Centrum voor Wiskunde en Informatica
REPORT
RAPPORT
Wavelet Methods in (Financial) Time-series Processing
Z.R. Struzik
Information Systems (INS)
CWI
P.O. Box 94079
1090 GB Amsterdam
The Netherlands
CWI is the National Research Institute for Mathematics
and Computer Science. CWI is part of the Stichting
Mathematisch Centrum (SMC), the Dutch foundation
for promotion of mathematics and computer science
and their applications.
SMC is sponsored by the Netherlands Organization for
Scientific Research (NWO). CWI is a member of
ERCIM, the European Research Consortium for
Informatics and Mathematics.
Copyright © Stichting Mathematisch Centrum
P.O. Box 94079, 1090 GB Amsterdam (NL)
Kruislaan 413, 1098 SJ Amsterdam (NL)
Telephone +31 20 592 9333
Telefax +31 20 592 4199
ZbigniewR.Struzik
CWI
P.O.Box94079,1090GBAmsterdam,TheNetherlands
email: Zbigniew.Struzik@cwi.nl
ABSTRACT
Webrie ydescribethemajoradvantagesofusingthewavelettransformfortheprocessingof nancialtime
seriesontheexampleoftheS&Pindex. Inparticular,weshowhowtouncoverlocalthescaling(correlation)
characteristicsoftheS&PindexwiththewaveletbasedeectiveHolderexponent[1,2 ]. Weuseittodisplay
the local spectral (multifractal)contents of the S&Pindex. In addition to this, we analyse thecollective
propertiesofthelocalcorrelationexponentasperceivedbythetrader,exercisingvarioustimehorizonanalyses
oftheindex.Weobservedanintriguinginterplaybetweensuch(dierent)timehorizons. Heavyoscillationsat
shortertimehorizonswhichseemtobeaccompaniedbyasteadydecreaseofcorrelationlevelforlongertime
horizons,seemtobecharacteristicpatternsbeforethebiggestcrashesoftheindex. Wendthatthiswayof
localpresentationofscalingpropertiesmaybeofeconomicimportance.
2000MathematicsSubjectClassication: 28A80,68T10,68P10,82D99
1999ACMComputingClassicationSystem: H1,I5,Jm,J2, E2
KeywordsandPhrases: econophysics,wavelettransform,Holderexponent,localcorrelation
Note: fullcolourversionofthispapercanbedownloadedfrom www.cwi.nl/~zbyszek
1. Introduction
Economicshasbeenproducingmoreandmorecomplicatedmodels,tryingtocapturedeviationsofthe
modelsituationfromreality. Butpatchingamodeltoincreaseitscomplexitymaynotbeanoptimal
wayofmodelling. Anyeconomicsystemisextremely complicatedbut, to alargedegree,thisis due
to the enormousnumber of degreesof freedom. Theinteractions betweeneconomic entities donot
needtobeverycomplicated. Approachesaimingat improvingthedescriptionofasystembymeans
ofhighordercorrectionsmay,therefore,havearatherslowconvergencetothesatisfactorymodel.
On the contrary, statistical physicsuses verysimple models of interactions between components
of a huge ensemble. Concepts from statistical physicshaveproven quite successful in the work of
econophysicists. Thepowerofmodern computersallowsnotonlythetesting ofsuchmodelsbutalso
theanalysis ofdatafor thepresenceofmultiscalingcharacteristics,in particularlocally in thedata.
Ananalysisofcorrelationscanbelinkedtothemodelsofinteractionbetweenthesystemcomponents,
andsocanthemultiscaledistributionanalysis.
Physicshasalongtraditionofdealingwithsystemswithanextremenumberofdegreesoffreedom,
where entities are coupled with very simpleinteraction rules. This is probably why economicshas
recentlyexperienced great interestamong physicists, followingthe pioneering work of Gene Stanley
et al[3],RosarioMantegna,[5],AlainArneodoet al,[4],MarcelAuslooset al[6] and,nottoforget,
theearlyworkofBenoitMandelbrot[7].
