Chapter 17. Hybrid Intelligent Decision Support Systems and Applications for Risk Analysis

N.Kasabov,L.Erzegovesi,M.Fedrizzi,A.Beber,andD.Deng

Dept.ofInformationScience,Univ.ofOtago,Dunedin,NewZealand.E-mail:nkasabov,ddeng@infoscience.otago.ac.nz

Dept.ofInformaticsandFacultyofEconomics,Univ.ofTrento,Italy

Abstract.Decisionmakinginacomplex,dynamicallychangingenvironmentisadifficulttaskthatrequiresnewtechniquesofcomputationalintelligenceforbuild-ingadaptive,hybridintelligentdecisionsupportsystems(HIDSS).Here,anewap-proachisproposedbasedonevolvingagentsinadynamicenvironment.Neuralnetworkandrule-basedagentsareevolvedfromincomingdataandexpertknowl-edgeifadecisionmakingprocessrequiresthis.Theagentsareevolvedfrommeth-odsincludedinarepositoryforintelligentconnectionistbasedinformationsystemsRICBIS(http://divcom.otago.ac.nz/infosci/kel/CBIIS.html)withtheuseoffinancialmarketdatacollectedinanon-linemode,andwiththeuseofmacroeconomicdatapublishedmonthlyintheEuropeanCentralBankBulletin.RICBISincludesdifferenttypesofneuralnetworks,includingMLP,SOM,fuzzyneuralnetworks(FuNN),evolvingfuzzyneuralnetworks(EFuNN),evolvingSOM,rule-basedsys-tems,datapre-processingtechniques,standardstatisticalandfinancialtechniques.AcasestudyprojectonriskanalysisoftheEuropeanMonetaryUnion(EMU)isconsideredandaframeworkofasystemEMU-HIDSSispresented,whichdealswithdifferentlevelsofinformationandusers,e.g.thewholeworld,Europe,clustersofnations,asinglenation,companies/banks.Itcombinesmodulesforfinaldecisionmaking,globalandnationaleconomicdevelopment,exchangeratetrendpredic-tion,stockindextrendprediction,etc.Someexperimentalresultsonrealdataarepresented.

Keywords.Intelligentdecisionsupportsystems,riskanalysis,connectionist-basedinforma-tionsystems.

1Introduction

Complexdecisionmakingproblemsveryoftenrequireconsideringenormousamountofinformationdistributedamongmanyvariables.Thedecisionsupportsys-tems(DSS)builttosolvetheseproblemsshouldhaveadvancedfeaturessuchas[19]:

Goodexplanationfacilities,preferablypresentingthedecisionrulesused;Dealingwithvague,fuzzyinformation,aswellaswithcrispinformation;

Dealingwithcontradictoryknowledge,e.g.,twoexpertspredictdifferenttrendsinthestockmarket;

Dealingwithlargedatabaseswithalotofredundantinformation,orcopingwithlackofdata.

Havinghierarchicalorganisation,asdecisionmakingisusuallynotasingle-levelprocess,butinvolvesdifferentlevelsofprocessing,comparingdifferentpossiblesolutions,usingalternatives,sometimesinarecurrentway.

Techniquesofcomputationalintelligence,suchasartificialneuralnetworks,fuzzylogicsystems,geneticalgorithms,advancedstatisticalmethods,hybridsystems,alongwithtraditionalstatisticalandfinancialanalysismethods,havebeenwidelyappliedondifferentproblemsfromfinanceandeconomics.Heresomeofthemarereferencesandexplained.

1.1SpecialisedandAdvancedStatisticalandEconometric

Models

In[45]amodelforriskanalysisoncrashesofcurrenciesisintroduced.Itconsid-ersacrashsituationwhenacurrencyhasbeendevaluatedmorethan10%withinamonthwhencomparedtotheUSdollar.Fourparametersareusedtoevaluatethelikelihoodofacrash(avalueof0or1)inthefollowingmonth:measureofcurrencyovervaluation;expecteddomesticoutputgrowth;ratiooffederalreservesversusex-ternaldebt;measuresofinternationalfinancialcontagion.Thelatterismeasuredinriskappetite(achangeininvestorspreferencesbeforecrashesoccur),andclus-ters(whethercrashesofcurrenciesfromthesameblockofcountrieshaverecentlyoccurred).Othermodelsarepresentedin[6].

Themethodsinthisgrouprequireareliableknowledgeontheunderlyingrulesofthefinancialandeconomicsystemsandcannotbeeasilyadjustedtoaneweco-nomicorfinancialsituation.Theyareapplicablewhencrisesoccuraccordingtorecurrentandknownpatterns.

1.2SymbolicandFuzzyRule-BasedSystems

Rule-basedsystemsandfuzzyrule-basedsystemsinparticularhavebeenusedinfi-nancialandeconomicdecisionmaking(seepapersin[12][13]).Themainadvantageofrule-basedsystemsisthattheirfunctioningisbasedonhumanexpertrules.Themaindisadvantageisthatthesesystemsarenotflexibleenoughtoreacttochangesinthereality.Changingtherulesusuallyisdifficult,andnobodycanarticulatetheperfectsetofrulesthatdonothavetochangeinafuturetime.

1.3ArtificialNeuralNetworks(ANN)

ExtendedstudiesofusingANNforfinancialDSShavebeendoneinseveralbooks(seeforexample:[12][13][49][54]).Generally,thesocallednon-parametricmod-elsprovedtooutperformthestatisticalmethods,especiallywhentheunderlyingruleswerenotknown,ortheychangeovertime.DifferenttypesofANNwereused,mainlymulti-layerperceptrons(MLP),radialbasisfunctions(RBF),andself-organisingfeaturemaps(SOM).

AMLPmodelandaRBFmodelonoptionpricing[16]provedtooutperformseveralothermodels,thatincludeadirectapplicationofBlack-Scholesformula,ordinaryleastsquares,andprojectionpursuit.ThetestdatawasthedailyS&P500futuresandoptionprices.Twoinputvariables–theratiostockprice/strikepriceandtimetomaturity,andoneoutputvariable-optionprice,wereused.Thismodelwasfurtherextendedintoamulti-modularANNmodelwiththeuseofhints[10].

In[5]asubstantialstudyofapplyingSOMtofinancialmarketsispresented.In[3]emergingandnewmarketsaremappedintoaSOMthatshowsgroups(clusters)ofsimilareconomies.

ThesofardevelopedandusedANNmodelsprovedtobeefficient,buttheydonotalloweasilyfordynamicon-linetraining,forchangingtheparametersinanon-linemode,forcombiningdataandknowledge(rules)intoonesystem.

1.4GeneticAlgorithms

Geneticalgorithms(GA)areheuristicmodelsthatarebasedongeneratingpossiblesolutionstoaproblemandevaluatingtheir“goodness”basedonagoodness(fitness)function(e.g.goalfunction)thathastobespecifiedinadvance.GAuseterminologyfromthenaturalselectionandevolutionofspecies.ThereareseveralstudiesofusingGAforeconomicandfinancialdecisionmaking(seealsochaptersinthisvolume).SoftwarebasedonGAsimulation,thatworksdirectlyondatainanExcelformathasbeenproduced[8].

