdof5%andamaximumsizeof70%. Theminimum
trading size for eah position is 1000 USD whih implies that the portfolio may
remainunhangediftherebalanementssuggestedbytheoptimizationareinferior
to this threshold. Transation osts are set to 2.5 bp. Figure 2.4 illustrates the
resultsthatatthis stagearenotappealing.
The wealth pathsin Figure 2.4 reveal an alternationbetweenprotable and un-
derperformingperiods. This motivatesthe introdutionof the possibility forthe
portfolio toget partially orompletely divested, ahievedbyintroduing anarti-
ialpositionin USD.Forthisposition weallowamaximumweightof100%,i.e.
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08 0.8
0.85 0.9 0.95 1 1.05 1.1
DD Omega Mom Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 15−Jan−2008 08:00:00 −− 31−Dec−2008 DD Omega Mom Bench
Figure2.4:Wealthpathsandnalwealthdistributionfor
DD
,OmegaandMom
.Hourlypries, dailyrebalanementsand80datapointsforthehistorialwindow.the portfolio mightbe fully or partially in ash. The resulting performanes are
shownin Figure2.5. Weobservethat drawdown andOmegaimprovein termsof
nal wealth densities. Fordrawdownin partiular, weobserveverylowvolatility
eventhough its behavior might look trivial, as we stay in ash for long periods
(the wealth path appears as straight line for most of the time). However mini-
mizing drawdown preserves the initial apital even in periods of adverse market
onditions, where the benhmark and all other strategiesare nonprotable, thus
drawdownseemstobeausefulomponentof aninvestmentmodel.
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.8 0.85 0.9 0.95 1 1.05 1.1
DD Omega Mom Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 15−Jan−2008 08:00:00 −− 31−Dec−2008 DD Omega Mom Bench
Figure2.5: Wealthpathsandnalwealthdensitiesfor
DD
,OmegaandMom
. Theportfoliois allowedtogofullyintoashposition.Inthefollowingwewilloptforlessonservativestrategies,thatiswewillnotallow
ourportfoliostobeompletelyoutofthemarket. Thusashwillbesubjettothe
sameholding size onstraintsas theother assets. As aonsequenetheportfolio
willbeatleastinvested30%.
Onemightwonderwhyaportfoliowhihminimizesdrawdownanenteraposition
otherthanash,asthereisnodrawdownwithash. Thereexisttworeasonsforthis,
one is market dependent, i.e., historial diversied portfolios without drawdown
mayexistandbehosenbytheoptimization;theotherdependsontheoptimization
tehnique: weinterrupttheoptimization proedure afteragivennumberofsteps
whihdoesnotneessarilyproduetheglobaloptimum. Thisis doneonpurpose
in ordertoavoidovertting.
Filtering out noise
It isawellknownfat thatnanialdata ontainmuh noise. Inafurther inves-
tigation we analyzethe impat that ltering outnoise from the data has onthe
performane of the portfolios. Filtering is done by approximatingthe matrix of
observedquotes byasumof rank-one matriesomputedfrom the leftand right
singular vetors of the singular value deomposition. Comparing the results for
fully invested portfolios in the upper panels of Figure 2.6 with 2.4 reveals that
ltering datasigniantlyenhanesoverall performane forallstrategies. This is
alsotheaseforthepartiallyinvestedportfolios(seelowerpanelin Figure2.6).
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.8 0.85 0.9 0.95 1 1.05 1.1
DD Omega Mom Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 16−Jan−2008 08:00:00 −− 30−Dec−2008 DD Omega Mom Bench
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.8 0.85 0.9 0.95 1 1.05 1.1
DD Omega Mom Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 16−Jan−2008 08:00:00 −− 30−Dec−2008 DD Omega Mom Bench
Figure2.6: Wealthpathsand nalwealth densitiesfor
DD
,OmegaandMom
forltereddata.Upperpanel: portfoliofullyinvested.Lowerpanel: maximumash70%.