Whereas thephysicists derivethe nalmodelfrom thecharacteristicsofthe dataanalysed, using
theirlearnedphysicsknowledgeandscienticintuition, thelatestcomputersciencetrendistoassist
thedata analyserin model discovery. Data miningis thename forthis recentdirection in machine
data mining/model discoveryapproachesin theeconomicalsciences willgrowrapidlyin the coming
decade.
Therefore,workingwithoutanaprioriassumedmodelischaracteristicofthemodernapproachto
economics. Instead,themodelistobeinferredfromthedata. Thedataisanalysedintermsofvery
genericanalysismethodslike,forexample,waveletdecomposition. Thewavelettransformcomponents
are then analysed and, of course, simple or complex models can be tted to such decomposition
components. Thescalingofmomentsordistributionscanbetestedandahypothesisdrawn,butthe
rststepis dataanalysisinthemostgenericterms.
Thepossibilityofdoinganalysislocallyisanotherveryattractiveoption. Thestationarityofalmost
any statisticalcharacteristicfailswhen appliedto thenancialdata. Asimilar situationpertains in
other complex phenomena (take, for example, the human heartbeat [9, 10]). But where the
non-stationarities occur, interestingness begins, and with tools like the wavelet transform, capable of
tamingthenon-stationarities(trends),interesting(local)patterns canbediscoveredinthedata.
1
In section 2, we brie y introduce the wavelet transformation in its continuous form, we describe
therequirementsforthewaveletinsection3anddiscusstheadvantagesfortimeseriesprocessing. In
particular,wefocusin section4ontheabilityofthewavelettransformationtocharacterisescale-free
behaviourthroughtheHolderexponent. Wedescribeinbriefatechnicalmodelenablingustoestimate
thescale-freecharacteristic(theeectiveHolderexponent)forthebranchesofamultiplicativeprocess.
Amoreextensivecoverageofthismethodisavailablein[1,2]. Insection5,weusethederivedeective
HolderexponentforthelocaltemporaldescriptionoftheS&Pindex. Section6providesanextension
to the uctuation analysis of the eective Holderexponent of theS&P index. Section 7 closesthe
paperwithconclusions.
2. Why Wavelets?
Figure1: ContinuousWaveletTransformrepresentationoftherandomwalk(Brownianprocess)time
series. The wavelet used is the Mexican hat - the second derivative of the Gaussian kernel. The
coordinateaxesare: positionx,scalein logarithmlog(s),andthevalueofthetransformW(s;x).
1
Thewavelettransformis aconvolutionproductofthesignalwiththescaledandtranslatedkernel
-thewavelet (x).[11,12]Thescalingandtranslationactionsareperformedbytwoparameters;the
scaleparameters`adapts'thewidthofthewaveletkerneltotheresolutionrequiredandthelocation
oftheanalysingwaveletisdeterminedbytheparameterb:
(Wf)(s;b)= 1 s Z dxf(x) ( x b s ) (2.1)
wheres;b2Rand s>0forthecontinuousversion(CWT).
The3Dplotingure1showshowthewavelettransformrevealsmoreandmoredetailwhilegoing
towardssmallerscales,i.e. towardssmallerlog(s)values. Thewavelettransformissometimesreferred
toasthe`mathematicalmicroscope'[4],duetoitsabilitytofocusonweaktransientsandsingularities
inthetimeseries. Thewaveletuseddeterminestheopticsofthemicroscope;itsmagnicationvaries
withthescalefactors.
WhetherwewanttousecontinuousordiscreteWT,seegure2,islargelyamatterofapplication.
Forcoding purposes, onewantsto usethesmallest numberofcoeÆcients which canbecompressed
bythresholdinglowvaluesorusingcorrelationproperties. Forthispurposeadiscrete(forexamplea
dyadic)schemeofsamplingthescales,positionb spaceisconvenient. Suchsampling oftenspansan
orthogonalwaveletbase.
For analysis purposes, oneis not somuch concernedwith numerical ortransmissioneÆciency or
representationcompactness,but ratherwithaccuracy andadaptivepropertiesoftheanalysing tool.