ThemainadvantageofGAsisthattheydonotrequiremuchknowledgeontheunderlyingrules,formulas,etc.,butagoodnessfunctiontoevaluatehowgoodsolutionsare.ThemaindisadvantagesofGAsare:(i)theyarecomputationallyslow;(ii)theydonotnecessarilyprovidewiththebestsolutionastheyareheuristicallybased;(iii)theydonotworkinon-lineandrealtimemodes.

1.5ModelsBasedonDynamicSystemAnalysis

In[50]thestockmarketismodelledasacomplexdynamicsystemthatcanbeinoneoffourstatesprojectedinatwo-dimensionalspaceofgroupthinking-fundamentalbias:randomwalk,chaoticmarket,coherentbullmarket,coherentbearmarket.In[38][39]astockindexpredictionproblemisconsideredasequivalenttothepredic-tionofachaoticprocesswithdifferentcharacteristicsatdifferenttimescalesandahighfrequencyindexprediction(e.g.dailyprediction)beingattemptedafterthelowfrequencyone(e.g.,monthly,orannualprediction)isperformed.

1.6HybridSystems

Hybridsystemscombineseveraloftheabovemethodsintoonesystem[19][20].Theyachievethiscombinationeitherina“loose”way,e.g.differentmodulesinthesamesystemusedifferentmethods(seeexamplesofsuchsystemsin[12][13]),orina“tight”way-methodsaremixedatalowlevel,e.g.fuzzyneuralnetworks([18][29][42]).Thesemethodsarethemostpromisingamongthemethodsdiscussedabove,asthehybridsystemsintegratetheadvantagesofallthemethodscombined,e.g.dealingwithbothdataandexpertrules,usingbothstatisticalformulasandheuristicsorhints.

1.7WhatisMissingintheMethodsfromAbove?

Thesixgroupsofapproachespresentedabovehavebeenpartiallysuccessful,withmajorproblemsaslistedbelow:

Theydonotconsiderthecomplexityoftheprobleminawholewithmanyhier-archicallevelsfordecisionmakingthatincludeapplyinglow-levelprocessingandhigherlevelexpertknowledge;or

Theydonotconsideruncertaintiesatdifferentlevelsofinformationprocessingandcombiningthemorpropagatingtheminataskdependentwaytothefinaldecisionmakingblock;or

Theydonotapplysufficientvarietyoftechniquesandchoosethemostappro-priateforeachsub-task;or

Theydonotofferadjustmentofvariablesets,optimisationcriteria,rules,eveniftherealsituationchangesovertime.

2FromDSStoHIDSS

Hybridsystemscombinedifferenttechniquesofcomputationalintelligencewithtraditionalstatisticalmethods.HybridsystemsareespeciallysuitableforbuildingDSS.

Stockmarketindexpredictionisagoodexampleofacomplexproblemthatrequiresahybridsystem,asitisshownonthecasestudyoftheNZSE40stockindexin[19][38][39].SeveralmodulesareincludedintheDSSsystempresentedthereasthereareseveraltaskswithintheglobalone:datapre-processing(e.g.normalisation,movingaveragescalculation,etc);predictingthenextvaluefortheindex;predictinglonger-termvaluesfortheNZES40,finaldecisionmakingthattakesintoaccountrulesonthepoliticalandeconomicsituation;extractingtradingrulesfromthesystem-seeFig.1.Aneuralnetworkisusedtopredictthenextvalueoftheindexbasedonthecurrentandtheprevious-dayvalues.Thepredictedvaluefromtheneuralnetworkmoduleiscombinedwithexpertrulesonthecurrentpoliticalandeconomicsituationinafuzzyinferencemodule.Thesetwovariablesarefuzzybynature.Thefinaldecisionisproducedasafuzzyone,andasacrispone

-afteradefuzzificationprocess.AnothermoduleintheDSSfromFig.1isdevotedtoextractingfuzzytradingrules,whichareusedtoexplainthecurrentbehaviourofthemarket.

Anenvironment,calledFuzzyCOPE/1,thatcanbeusedtocreatehybridsys-tems,isdescribedin[19]andavailablefrominternetURLhttp://kel.otago.ac.nz/.Itconsistsofthefollowingmodulesthathavecompatibleinterfacesandcanbecon-nectedinaDSSasadecisionmakingsequencethatrepresentsthelogicoftherealdecisionmakingprocess:dataprocessingmodules(normalisation,fuzzification,fil-tering,etc.);multi-layerperceptron(MLP);self-organisingmap(SOM);fuzzyneu-ralnetwork(FuNN)asintroducedbyKasabovin[18][19][28];fuzzylogicinferenceengine(FLIE)basedonsimplefuzzyrulesofZadeh-Mamdanitype(Zadeh,1965);productionrule-basedsystemCLIPSandFuzzyCLIPSinparticular.TheFuzzy-COPE/1environmenthasbeenextendedtoFuzzyCOPE/3withtheinclusionofnewMLPandSOMlearningmodes,andnewmodesforlearning,ruleextractionandruleinsertioninFuNNs.ExamplesofhybridsystemsbuiltwiththeuseofFuzzy-COPEaregivenin[19].ThetwoenvironmentsFuzzyCOPE/1andFuzzyCOPE/3areavailablefromthefollowingWWWsiteandcanbeusedforbuildinghybridDSS:http://divcom.otago.ac.nz/infosci/kel/CBIIS.html(SoftwareFuzzyCOPE).

Fig.1.AnexampleofahybridDSSforstocktrading(from[19])

Thehybridsystemenvironmentsdevelopedsofar,andthehybridsystemsbuiltwiththem,havebeenveryusefultechniques,butthecomplexityandthedynamicsofthereal-worldproblems,especiallyinfinanceandeconomicsatpresent,requireevenmoreadvancedandsophisticatedmethodsandtoolsforbuildinghybridintel-ligentdecisionsupportsystems(HIDSS).Suchsystemsshouldbeabletochangeastheyoperate,alwaysupdatingtheirknowledge,andtorefinethemodelthroughinteractionwiththeenvironment.Somemajorrequirementstothepresentdayintel-ligentsystems(IS),andtotheHIDSSinparticulararegivenin[21]-[25].

Aframeworkforbuildingadaptiveintelligentsystems,calledECOS(evolvingconnectionistsystems)hasbeenrecentlyintroducedin[21]-[25],alongwithitsar-

chitecture,learningandevolvingalgorithms,ruleextractionandruleinsertionalgo-rithms,ofanevolvingfuzzyneuralnetworkEFuNN[23]-[25].EFuNNscanlearninanincremental,adaptivewaythroughone-passpropagationofanynewdataex-amples.EFuNNsaremuchfasterthanFuNNsandMLPsandcanlearndatainanon-linemode.EFuNNsdonothaveafixedstructure,onthecontrary–theystartevolving/learningwithnorule(hidden)nodesandtheygrowasdataispresentedtothem.Pruningofnodesandnodereductionisachievedwiththeuseoffuzzyprun-ingrules,e.g.:

IFanodeisnotmuchusedinadefinedperiod,ANDitisold,THENprobabilitytoprunethenodeishigh.