Combining objetives
The nal wealth densities show that drawdown dominates. However when in-
spetingthe wealth path,theother strategiesshow,forertain periods, desirable
properties. Therefore, ombiningseveralobjetivesshould resultinbetteroverall
performane. Inthefollowingourobjetivefuntion
Φ(x)
willbeafuntionofthesinglestrategiesdesribedearlier,i.e.
Φ(x) = Ψ( D t , Omega, Mom t ) .
where
Ψ
isalinearombinationofdrawdown,Omegaandmomentum. Figure2.7illustratesresultsforsuh ombinations. Onlydrawdowninonjuntionwithmo-
mentumperformswell. ThereasonwhyOmegadoesnotontributeinperformane
mightresideinthefatthatOmegaisnotpathdependentbut
Mom
anddrawdownare. Also,Omegadoesnotusetheinformationprovidedbythewealthpath, that
drawdownandmomentumdo.
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.8 0.85 0.9 0.95 1 1.05 1.1
DDMom DDOm DDOmMom Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 15−Jan−2008 08:00:00 −− 31−Dec−2008 DDMom
DDOm DDOmMom Bench
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.8 0.85 0.9 0.95 1 1.05 1.1
DDMom DDOm DDOmMom Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 15−Jan−2008 08:00:00 −− 31−Dec−2008 DDMom
DDOm DDOmMom Bench
Figure 2.7: Wealth paths and nal wealth densities for ombined objetivesand ltereddata.
Upperpanel: fullyinvestedportfolio. Lowerpanel: partiallyinvestedportfolio.
We havetwoalternatives,either bealwaysfullyinvested or allowash positions.
The rstoptionyieldshigher average nal wealth,see upperpanelof Figure 2.7,
theseond ismoreonservative,yieldinglessreturn,but alsolowervolatilityand
almostalwayspreservestheinitialwealth(lowerpanelofFigure2.7).
Inreasing holdingperiods/short positions
Retaining the best performingstrategies from above weontinue to enhane our
portfoliosbyinreasingthelengthoftheholdingperiods (i.e.reduingrebalane-
ment frequeny), introduing simulated priesfor the optimization and allowing
for short positions up to 30 perent of the value of the portfolio. The numeri-
alproedureforomputingsuhportfoliosremainsunhanged. Notethat,unlike
stokmarkets,urrenymarketsdonotriskshortsqueezesbuttheinvestorisstill
exposedto interestratemovements. Thisis why alimithasbeenimposed tothe
short book. Thepriegeneration proessfollowsideasfromDembo(1991)andis
explainedin moredetailin GilliandShumann(2011).
Altogether this leads to a signiant improvement for nal wealth as well as to
lowernal wealth volatility. Figure 2.8 summarizesthese resultsfor thestrategy
ombiningdrawdownwithmomentum. Bestresultsareobtainedfor theportfolio
allowingshortpositionswhihreahesamediannal wealth ofabout12(beating
thebenhmarkby17).
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.9 0.95 1 1.05 1.1
1.15 DDMom
Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 15−Jan−2008 08:01:00 −− 31−Dec−2008 DDMom
Bench
Feb08 Apr08 May08 Jul08 Aug08 Oct08 Dec08
0.9 0.95 1 1.05 1.1
1.15 DDMom
Bench
−15 0 −10 −5 0 5 10 15
0.1 0.5 0.9 1
Period returns: 15−Jan−2008 08:00:00 −− 31−Dec−2008 DDMom
Bench
Figure2.8:Wealthpathsandnalwealthdensitiesforombinedobjetives,simulatedpriesand2
rebalanemetsperweek. Upperpanel:Longonlyportfolios. Lowerpanel: Long/shortportfolios.
2.4.3 Simulation results for tehnialtrading
As explained in Setion 2.3.2 the model generates `buy', `sell' and `hold' signals
whiharetranslatedinto xed-sizetradingorders. Asin theassetalloationase
thereisaminimumtradingsize.Theminimumandthemaximumholdingsizeare
thoseusedin theoptimizationapproah. Thereisnoadditionalonstraintforthe