Therefore, in analysis tasks, continuouswavelet decomposition is mostlyused. Thespace of scales
andpositionb,isthensampledsemi-continuously,usingthenestdataresolution available.
b
s
b
s/2
s/4
s/8
b/8
Figure2: Continuoussamplingoftheparameterspace(left)versusdiscrete(dyadic)sampling(right)
For decomposition,a simplebasis function is used. The wavelet , see Eq. 2.1 took its name from
its wave-likeshape. It has to cross the zero value line at least once since its mean value must be
zero. Thecriterion of zeromean is referredto asthe admissibility ofthe wavelet, and is related to
thefact thatonewantstohavethepossibilityofreconstructingtheoriginalfunctionfromitswavelet
decomposition.
signalbeingdecomposed. Thisistheso-calledbandpasslteringofthesignal. Onlyacertainband
offrequencies(levelofdetail)iscapturedbythewaveletsworkingatonescale. Ofcourse,atanother
scaleadierentset ofdetails (band offrequencies) iscaptured. Butother frequencies(inparticular
zerofrequency)arenottakeninto thecoeÆcients. Thisistheideaofdecomposition.
Kernels like the Gaussian smoothing kernel are low-pass lters, which means they evaluate the
entireset of frequenciesup to the currentresolution. This is the ideaof approximationat various
resolutions.
The reader may rightly guess here that it is possible to get band pass information
(wavelet-coeÆcients) from subtracting two low-pass approximations at various levels of resolution. This is,
in fact,aso-calledmulti-resolutionschemeof decomposition intoWT components. Butback tothe
admissibility-reconstructionfrommultipleresolutionapproximationswouldnotbepossiblesincethe
samelowfrequency detailwould bedescribed in several coeÆcientsof the low passdecomposition.
This, ofcourse,is notthecaseforwavelets;theyselectonlyanarrowband ofdetailwithverylittle
overlap(in theorthogonal casenooverlapat all!). Inparticular, if one requires that thewavelet is
zeroforfrequencyzero,i.e. itfullyblockszerofrequencycomponents,thiscorrespondswiththezero
meanadmissibilitycriterion.
3. The Wavelet
Theonlyadmissibility requirementforthewavelet isthatithaszeromean-itisawavefunction,
hencethenamewavelet.
Z 1
1
x (x)dx=0 (3.1)
However, in practice, wavelets are often constructed with orthogonality to a polynomial of some
degreen. Z 1 1 x n (x)dx=0 (3.2) Mexican Hat
-0.4
-0.2
0
0.2
0.4
0.6
-0.1
-0.05
0
0.05
0.1
scaling function
-0.4
-0.2
0
0.2
0.4
-0.1
-0.05
0
0.05
0.1
wavelet, m=1
-0.6
-0.4
-0.2
0
0.2
0.4
-0.1
-0.05
0
0.05
0.1
wavelet, m=2
Figure 3: Left: thesmoothingfunction: theGaussian. Centre: waveletwithonevanishingmoment,
therstderivativeoftheGaussian. Right: waveletwithtwovanishingmoments,thesecondderivative
oftheGaussian.
Thispropertyofthewavelets-orthogonalitytopolynomialsofdegreen-hasaveryneapplication
insignalanalysis. Itisreferredtobythenameofthenumberofvanishingmoments. Ifthewaveletis
degreeP 0
. This canbe done, forexample, by the rstderivative of theGaussian kernelplotted in
gure3. SimilarlythesecondderivativeofthesameGaussiankernel,whichisoftenusedupsidedown
and thenappropriatelycalled theMexicanhatwavelet,hastwovanishing momentsandin addition
to constantscan also lterlinear trendsP
1
. Ofcourse, if thewavelethas m vanishingmoments, it
canlterpolynomialsofdegreem 1,m 2,... 0.
4. The
H
older Exponent
Theuseofvanishingmomentsbecomesapparentwhenweconsiderlocalapproximationstothe
func-tiondescribingourtimeseries. Supposewecanlocallyapproximatethefunctionwithsomepolynomial
P n
,buttheapproximationfailsforP n+1
. OnecanthinkofthiskindofapproximationastheTaylor
se-riesdecomposition. InfacttheargumentstobegivenaretrueevenifsuchTaylorseriesdecomposition
doesnotexist,butitcanserveasanillustration.