EFuNNshavethefollowingadvantageswhencomparedwithtraditionalMLPorSOMnetworks([24]:

1.theycanlearninanon-linemodeanynewdataastheyaremadeavailableovertime;

2.theycanworkinacomplexenvironmentwithchangingdynamics,e.g.astockindexsystemcanbeinarandomwalkstate,thenitmovestoachaoticstate,andthen-toquasiperiodicstate,andanEFuNNthatpredictsfuturestockvalues,learnsallthetimethenewbehaviourwithoutanyhumaninterventionforparameteradjustment.

3.theycanbeusedtomixexpertrulesanddataastherearealgorithmsforruleinsertionandruleextraction;

4.theycanclusterdatainanon-linemodewithoutpre-definingthenumberofclusters,orthedimensionalityandthesizeoftheproblemspace;

5.theycanbeusedforbothsupervisedandunsupervisedlearningmodes;assuper-visedsystemstheycanbeusedtopredictfuturevaluesoftheoutputvariables.ExamplesofusingFuNNsandEFuNNsforadaptive,intelligentdecisionsup-portsystemsforstockpredictionandloanapprovalaregivenin[26].Otherex-amplesinclude:imagerecognition[34];speechandlanguagerecognition[24][33];mobilerobotcontrol[24].

SimulatorsofEFuNNareavailablefromhttp://divcom.otago.ac.nz/infosci/kel/CBIIS.html(Software).

AnotherECOSalgorithm,calledEvolvingSelf-organizingMap(ESOM)[4],isproposedasavariationoftheSOMalgorithmbasedontheECOSprinciples.ESOMusesalearningrulesimilartoSOM,butitsnetworkstructureisevolveddynamicallyfrominputdata.SimulationshaveshownthatESOMlearnsfasterthanSOMwithasmallerquantisationerrorforfeaturevectors.

ECOS-basedmodulessuchasEFuNNsandESOMarepartoftheNewZealandRepositoryofIntelligentConnectionist-BasedSystems(RICBIS),whichalsointe-gratesmodulesfromtheFuzzyCOPEenvironments,aJavaversionofrule-basedsystemCLIPS(JESS),andaplatformindependentinterfacerunningasaJavaap-plet,whichallowsfordynamiccreationofnewmodulesduringtheoperationofanECOS,oranHIDSSinparticular.RICBISisavailablefromthefollowingURL:http://divcom.otago.ac.nz/infosci/kel/CBIIS.html(Software).

AnewexpertsystemarchitecturecalledAdaptiveIntelligentExpertSystems(AIES),basedondynamicgenerationofinterconnectedmodules(agents)fromtheRICBIS,isexplainedin[34].Itisinsharpcontrasttotheconventionalexpertsys-temsandDSSthatusuallyhaveafixedstructureofmodulesandafixedrulebase.AlthoughtraditionalexpertsystemsandDSShavebeensuccessfulinsomespecificandrestrictedareas,therewasno,orlittleflexibilityleftfortheexpertsystemtoadapttochangesrequiredbytheuser,orbythedynamicallychangingenvironmentinwhichtheexpertsystemandtheDSSrespectivelyoperated.

AIES,orHIDSSinparticular,consistofaseriesofmoduleswhichareagent-basedandaregenerated“onthefly”astheyareneeded.Fig.2showsageneralarchitectureofanAIES[34].Theuserspecifiestheinitialproblemparametersandtaskstobesolved.TheAIESthencreatesModulesthatmayinitiallyhavenorulesormaybesetupwithrulesfromtheExpertKnowledgeBase.TheModulescombinetheruleswiththedatafromtheEnvironment.TheModulesarecontinuouslytrainedwithdatafromtheEnvironment.RulesmaybeextractedfromtrainedFuNNsorEFuNNsforlateranalysis,orforthecreationofnewModules.TheModulesdy-namicallyadapttheirrulesetsastheenvironmentchangessincethenumberofrulesisdependentonthedatapresentedtotheModules.Modules(agents)aredynami-callycreated,updatedandconnectedinanon-linemode.TheycanberemovediftheyarenomoreneededatalaterstageoftheoperationoftheAIES.

User task specificationIntelligent Design InterfaceResultsSolutionKnowledgeBaseAgent 1Agent 2ModuleGenerationRepository of ModulesAdaptive Learning Data TransformationNeural NetworksFuzzy LogicGenetic AlgorithmEFuNNsData / EnvironmentDatabasesWeb pagesTime series Fig.2.Ablockdiagramofanagent-based,adaptiveintelligentdecisionsupportsystem–HIDSSthatusesthearchitectureofanAIES.

AverycomplexproblemofriskanalysisoftheEuropeanMonetaryUnion(EMU)system,establishedin1999tounifythecurrencyandtheeconomicdevel-opmentofelevenEuropeancountries,istheproblemdiscussedandhandledhere.TherestofthematerialherepresentsfirsttheproblemandthenaframeworkofaHIDSSforitssolution.Itthendevelopssomespecificmodulesanddiscussessomepreliminaryexperimentalandimplementationissues.

3TheProblemofRiskAnalysisintheEuropeanMonetaryUnion

Sinceitsconception,therehavebeenalotofmaterialspublishedondifferentissuesconcerningEMU.Globalpolicyrequirementsexist,suchasforeachparticipatingcountrytohaveadeficitlessthan3%ofitsGDPandexternaldebtlessthan60%oftheGDP.IntheEMUframework,theEuropeanCentralBank(ECB)isinchargeofthemonetarypolicy,withpriorityandresponsibilityforinflationcontrol.TheECBpublishesamonthlybulletincontainingarichsetofreal,monetaryandfinancialdataregardingtheEMUeconomies,othercountriesthatwouldbemembersoftheEMU,USAandJapan,inordertofollowtheevolutionoftheEMUinaworld-widecontext.Importantfinancialandeconomicparametersarerecordedandanalysedmonthly,quarterly,andannually,e.g.:Reservesandassets(gold,foreignexchange,other);Liabilities;StockmarketindexesoftheEMU,eachcountryseparatelyandthemajorworldindexes(DowJonesEUROSTOXX,S&P500–USA,Nikkei225–Japan);Interestrates;ExchangeratesEuro/US$,Euro/JY;Governmentbondyields(2,3,5,7and10years);Indexofconsumerprices;Industryandcommodityprices;GDP;Employment/unemployment;Saving,deficit/surplusratio(asa%ofGDP);Grossnominalconsolidatedebt(asa%ofGDP);Balanceofpayments(goods,services,income,capitalaccount).

OfaparticularinterestistheanalysisoftheEMUasadynamicclusterofeconomiesintermsofvolatility,variations,change,tendencies,andprediction(e.g.,monthly,quarterly,yearly).Therearesmallersub-clustersthatevolvewiththeeco-nomicdevelopmentofdifferentgroupsofEuropeancountriesandworldeconomiesthatshouldbealsomodelledandpredictedinrelationtotheEMUcluster.Alltheseclustersmovequicklyinadynamicallychangingproblemspace,thusmakingtheproblemoftheirpredictionandriskanalysisextremelydifficult.

Theproblemthispaperisdealingwithasacasestudy,canbedescribedasriskanalysisoftheEMUsystem.Here,moredetailsaregiven.