For the sakeof illustration, let us assume that the function f can be characterised by the Holder
exponenth(x
0 )inx
0
,andf canbelocally describedas:
f(x) x0 =c 0 +c 1 (x x 0 )++c n (x x 0 ) n +Cjx x 0 j h(x 0 ) =P n (x x 0 )+Cjx x 0 j h(x 0 ) : Theexponenth(x 0
) iswhat `remains' after approximatingwith P
n
andwhat doesnotyet `t' into
anapproximationwithP
n+1
. Moreformally,ourfunctionortimeseries f(x)islocally describedby
thepolynomialcomponentP
n
andtheso-calledHolderexponenth(x
0 ) . jf(x) P n (x x 0 )jCjx x 0 j h(x0) : (4.1)
It is traditionally considered to beimportant in economics to capture trend behaviourP
n . It is,
however,widelyrecognisedinothereldsthatitisnotnecessarilytheregularpolynomialbackground
but quite often the transient singular behaviour which can carry important information about the
phenomena/underlying system`producing'thetimeseries.
One of the main reasons for the focus on the regular component was that until the advent of
multi-scaletechniques(likeWT)capableoflocallyassessingthesingularbehaviour,itwaspractically
impossibletoanalysesingularbehaviour. Thisisbecausetheweaktransientexponentshareusually
completelymaskedbythemuch strongerP
n .
However,waveletsprovidea remedyin this case! The reader hasperhapsalready notedthe link
with thevanishing moments ofthe wavelets. Indeed, if thenumberof the vanishing momentsis at
leastashigh as thedegree ofP
n
, thewavelet coeÆcientswill capturethe localscaling behaviourof
thetimeseriesasdescribedbyh(x
0 ).
Infact,thephrase`ltering'withreferencetothepolynomialbiasisnotentirelycorrect. Theactual
lteringhappens onlyfor waveletsthe support of which is fully incorporatedin thebiased interval.
Ifthewaveletisattheedgeofsuchaninterval(wherebiasbegins)orifthecurrentresolutionofthe
waveletissimplytoolargewithrespecttothebiasedinterval,thewaveletcoeÆcientswillcapturethe
informationpertinenttothebias. Thisisunderstandablesincetheinformationdoesnotget`lost'or
`gained'in theprocessof WTdecomposition. The entire decomposed function canbereconstructed
from the wavelet coeÆcients,including the trendswithin the function. Whatwaveletsprovidein a
uniquewaytotameand managetrendsin alocalfashion, throughlocalisedwaveletscomponents.
Above, we have suggested that the function can locally be described with Eq. 4.1. Its wavelet
transformW
(n)
f withthewaveletwithat leastnvanishingmomentsnowbecomes:
W (n) f(s;x 0 )= 1 s Z Cjx x 0 j h(x 0 ) ( x x 0 s )dx=Cjsj h(x 0 ) Z jx 0 j h(x 0 ) (x 0 )dx 0 :
Therefore,wehavethefollowingpowerlawproportionalityforthewavelettransformofthe(Holder)
W (n) f(s;x 0 )jsj h(x0) :
From the functional form of the equation, one can immediately attempt to extract the value of
thelocalHolderexponentfromthescalingofthewavelettransformcoeÆcientsinthevicinityofthe
singularpointx
0
. This is indeed possiblefor singularities which are isolatedoreectivelyisolated,
that isthat canbeseenasisolatedfrom thecurrentresolution oftheanalysingwavelet. Acommon
approachtotracesuchsingularitiesandtorevealthescalingofthecorrespondingwaveletcoeÆcients
istofollowtheso-calledmaximalinesoftheCWTconvergingtowardstheanalysedsingularity. This
approachwasrstsuggestedbyStefanMallatet al[13]andlaterusedandfurtherdevelopedamong
othersinRefs[14, 4,15].
Ingure4,weplottheinputtimeserieswhichis apartoftheS&Pindex containingthecrashof
'87. In thesame gure,weplot correspondingmaxima derivedfrom theCWT decomposition with
the Mexican hatwavelet. The maxima correspondingto the crash stand outboth in the top view
(they are the longestones) and in theside log-logprojectionof allmaxima (they havea valueand
slopedierentfromtheremainingbulkofmaxima). Theonlymaximahigherin valuearetheendof
thesamplenitesizeeectmaxima. Theseobservationsindicatethat thecrashof'87canbeviewed
asan isolatedsingularity in theanalysed recordof theS&P index for practicallytheentire wavelet
rangeused.