WiththeEMUineffect,countriessharingtheEuroasacommoncurrencyshouldovercometheriskofcurrencycriseswithintheEuroarea.However,theEuropeanmonetaryunificationhasnotruledoutpossibleepisodesoffinancialin-stabilityinEurope.TheMaastrichtTreatyimposesrigidconstraintsonpublicbud-getdeficits.Theseconstraintsareaimedatpreventinganexcessivedebtburdenonnationalgovernments,whichcouldleadtoaweakerEuro.Ontheotherside,EMUgovernmentscouldputpressureinordertoeasesuchconstraintsinthepresenceofexternalshocks,suchasthecrisisintheBalkans.Moreover,Europeanfinancialmarketsarenotimmunefromshocksoriginatingintheworldeconomy.

Inanextremescenario,riskofunilateralwithdrawalbyweakermembercoun-tries(otherwisecalledbreakawayrisk)cannotbeexcludedapriori,sincethepo-liticalcostsassociatedwithrespectofrigidfiscalandmonetaryconstraintsinananchoredregimecouldmakewithdrawalimperative.Anywithdrawalwouldbeadisruptiveevent,anticipatedand/orfollowedbyinstabilityandcrashesincreditandassetmarkets.CredibilityofEMUmembershipwillbeassessedandpricedbyfinan-

cialmarkets.IntheEMU,expectationsofbreakaway,unlikeexpectedrealignmentsorwithdrawalsintheExchangeRateMechanismoperatingfrom1979to1998undertheEuropeanMonetarySystems,willnolongertranslateintowiderinterestratedif-ferentialamongcurrencies,butintovariationofcreditspreadsappliedtosovereigndebtfromdifferentcountries.Holdersoffinancialassetswillbearanewsortofmacrorisk,whichwillbedifferentfromplaincurrencyriskandmoredifficulttoidentifyandmeasure.Studiesoncurrencycrises(e.g.[6])mustbereinterpretedandextendedtothenewcontext.Earlywarningsystemsusedbycentralbanksandspec-ulators,fedwithsignalsofrealandfinancialdis-equilibrium,mustberedesigned.Thegoalofthisprojectistodevelopacomputationalmodelforanalyzingandanticipatingsignalsofabruptchangesofvolatilityinfinancialmarkets.Thesys-temwillbeaimedatassessingthepossibilityofspeculativeattacksagainstspecificEMUmembercountries,prospectiveEMUmembersortheEMUareaasawhole.Potentialusersofthesystemincludemonetaryauthorities,assetmanagers,tradersonmoney,debt,currencyandstockmarketsandcorporatefinancialmanagers.Theconceptualmodelunderlyingthecomputationalmodelwillbederivedfromarepresentationoffinancialmarketsascomplexdynamicsystems,whosestochasticbehaviorisinfluencedbyexogenousshocksandendogenousuncertainty,thelattercausedbyinteractionamongmarketparticipants(degreeofconsensusandtendencytocrowdbehavior).InspirationforthisapproachcamefromapaperbyTonisVaga([50]).Thesystemwillbefedwithinformationfromdifferentsources,namely:-macroeconomicandmacro-financialindicators,suchastherealexchangerateoftheEuroagainstUSDollarandJapaneseYen,inflationdifferentials,Governmentdeficit,aggregateliquidityandsolvencymeasures(debt/assetratio,reserve/debtra-tio);

-riskspreadondebtsecuritiesissuedbysovereignandprivateborrowersinEMUcountries;

-riskappetiteofinvestorsinsecurities,measuredonthebasisofcorrelationbetweenreturnsin“risky”and“safe”markets(riskappetiteishighwhenriskiermarketsarerallyingversus“safe”marketsandthecorrelationishighlypositive);

-returnsandhistoricalvolatilityinfinancialmarkets(stock,currency,bond,money,derivatives);

-signalsoftrend,reversalandchangeofregimefromtechnicalanalysisoffinancialprices(movingaverages,resistanceandsupportlevels,relativestrengthindicators,etc.);thesesignalswillserveasproxyvariablesforendogenousuncertainty;

-impliedvolatilityinoptionmarketsandexpecteddistributionsextractedfromthem;

-recentepisodesofinstabilityinothercurrencyareasthatcanexertacontagioneffectontheEMUarea.

Thelogicofthesystemwillbedesignedasanextensionofexistingeventriskmodelsappliedtoforeignexchangemarkets([45]).Thesystemwillbetestedonasetofrecentepisodesofmarketcrash,andthenextendedtakingcareofuniquefea-turesofthenewEMUmonetaryregime.Thesystemwillproducearichinformativeoutput,consistingofdescriptivereportsandwarningsignals.

Firstly,thesystemwillprovideanintelligentinterfacetoinformationcurrentlyanalyzedbyeconomistsandtraders.Userswillbeabletonavigatethrougharichsetofeconomicandfinancialdatapresentedintablesandgraphs.ThepresentationwillfocusonphenomenapertainingtotheeconomicperformanceintheEMUarea,em-phasizingdivergencesamongcountriesandsustainabilityofMaastrichtconstraintsbothatthenationalandtheEMUlevel.

Secondly,thesystemwillprovidesignalsandindicatorsreflectingthelikelihoodofacrashinEMUfinancialmarkets.Anextensivesetofsymptomsoffinancialfragilitywillbemonitoredincredit,bondandstockmarkets.Neweventswillbecheckedagainsttypicalpatternsofevolutionoffinancialcrises.

Thesystemwillprovideavaluablesupporttoanalystsanddecisionmakersintwoways:(1)selectingrelevantinformationtobesubsequentlyanalyzedbyhumanexpertsand(2)extractingandsynthesizingsignalsfromavastarrayofinformationsources.

4CollectingRelevantDataforRiskAnalysisontheEMU.TheTrentoFinancialDataDictionary

Thedatacollectionisanimportantpartoftheresearchproject,becausetheresultsandthereliabilityofthehybridsystemdependcruciallyonthequalityoftherawdata.Infactahugeamountofdataisnecessarytomonitortheso-calledbreakawayriskintheEMU,inordertodetectthesolidityoftheEMUsystem,thedegreeofasymmetryandthepotentialexternalshocks.

Hugeamountofdataisbeingcollectedthatincludesdifferentlevelofinforma-tion:macroeconomicinformation;financialdata.MacroeconomicdataisavailablefrommanysourcesincludingEUROSTAT,theEuropeanCentralBankMonthlyBulletin,OECSandIMF.Avastamountofmacroeconomicdataismadeavail-ablethroughDatastream.Financialdataisavailablefromdifferenton-linesources,suchasDataStreamandReuter.Differenttime-scalesofdataarepresentinthedatarepository(e.g.,yearly,quarterly,monthly,daily,irregulartimeintervals,etc)

Itisnoteasytodealwiththisrelevantamountofdatawithoutpreparingatleastageneralinitialstructure.Forthisreasonwehavebuilttwomainprospects:the“sourcesdictionary”andthe“datadictionary”.