-2
0
2
4
6
8
10
12
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
minima
maxima
crash ’87 maxima
log(s)
x
input time series(x)
S&P 500
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
edge singularity maxima
crash singularity maxima
Wf(log(s))
log(s)
Figure4: Left: theinputtimeserieswiththeWTmaximaaboveitinthesamegure. Thestrongest
maximacorrespondtothecrashof'87. Theinputtimeseriesisde-biasedandL1normalised. Right:
weshowthesamecrashrelatedmaximahighlightedintheprojectionshowingthelogarithmicscaling
ofallthemaxima.
This is, however,(luckily)an unusual event andin general in time serieswehavedensely packed
singularities which cannot be seenasisolatedcasesfor awiderrangeof wavelet scales. Therelated
Holder exponent can then be measured either by selecting smaller scales or by using some other
approach. A possibility we would like to suggest is using the multifractal paradigm in order to
estimatewhatwecall theeectiveHolderexponent. Thedetailed discussionofthisapproachcanbe
foundin[1,2],butletusquicklypointoutthattheeectiveHolderexponentcaptureslocaldeviations
fromthemeanscalingexponentofthedecompositioncoeÆcientsrelatedtothesingularityinquestion.
Densesingularitiescanbeseenasevolvingfromamultiplicativecascadingprocesswhichtakesplace
acrossscales. TheCWThasbeensuccessfullyused in revealingsuchaprocess andin recoveringits
characteristics. Inshort,allthatis thenneededtoevaluatethelocaleectiveHolderexponentfora
singularity ataparticular scaleis thegainin process density acrossscales withrespectto thescale
gain. ^ h shi s lo = log(Wf! pb (s lo )) log(Wf! pb (s hi )) log(s lo ) log(s hi ) ; where Wf! pb
(s) is the value of the wavelet transform at the scale s, along the maximum line !
pb
correspondingto thegivenprocessbranch bp. Scales
lo
correspondswithgenerationF
max
,whiles hi
correspondswithgenerationF
0
,(simplythelargestavailable scaleinourcase).
5. Employing theLocalEffective
H
olderExponentinthe Characterisationof Time
Series
Suchanestimatedlocal
^ h(x
0
;s)canbedepictedinatemporalfashion,forexamplewithcolourstripes,
aswe havedone in gure 5. The colour of the stripesis determined by the value of the exponent
^ h(x
0
;s)anditslocationissimplythex 0
locationoftheanalysedsingularity(inpracticethisamounts
tothelocationofthecorrespondingmaximumline). Colourcodingisdonewithrespecttothemean
value,whichis setto thegreencolourcentralto ourrainbowrange. All exponentvalueslowerthan
themeanvaluearegivencoloursfromthe`warmer'sideoftherainbow,allthewaytowardsdarkred.
Allhigherthanaverageexponentsget`colder'colours,downtodarkblue.
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
histogram eff. h of S&P index
h
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
histogram eff. h of fBm H=0.6
h
Figure 5: Left: example time series with local Hurst exponent indicated in colour: the record of
healthy heartbeat intervals andwhite noise. The backgroundcolour indicates the Holder exponent
locally,centredattheHurstexponentatgreen;thecolourgoestowardsblueforhigher
^
handtowards
redforlower ^
h. Right: thecorrespondinglog-histogramsofthelocalHolderexponent.
bothsmootherandroughercomponents. Inparticular,onecanobserveanextremalredvalueatthe
crash of'87 coordinate, followedbyveryroughbehaviour(a rather obviousfact, but to thebest of
our knowledge notreported to date in therapidly growingcoverage of this time series record), see
e.g.[16].
The second example time series is a computer generated sample of fractional Brownian motion
with H = 0:6. It shows almost monochromatic behaviour, centred at H =0:6; thecolour green is
dominant. There are, however, several instances of darker green and light blue, indicating locally
smoothcomponents.
It isimportantto notice that h=H
BrownianWalk
, theHolderexponentvalueequalto theHurst
exponentofanuncorrelatedBrownianwalk,correspondswithnocorrelationintimeseries.