Thefirstprospecthasathreedimensionalstructure.Inthex-axiswehaveplacedthecountriesrelevantfortheEMUriskanalysis;wehavechosentheelevenpartic-ipatingcountries,thepotentialentrants,thebigworldplayers,andsomeemergingcountriestoproxyforanycontagioneffect.Inthey-axiswehavesetallthetypesofdatawhichcouldbeusefulfortheresearchproject.Thethirddimensioniscon-cernedwiththesourceofthegatheredinformation;everysinglevariablerelatedtoeveryinvolvedcountryhasbeencollectedbyaspecificsource.Thisaspectisimpor-tanttoevaluatemainlythereliabilityofthedata,butalsothedegreeofhomogeneityinthedataset.Infact,choosingthedatasourcewhichcoversthemostpartoftheinvolvedcountriesallowstoachievehighdegreeofsimilarityinthemodellingpro-cedureandintheupdatingtimingandmethod.Consideringthespecificpurposeof

theanalysis,werelyuponofficialstatisticsfromEurostatformostoftherequireddata.

Thesecondprospect,named“datadictionary”,isformedasatablewithcolumnsreportingthespecificfeaturesofeachdataseries;themostimportantonesarethefirstavailabledate,thefrequencyofcollection,andthemnemoniccode,whichisanalphanumericexpressionusefultoautomatetheinformationdownloadingprocess.Wehavedistinguishedtwobroadcategoriesofdatathatarebeingcollected.Thefirstlevelofinformationisconcernedwithmacroeconomicvariables;thesecondlevelismorespecificandregardsfinancialmarketdata.Athirdlevelofinforma-tion,connectedwithqualitativeknowledge,forexamplethepoliticalsituationorparticularnews,isnotformallyconsideredinthisfirstmodelimplementation.Thetypicalfeatureofthefirstgroupofvariables–macroeconomicdata–isthelowfre-quencyofcollection,whichismonthly,quarterlyorevenyearly,dependingonthespecificsector;anotherimportantfeatureisthepotentialvarietyofsourcesforthesamekindofdata.Athirdpropertyisthepotentialdifferentcalculusandupdatingprocedureforthesamesortofindices.Themostrelevantmacroeconomicdataaregenerallyavailableeitherashistoricaltimeseries,orasaconsensusforecast.

Thefinancialmarketvariablesarecollectedmorefrequently,evenonadailybasis,andareusuallyreleasedofficiallybytheexchanges.Inourresearchprojectwehavetriedtoconsidernotonlythepastevolutionofthevariousfinancialseries,butalsothemarketexpectationsimpliedinoptionprices;thereisagrowingliteraturedealingwiththisaspectwithtwomainapproaches.Thesimplerapproachisrelatedtothecalculationofimpliedvolatilityatdifferentmonetarydegreesasaforecastforexpectedvolatility.Inthesecondapproachtheunderlyingriskneutraldensityfunctionisextractedfromoptionmarketprices;itisveryusefulsubsequentlytomonitortheevolutionofthevariousmomentoftheprobabilitydistribution.

5TheEMU-HIDSS:ArchitectureandFunctionality

5.1AGeneralFrameworkoftheEMU-HIDSS

HereapreliminarydesignoftheEMU-HIDSSispresented(seeFig.3).Itisamulti-level,multi-modularstructurewheremanyneuralnetworkmodules(denotedasNNM),rule-basedmodulesandothermodulesareconnectedwithinter-,andintra-connections.EMU-HIDSSdoesnothaveaclearmulti-layerstructure,butratheramodular,“open”structure.Itisanevolvinghybridconnectionist-basedsystem([22]).

ThemainpartsoftheEMU-HIDSSaredescribedbelow.

(1)Featureselectionpart.Itperformsfilteringoftheinputinformation,featureex-tractionandforminginputvectors.Typicalfeaturesextractedfromtheinputdataeitherinanon-linemodeorfromthealreadystoreddatainfilesare:

Basicstatisticalparameters;

Probabilitydistributionandclusterinformation;

Movingaverages;

Wavelettransformations;

PowerspectrumandFFTfrequencycharacteristics;Mainfrequencies;Lyapunovcoefficients;Fractaldimensions;

Firstandsecondderivatives;Skewnessmeasures.

Theabovetransformationsareperformedinthepre-processing(featureextraction)modulesofthesystemappliedtocertaininformationinputstreams.

(2)Learningandmemorypart,whereinformation(patterns)arestored.Itisamulti-modular,evolvingstructureofneuralnetworkmodules(NNM).ThesemodulescanbebuiltwiththeuseofMLP,SOM,ESOM,FuNNs,EFuNNs,etc.Thereareseverallevelsofprocessinginthesemodulesintermsoftiming:

Dailyupdatedmodules,thesearemodulesthatdealwithdailyfinancialpre-dictionanddailyinputdata,e.g.MIB30prediction,Euro/US$exchangerateprediction,etc.

WeeklyupdatedmodulesMonthlyupdatedmodules

Yearlyupdatedmodules,e.g.longtrendprediction,andintermsoftheproducedresultsthatarepassedtothenextlevelhigherdecisionmodules:OnedayaheadpredictionresultsMonthlypredictionresultsYearlypredictionresults

Longertermpredictedresults

ThispartofthesystemwillincludeseveralmodulestodealwithdifferentlevelsandscalesofpredictionforeachoftheEuropeancountries,thebigeconomiesandtheemergingeconomies.Differentmoduleswilldealwith:

Predictingvalues(differencesinvalues)

Predictingshorttermtrends,e.g.oneweektrendifastockvaluewillbegoingupextremelyhigh,ormoderatelydown.

Predictinglongtermtrends,e.g.one-yeartrendifastockvaluewillbegoingupextremelyhigh,orcriticallydown.

(3)Higher-leveldecisionpartthatconsistsofseveralmodules,eachtakingdeci-siononaparticularproblem.ThemodulesreceiveinputfromtheNNMs,inputsformothervariablesinthedata,qualitativeinputfromusers,andmakedecisionsonpossiblecriticalsituationsthatmightoccurintheEMU.ThesemodulescansendafeedbacktotheNNMandthefeatureextractionpartofthesystemintermsofrequiringmoreinformation,differentscenariostobeexplored,differentfeaturesextracted,etc.Themodulesherearemainlyrulebasedwiththeuseofproduction

systems,flatfuzzyrules,FuNNsandEFuNNsthatcanrepresentbothfuzzyrulesanddata.Thereareseveralgroupsofinthispartthatinteractbetweeneachother,forexample:

(a)Agroupofmodulesthatdealwithaglobalriskevaluationproblems,e.g.:

ModuleevaluatingthedegreeofstabilityintheEMU;

Moduleevaluatingthedegreeofsymmetry/asymmetrybetweentheeconomieswithintheEMU;

ModuleevaluatingthepoliticalsustainabilityintheEMU;

ModuleevaluatingthedegreeofsuitabilityofanewcountryjoiningtheEMU;ModuleevaluatingthedegreeofinstabilityintheEMUbasedonexternalfac-tors(USA;Asia,Japan,India,Russia,wars,etc.)others

(b)Agroupofmodulesthatdealwithimportanteconomicfactorsandtheirresultscanbeusedeitherseparately,orbythemodulesoftype(1)above:

GDP

RateofunemploymentInternaldebtExternaldebt

ShorttermglobaleconomictrendsLongtermglobaleconomictrends

Solvencyratioofhouseholds,business,banksandgovernmentIndicationofreallocationofinvestmentsbyglobalassetmanagers

SharpmovementsofacertaincommodityinashortorinalongtermpatternSharpfallsinshorttermorlongtermtrends

Evaluatingandindicationofconsecutivephaseshappeningoveraperiodoftime,forexampleaneconomyhasbeeninthreeconsecutivephasesthatsignalacriticalsituationforthiseconomyandwillbeinfluentialfortheEMU

(4)Actionmodules,thattaketheoutputfromthehigher-leveldecisionmodulesandproduceoutputresultsorsendoutput(control)informationinanon-lineorinanoff-linemodetoinstitutionsthatshouldbealertedonacriticalsituation.