2
Anideal
randomwalkwouldhaveonlymonochromaticcomponentsofthisvalue. Ofcourse,anideal,innitely
long record of fractional Brownian motion of H = 0:6 is correlated, but this correlation would be
stationaryinsuchanidealcaseandno uctuationsincorrelationlevel(incolour)wouldbeobserved.
Bythesameargument,wecaninterpretthevariationsin hasthelocal uctuationsofcorrelationin
theS&Pindex. Themoreredthecolour,themoreunstable,themoreanti-correlatedtheindex. And
themoreblue,themorestableandcorrelated.
Totherightofgure5,thelog-histogramsareshownoftheHolderexponentdisplayedinthecolour
panels. Theyaremadebytakingthelogarithmofthemeasureineachhistogrambin. Thisconserves
the monotonicity of the original histogram, but allows us to compare the log-histograms with the
spectrumofsingularities D(h). Thelog-histogramsare actuallyclosely relatedto the(multifractal)
spectraof the Holderexponent[2]. Themultifractal spectrum of the Holderexponent isthe `limit
histogram' D
s!0
(h) of the Holderexponentin the limitof innite resolution. Of coursewecannot
speakofsuchalimitotherthantheoreticallyand,therefore,alimithistogram(multifractalspectrum)
hastobeestimatedfrom theevolutionofthelog-histogramsalongscale. Fordetailssee [2].
Letuspointoutthatthewidthofthespectraaloneisarelativelyweakargumentinfavourofthe
hypothesis of the multifractality of the S&P index. The log-histogram of the S&P is only slightly
wider than the log-histogram of arecord of fractional Brownian motion of comparable length, see
gure5. An interestingobservationis,however,thatthecrashof'87isclearlyanoutlierinthesense
ofthelog-histogramofHolderexponent(andthereforein thesenseoftheMF spectrum). Theissue
ofcrashesasoutliershasbeenextensivelydiscussedbyJohansen[17]andrecentlybyL'vovetal[18].
Herewesupport thisobservationfrom anotherpointofview.
3
6. Discovering Structure Through the Analysis of Collective Properties of
Non-stationary Behaviour
Non-stationarities are usually seen asthe curse of the exact sciences, economics notexcluded. Let
us herepresentadierentopinion: where thenon-stationarities occur, interestingnessbegins!
Non-stationaritiescanbeseenasadeparturefrom some(usually)simple`model'. Forexample,this can
bethefailureofthestationarityoftheeectiveHolderexponent(seegure5).
Insomesense,theyindicatethatthesimplemodelusedisnotadequate,butthisdoesnotnecessarily
meanthatoneneedstopatchorreplacethislowlevelmodel. Onthecontrary,theinformationrevealed
bysuchalowlevelmodelmaybeused todetecthigher orderstructures. Inparticular,correlations
in thenon-stationaritiesmayindicate theexistenceof interestingstructures. An intriguingexample
of such an approach in the nancial domain is the work by A. Arneodo et al, where a correlation
structurein S&Pindexhasbeenrevealed[19].
The simplest way of detecting structure, we suggest, is detecting uctuations or the collective
behaviour of the local eective h. This has already been successfully applied in human heartbeat
analysis.[10]. Here wewillpresentsomepreliminary resultsfortheS&Pindex.
2
Theoreticallythisish=0:5,butnitesizesampleeectsusuallyaddsomedegreeofcorrelation,slightlyincreasing
thisvalue. 3
Thenon-stationarybehaviourinhcanbequantied,andforthispurposeweusealowpassmoving
average lter (MA) to detect/enhance trends. This processing is, of course, done on the Holder
exponentvaluesetfh
i
(f(x))g, notontheinputsignalf(x). An-MAlteringofnbaseisdenedas
follows: h MAn (i)= 1 n i=n X i=1 h i (f(x)); (6.1) whereh i
(f)arethesubsequentvaluesoftheeectiveHolderexponentof thetimeseriesf.
Let us now go back to the S&P index and its eective Holder exponent description. Dierent
windowlengthsin ourMA lterrepresentdierenthorizonsforthetrader. Iftheindex isallthatis
available,in orderto evaluatetheriskassociatedwiththetrading(orin otherwords,topredictthe
riskofanindexcrash),thetradermightwanttoknowhow`stable'themarket/indexisonadailyor
monthlytimescale. Infact acomparisonbetweenthetwoindicatorsofstabilitymightbeevenmore
indicative.