(5)Self-analysis,andruleextractionmodules.Thispartextractscompressedab-stractinformationfromtheNNMsandfromthedecisionmodulesindifferentformsofrules,abstractassociations,etc.HereFuNNs’andEFuNNs’ruleextractioncapa-bilitieswillbeutilised.

InitiallytheEMU-HIDSSwillhaveapre-definedstructureofmodulesandveryfewconnectionsbetweenthemdefinedthroughpriorknowledge.Gradually,thesys-temwillbecomemoreandmore“wiredthroughself-organisation,andthroughcre-ationofnewNNMandnewconnections.

Eachofthemodulesinthesystemarebuilt,orautomaticallygeneratedfromtheagentmodulesavailablefromRICBIS,e.g.:dataprocessingmodules(e.g.normal-isation,movingaverages,FFT,filtering,wavelettransformation,fractalanalysis,chaosanalysis,etc);productionrulesinJESS;fuzzyinferencerules,MLP,SOM,ESOM,FuNNs,EFuNNs,Hidden-MarkovModels,etc.

Learning ModulesFeatureSelectionModulesDecisionPartActionPartEnvironment(Critique)AdaptationNNMInputData• • • NNMAction ModulesHigherLevelDecis.ModulesResultsNewInputsNNMSelf analysis, Rule extractionModulesFig.3.AblockdiagramofthegeneralframeworkoftheEMU-HIDSSasanevolvingconnectionist-basedsystem(adaptedfrom[22]).

5.2EMU-HIDSS/1

Here,thefirstversionoftheEMU-HIDSSispresentedthatincludesasmallnumberofmodulesandgroupsconnectedbetweeneachotherasdescribedbelow.Group1:Modulesforhigher-levelriskanalysis

Herearemodulesthatmakedecisiononthediscrepancy/riskofasinglecountrytodevelopinadirectionawayfromtheexpecteddevelopmentoftheEMU.Fuzzyproductionrulesareusedfortheimplementationofthesemodules,suchas:

IFacountryispoliticallyunstable,orinwar,ANDthetrendofitsmacroeconomicdevelopmentinthelasttwoperiodsisawayfromthecentreoftheEMUcluster,THENtheriskthatthiscountrywillgoevenfurtherawayfromtheEMUishigh.Group2:ModulesfordiscoveringtrendsinthemacroeconomicdevelopmentoftheEMUclusterrelatedtothedevelopmentofotherclustersandothercountries.

HeretheconceptofEMUclusterisintroducedbasedontheEMUaggregateddataprojectedfromamultidimensionalspaceintoatwo(orathree)dimensionaltopologicalmapwiththeuseofself-evolving,self-organisingmaps.Thevectorsoftheused8parametersforthelast5-6yearofalltheEMUcountries,theotherEuropeancountries,someemergingmarketcountries,andalsotheUSAandJapan,aremappedintooneSOMorESOM.

Theneuron(thepointinthetwo-dimensionalmapspace)wherethecurrent-periodEMUvectorisprojectedisconsideredtobethecenteroftheEMUcluster.TheclusterincorporatesallpointswherethedataoftheindividualEMUcountriesaremapped.Theformoftheclusteranditsmovementfromoneperiodtothenextonecanbeobservedonthemapsandtheinformationwillbequantifiedbasedonthedistancebetweenthepoints.TheshapeoftheEMUclusteranditsdynamicscansuggestfurtherpoliticalandeconomicdevelopmentintheEMU.

ThemovementofthecenteroftheEMUclustercanbecomparedwiththemove-mentofthepointswherethedifferentcountriesaremapped.Aquantitativemeasureonthedifference(thedistanceinthetopologicalspace)betweendifferentcountriesfromtheEMU,andoutsidetheEMU,isevaluatedthatisusedasaseparateinfor-mationresultsaswellasaninputinformationtothehigher-leveldecisionmakingmodules.

DifferentclustersareformedonaSOM:theEMUcluster;theemergingeconomiescluster(e.g.,Poland);theclusterofthenon-Europeandevelopedcoun-tries(e.g.theUSA,Japan,Canada,Australia,NewZealand);theclusterofunder-developedcountries;theclusterofthedevelopednon-EMUcountries(e.g.theUK);theclusterofthedevelopingnon-EMUcountries(e.g.,Bulgaria,Romania).Avectoroffuzzymembershipdegreestowhicheachcountrybelongstoeachoftheclustersiscalculatedandtracedovertime.

Modulesfromgroup2coverdifferenttime-scales:annualmacroeconomicmap-ping,quarterlymacroeconomicmapandmonthlymacroeconomicmapping.

Thefollowingvariablesdescribethemacroeconomicstateofacountryinagivenperiodandallthevectorsforalltherelevantperiods(years,quarters,months)areusedintheunsupervisedtraining:GDP,debt,deficit,inflationrate,interestrate,unemployment,balanceofpayment,productiongap.

Group3:ModulesforevaluatingtrendsintheexchangerateEuro/US$

ThemainmoduleherepredictsthetrendoftheexchangeratebetweenEuroandtheUSdollar,butothermodulesdealwithnationalcurrenciesthatarenotpartoftheEMU.

Thefollowing10inputvariablesforexamplecanbeusedtopredicttherateR(t+1)ofEuro/US$,wheretisthecurrentperiod:R(t),R(t-1),Euro/JY(t),Euro/JY(t-1),ratioinflationrateinEMU/inflationrateintheUSAforboth(t)and(t-1)periods;ratiointerestratesinEMU/interestratesinUSAforboth(t)and(t-1)periods;ratioGDPinEMU/GDPinUSAforboth(t)and(t-1)periods.

Twotypesofmodelsareused-FuNNsandEFuNNs.Thetwomodelsusediffer-enttechniquesforextractingrulesandthemeaningoftheextractedrulesisdifferent.

IntherulesextractedfromFuNNstheconditionandconclusionelementshaveim-portancefactorsattachedtothempointingtotheimportanceofthedifferentpartsofarule.TherulesextractedfromEFuNNsarealsofuzzy,buttheypointtotheclus-tersintheinputandtheoutputspacethatarelinkedtogetherintherule.EFuNNsrequirelesstimefortrainingandcanbeupdatedveryquicklywithnewdatainanon-linemode.

Thepredictedtrendsoftheexchangeratescanbeusedeitherasseparateoutputresults,orasinputvaluesforthehigher-leveldecisionmodulesforbothquarterlyandmonthlytrendsprediction.

Group4:Modulesforevaluatingtrendsinmajorstockindexesandstockmarkets

Hereamapofthedifferentstatesofastockmarketaccordingto[50]iscreated.Thetransitionbetweenrandomwalkstate,chaoticstate,oracoherentstatewillbemodelledbytheuseofdifferenttechniques,thatinclude:hiddenMarkovmodels;evolvingfuzzyneuralnetworksEFuNNs;productionrules.

AnothermodulefromthisgroupevaluatesthevolatilityofmonthlytrendsintheDowJonesEuroSTOXX50index(DJE).Othermodulesevaluateweeklytrendsanddailyvalues.Thefollowing14inputvariablescanbeusedtoevaluatetheDJE,whereisthecurrenttimeperiod:DJE;DJE;;;Euro/US$;Euro/US$;Euro/JY;Euro/JY;Inflationrate;Inflationrate;GDP;GDP;Interestrate;Interestrate.

OthermodulesevaluatethetrendsinmajorEuropeanstockmarkets,suchastheItalianMIB30index(Milano).

6ImplementationandCurrentExperimentalResultswiththeEMU-HIDSS

TheimplementationoftheconceptualmodeloftheEMU-HIDSSfromsection5isaverycomplicatedtaskandalong-termobjective.Here,differentmodulesfromtheEMU-HIDSS/1conceptualmodelthatfollowthegeneraldescriptionandthelogicallinkspresentedintheprevioussection,aredevelopedandresultsareexplained.

6.1Group1ModuleforStatistically-BasedHigher-LevelRisk

Analysis

ThismoduletakesdynamicinputinformationfromtheclustermapsofthepreviouslevelofprocessingandcalculatestheEuclideandistancebetweeneachcountry’srepresentationvectorandthecenteroftheEMUclusteroverconsecutiveperiodsoftime(years,quarters).InthiswaythecountriesthataremovingawayfromtheEMUclusterareindicatedalongwiththespeedatwhichtheyaremoving.ForexamplethedistancebetweentheclustercenterofthemainEMUcountriesandItalyfor1997canbeevaluatedas3.8andfor1998as3.3,whilethesamedistancebetweenITandJPforthesameperiodscanbeevaluatedas1.2and0.7respectively(seeFig.4).

Fig.4.Theannualmapofthe15countriesaccordingto5characteristics(DBT/GDP,DEF/GDP,Inflationrate,Interestrate,Unemployment).ThecontourshowsthecenteroftheEMUclusterfor1998.

6.2Group2,SOMModulesforVisualExplorationofthe

AnnualandQuarterlyMacroeconomicDevelopmentoftheEMUClusterRelatedtotheDevelopmentofOtherClustersandOtherCountries

Thefollowing5variablesdescribetheannualmacroeconomicstateofacountry:DBT/GDP,DEF/GDP,Inflationrate,Interestrate,Unemployment.TheSOMmodelwastrainedon15countriesdatafrom1992till1998.ItisseenhowthepointsofthemapofthemainEMUcountriesandUSAaremovingfromlefttoright.Fig.4showsalsothecontoursurroundingthecentreoftheEMUclusterfor1998.ItisobviousthatthefollowingEMUcountriesarewithinthecluster:OE,NL,DK,IR,UK,SD,BD,FRinadditiontotheUSAandtheUK.Butfourcountriesareoutsideit(IT,BG,GR,ES)withonlyESmovingintotherightdirectiontowardstheEMUclustercenter.

Fig.5showsthedirectioninwhichItaly(IT)ismovingovertheyearsfrom1992till1998.

Fig.5.Iso-graphsofthe15countriesannualdevelopmentmapandthedirectionofthedevel-opmentofItaly.

AnotherSOMmoduleistrainedonquarterlydataofthefollowingthreevari-ables:Inflationrate,Interestrate,Unemployment.Fig.7showsthemapandthelineofthequarterlydevelopmentofGermany(BD).

TheSOMmodulesaresuitableforvisualexplorationasSOMsprovidewithanefficienttoolforvectorquantisation,i.e.turningann-dimensionalspaceusuallyintotwodimensionalspacepreservingthesimilaritybetweentheinputvectorsacrossalltheattributesastopologicaldistancebetweenpointsofthemap.SOMshavesomedifficulties,mainly:(a)theyhaveafixedstructure;(b)theycannottoleratemissingvalues;(c)learningusuallytakesalongtime.

InordertoovercomethelimitationsoftheSOMs,hereevolvingSOMmodulesarealsoused.

6.3Group2,ESOMModuleforDynamicMappingof

MacroeconomicAnalysis

AnotherannualmapisevolvedwiththesamedatasetasintheSOMmodulesandisshowninFig.8.ThemapisfirstclusteredusingESOMalgorithm,andthenpro-jectedontoatwo-dimensionalplaneforvisualisationusingSammon’salgorithm.Hencebothclusteringanddatastructureareavailablewithinthemap.ThelayoutoflabellednodesisverysimilartothatofFig.4,buttheESOMmapgivesmore

(a)

(b)

Fig.6.Componentanalysis:(a)Thefirstcomponent(DBT/GDP)fromthemapoffigs.4and5showthatinthisrespectBG,ITandGRareinasimilarposition.SD97andSD98forman“island”intheEMUclusterwithagoodtendency.(b)TheInflationcomponentshowsadramaticincreaseinGreecefrom1996to1997and1998.

Fig.7.Thecompletemapofthequarterlydevelopmentofthe15countriesfrom1995till1999andthelineofthedevelopmentofGermany(DB).

Fig.8.TheESOMclustersofthemacroeconomiesofEMUcountries.

10.90.80.70.6SW940.50.40.30.2

Distribution of Rule Nodes in Input-Input Space (

IR98Training )

FN98

DK982

IR97US98

SD98

JP9224

DK97

UK98

US97IR96

8NL98DK96NL97

SD97

US96

IR94JP93

17OE97UK97OE92OE98ES98NL961SD96IR95DK95PT98US95JP94DK92

FR98BD98PT97IR92IR9322BD94ES97DK94

BD97DK93BD9216PT92US94

NL93PT96JP97

FR97BD96BD9514JP95BD93OE96FN96NL95NL94US93NL927FR92JP96OE93ES9223

UK96FR96ES96US9210

OE94OE9511FN95

FR95UK95

FN92PT9512PT94FR94FR935JP98FN94PT93UK92

18ES94

UK94

ES93

ES95

21SD95UK93SD92FN9320

FN97

BG98

BG97

GR984

3

IT98IT97

BG96

6GR97

15

BG95BG94

9

GR96

IT96IT95

BG92

BG93

IT92

SD94

IT93

IT94

19

GR94

13GR95

0.10

0

SD93

0.10.20.30.40.50.60.70.80.91

Fig.9.ClusteringofannualdatawithEFuNN.

explicitinformationonsimilarityofcountryperformancewhichcanberepresentedbydistancebetweenthenodes.

Byclippingweakconnectionsusingadistancethreshold,asshowninFig.8,clustersintheannualmacroeconomicperformancedatacanberevealed.Herewefindtwomajorclusters,theEMUclusterwithcountrieslikeFR,BD,FN,PTetc.plusUKandUS,andthefall-outclusterwithGR,IT,andBG.