This is exactly what wehave done using twodierent time scales (twotradinghorizons) for the
MAsmoothing,seegure6. ThesmoothedinputistheeectiveHolderexponentoftheS&P index.
Itcorrespondscloselywiththelogarithmofthelocal volatilityandassuchitre ectsthestabilityof
themarket.
Wemadethefollowingobservationsfromthisexperiment: theshorttimehorizonMAshowsastrong
oscillatorypatternin collectivebehaviouroftheh. These oscillationshavealreadybeenobservedby
Y.Liu et al [16] and by N.Vandewalle et al. [6] This is, however, not log-periodic behaviour in our
resultsanditdoesnotconvergetoamomentofcrash. Whatcanperhapsbeusedinordertohelpthe
trader in evaluatingthe growingriskis the interplayof the various timehorizons. Thesecond MA
lterhasatimehorizontentimeslongerand itshowspracticallynooscillations. However,its value
decaysalmost monotonically, in themomentjust before thecrash,reachingthelevelofcorrelations
characteristicfor therandomwalk(seegure6rightinserts). Notethat the crashesthemselvesare
notvisible intheinsertplots. Letusrecallthat themain advantageoftheeectiveHolderexponent
abovesometraditional measuresof volatility isthat it describesthe locallevelof correlationin the
timeseries. Ifthevalueofhisbelowh=H
BrownianWalk
,thismeanswehaveananti-correlatedtime
serieswhichintuitivelycorrespondswitharatherunstableprocess. Thehaboveh=H
BrownianWalk
indicates presence of correlations and generallycan be associated with `stability'. Pleasenote that
theoscillationsinMA50beforethecrashesbringthecollectivehupanddownbetweenthecorrelated
andtheanti-correlatedregimes. SimilarlyMA500steadily decaystowardstheanti-correlatedregime
just beforethecrashes.
7. Conclusions
The local eective Holder exponent has been applied to evaluate the correlation level of the S&P
index locallyat anarbitraryposition (time)andresolution (scale). In additionto this, weanalysed
collective properties of the local correlation exponent as perceived by the trader exercising various
time horizonanalyses ofthe index. A movingaveragelteringof Holderexponentbasedvariability
estimateswasused to mimicthevarious timehorizon analysisof theindex. We observedintriguing
interplay betweendierent timehorizons before thebiggest crashesof theindex. Wend that this
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
MA50(h)
MA500(h)
S&P500 detrended
and rescaled
crash #1
crash #2
h
maxima #
0.48
0.5
0.52
0.54
0.56
0.58
0.6
1500
1600
1700
1800
1900
2000
2100
2200
2300
MA50(h)
MA500(h)
slope of decay
h level of Brownian Walk
h(BW)
h
0.49
0.5
0.51
0.52
0.53
0.54
0.55
0.56
0.57
0.58
0.59
0.6
3500
3600
3700
3800
3900
4000
4100
MA50(h)
MA500(h)
slope of decay
h level of Brownian Walk
h(BW)
h
Figure 6: TheeectiveHolderexponentsmoothedwithtwowindows(MA50and MA500)is shown.
The two plots to the right show windows on the smoothed eective Holder exponent just before
crash #1and crash#2. (Thecrashesare visible in theleftgure but notin thewindows). Visible
oscillationsofMA50anddecayof MA500characteriseprecursorsofbothcrashes. Theaveragelevel
References
1. Z. R. Struzik, Local Eective Holder Exponent Estimation onthe Wavelet Transform Maxima
Tree, in Fractals: Theory and Applications in Engineering, Eds: M. Dekking,J. LevyVehel, E.
Lutton,C.Tricot,SpringerVerlag,(1999).
2. Z.R.Struzik,DeterminingLocalSingularityStrengthsandTheirSpectrawiththeWavelet
Trans-form,Fractals,8,No2,pp163-179,(2000).
3. M.H.R.Stanley,L.A.N.Amaral,S.V.Buldyrev,S.Havlin,H.Leschhorn,P.Maass,M.A.Salinger,
H.E.Stanley,CanStatisticalPhysicsContributeto theScience ofEconomics?,Fractals 4,3,pp
415-425,(1996).
H.E.Stanley, L.A.N.Amaral, D. Canning,P. Gopikrishnan,Y. Lee, Y. Liu,Econophysics: Can
PhysicstsContributetotheScienceofEconomics?,PhysicaA, 269,156-169,(1999).
4. A.Arneodo,E.BacryandJ.F.Muzy,TheThermodynamicsofFractalsRevisitedwithWavelets.
PhysicaA, 213,232(1995).
J.F.Muzy,E. Bacryand A.Arneodo,TheMultifractal FormalismRevisited withWavelets.Int.
J.of Bifurcation andChaos4,No2,245(1994).
5. R.N.Mantegna,H.E.Stanley,Scalingbehaviourinthedynamicsofofaneconomicindex,Nature
376,46-49(1995).
R.Mantegna,H.E.Stanley,AnIntroductiontoEconophysics(CambridgeUniversityPress,
Cam-bridge,2000).
6. N.Vandewalle,M.Ausloos,CoherentandRandomSequencesin FinancialFluctuations,Physica
A,246,454-459,(1997).
N.Vandewalle,Ph.Boveroux,A.MinguetandM.Ausloos,PhysicaA,(1998).
7. B.B.Mandelbrot, H.Taylor,On thedistribution of stockpricedierences, Operations Research
15,1057-1062(1962).B.B.Mandelbrot,Thevariationofcertainspeculativeprices,J. Business
36,394-419(1963).
8. PrinciplesofDataMiningandKnowledgeDiscovery,J.M.
_
Zytkow,J.Rauch(Eds.)LectureNotes
inArticialIntellingence1704,Springer1999.
D.J.Hand,Data mining: statisticsandmore? TheAmericanStatistician,52,112-118,(1998)
berger,ScalingBehaviourofHeartbeatIntervalsObtainedbyWavelet-basedTime-seriesAnalysis.
Nature,383,323(1996).
P.Ch. Ivanov, M.G. Rosenblum, L.A. Nunes Amaral, Z.R. Struzik, S. Havlin, A.L. Goldberger
andH.E.Stanley,Multifractalityin HumanHeartbeatDynamics.Nature399,(1999).
10. Z.R.Struzik,RevealingLocalVariabilityPropertiesofHumanHeartbeatIntervalswiththeLocal
EectiveHolderExponent,to appearin Fractals,March2001,
seealsoCWIReport,INS-R0015,June2000,available fromwww.cwi.nl/~zbyszek.
11. I.Daubechies,TenLectures onWavelets,(S.I.A.M.,1992).
12. M.Holschneider, Wavelets- AnAnalysisTool,(OxfordSciencePublications,1995).
13. S.G.MallatandW.L.Hwang,SingularityDetectionandProcessingwithWavelets.IEEETrans.
onInformationTheory38,617(1992).
S.G.Mallat and S. ZhongComplete SignalRepresentation withMultiscale Edges.IEEE Trans.
PAMI14,710(1992).
14. S.Jaard,MultifractalFormalismforFunctions: I.ResultsValidforallFunctions,II.Self-Similar
Functions,SIAMJ. Math.Anal., 28(4): 944-998,(1997).
15. R.Carmona,W.H.Hwang,B.Torresani,CharacterisationofSignalsbytheRidgesoftheirWavelet
Transform,IEEE Trans.SignalProcessing45, vol10,480-492,(1997)
16. Y.Liu,P.Gopikrishnan,P.Cizeau,M.Meyer,C.-K.Peng,H.E.Stanley,TheStatisticalProperties
oftheVolatilityofPriceFluctuations, arXiv:cond-mat/9903369v2,25Mar1999.
17. A. Johansen, D. Sornette, Stock Market Crashes are Outliers, arXiv:cond-mat/9712005v3, 16
Dec1997.
18. V.S. L'vov, A. Pomyalov, I. Procaccia, Outliers, Extreme Events and Multiscaling,
arXiv:nlin.CD/0009049,27Sept2000.
19. A.Arneodo, J.-F.Muzy,D. Sornette, `Direct'casualcascadein thestockmarket,Eur.Phys. J.