6.4Group2,EFuNNBasedModuleforPredictionand

ClusteringoftheAnnualandQuarterlyEconomicDevelopmentofthe15Countries

HereEFuNNsareusedtodevelopmodulesthatcanpredictvaluesforallthefiveselectedattributesintheannualdevelopmentandthethreeselectedattributesforthequarterlydevelopment.Suchmodulesaretheannualmoduleandthequarterlymodule.Thefirstonetakestwoinputvectorseachof5variables(atthetimemomentand)andcalculatesoneoutputvectorof5elementspredictingwhatthevaluesforthesevariableswillbe.Fig.9showstheclusteringoftheannualdata(displayedwiththefirsttwovariablesDBT/GDPandDeficit/GDP)achievedintherulenodesofanevolvedEFuNN.ItisclearthatIT,GRandBGformaclusterwithhighDBT/GDPvalueandDEF/GDPrunningfromlowtohighvalues.TheclusteringofdatasamplesarequitesimilartothatoftheSOMmodule,butitismuchmorequicklylearnedintheone-passlearningEFuNNmodule.

Fig.10showstheannualEFuNNpredictorforthe15countriesruninanon-linemodetopredicttheDBT/GDPvaluesannually.

Fig.10.AnnualEFuNNpredictorforpredictionofannualDBT/GDPvalues.

Fig.11.Quarterlypredictorforthe15countriesandtheon-linepredictionofthefirstvariable(Inflationrate).

RulescanbeextractedfromEFuNNsthatarefuzzy,andtheypointtotheclustersintheinputandtheoutputspacethatarelinkedtogetherintherule.EFuNNsrequirelesstimefortrainingandcanbeupdatedveryquicklywithnewdatainanon-linemode.

Fig.11showsthe15countriesquarterlypredictor,whichcanpredictvaluesofthethreeselectedvariablesInflationrate,Interestrate,UnemploymentforanyofthecountriesinthefollowingquartergiventhedataforthecurrentandpreviousquarterandalsotheannualDBT/GDPandthepreviousyearDBT/GDPareentered.Thefirstoutputvariableisshownaspredictedinanon-linemode.Thesystemimprovesitspredictionovertime.

6.5Group4,EFuNNModuletoEvaluateTrendsinthe

ExchangeRateEuro/US$

ThemainmoduleherepredictsthemonthlytrendoftheexchangeratebetweenEuroandtheUSdollar.Thefollowing10inputvariablesareusedinordertopredicttherateofEuro/US$,whereisthecurrentperiod:,,Euro/JY,Euro/JY,ratioinflationrateinEMU/inflationrateintheUSAforbothand

periods;ratiointerestratesinEMU/interestratesinUSAforbothandperiods;ratioGDPinEMU/GDPinUSAforbothandperiods.

6.6Group4,EFuNNModuletoEvaluateTrendsintheDJE501MajorStockIndex

ThismoduleevaluatesthemonthlytrendsintheDowJonesEuroSTOXX50index(DJE).Thefollowing14inputattributesareusedtoevaluatetheDJE,whereisthecurrenttimeperiod:DJE;DJE;;;Euro/US$;Euro/US$;Euro/JY;Euro/JY;Inflationrate;In-flationrate;GDP;GDP;Interestrate;Interestrate.

Besidesthemoduleslistedabove,severalothermodulesarecurrentlyunderdevelopment.

7ConclusionsandDirectionsforFurtherResearch

Aframeworkofhybridintelligentdecisionsystemispresentedinthepaper.Byapplyingarepositoryofintelligentinformationprocessingmodulesimplementedinanagent-basedarchitecture,acasestudysystemEMU-HIDSSisbuiltforriskanalysisandpredictionofevolvingeconomicclustersinEurope.

TheEMU-HIDSSisdesignedtobeusedatdifferentlevelsofanalysisanddecisionmakingabouttheEMUandabouttherelevantchangesintheeco-nomicclustersofEuropeandtheworld,thatincludes:theEuropeanUnionlevel;theglobalworldeconomieslevel;nationallevel;companyandbanklevel.DataandsomeofthedevelopedmodelsareavailablefrominternetURL

http://divcom.otago.ac.nz/infosci/kel/CBIIS.html(Software-FinancialRiskAnaly-sisandPrediction).

Acknowledgements

ThisworkissupportedbyaresearchgrantoftheDepartmentofComputerandMan-agementSciences,UniversityofTrento,andalsobyagrantPGSF-UOO808fundedbytheNewZealandFoundationforScience,ResearchandTechnology(FRST).

References

1.Amari,S.,Kasabov,N.,Eds.(1998)Brain-likeComputingandIntelligentInformationSystems.SpringerVerlag.

2.Bollacker,K.,LawrenceS.,GilesL.(1998)CiteSeer:AnautonomousWebagentforautomaticretrievalandidentificationofinterestingpublications.Proc.ofthe2ndInter-nationalACMconferenceonautonomousagents.ACMPress,116-123.

3.Deboeck,G.(1999)Investmentmapsforemergingmarkets.In:N.KasabovandR.KozmaEds.Neuro-fuzzytechniquesforintelligentinformationsystems.PhysicaVer-lag(Springer),373-395.

4.Deng,D.,Kasabov,N.(1999)EvolvingSelf-organizingMapanditsApplicationinGen-eratingAWorldMacroeconomicMap.Proc.ofICONIP/ANZIIS/ANNES’99Interna-tionalWorkshop,UniversityofOtago,7-12.

5.Deboeck,G.,Kohonen,T.(1998)Visualexplorationinfinancewithself-organizingmaps,SpringerVerlag.

6.Eichengreen,B.,Rose,A.,Wyplosz,C.(1995)Exchangemarketmayhem:thean-tecedentsandaftermathofspeculativeattacks.EconomicPolicy,251-312.7.MonthlyBulletin(1999)EuropeanCentralBank.8.EvolversoftwarepackageforExcel,http://www.palisade.com/html/evolver

17.Ishikawa,M.(1996)Structurallearningwithforgetting.NeuralNetworks,9,501-521.18.Kasabov,N.(1996)Adaptableconnectionistproductionsystems.Neurocomputing,

13(2-4)95-117.

19.Kasabov,N.(1996)FoundationsofNeuralNetworks,FuzzySystemsandKnowledge

Engineering.TheMITPress,CA,MA.

20.Kasabov,N.(1996)Learningfuzzyrulesandapproximatereasoninginfuzzyneural

networksandhybridsystems.FuzzySetsandSystems,82(2),2-20.

21.Kasabov,N.(1998)TheECOSFrameworkandtheECOLearningMethodforEvolving

ConnectionistSystems.JournalofAdvancedComputationalIntelligence,2(6),1-8.22.Kasabov,N.(1998)ECOS:Aframeworkforevolvingconnectionistsystemsandthe

ECOlearningparadigm.Proc.ofICONIP’98,Kitakyushu,Japan,IOSPress,1222-1235.23.Kasabov,N.(1998)Evolvingfuzzyneuralnetworks-algorithms,applicationsandbio-logicalmotivation.InYamakawaandMatsumotoEds.,MethodologiesfortheConcep-tion,DesignandApplicationofSoftComputing.WorldScientific.271-274.

24.Kasabov,N.(1999)Evolvingconnectionistandfuzzyconnectionistsystemsforon-line

adaptivedecisionmakingandcontrol.In:AdvancesinSoftComputing-EngineeringDesignandManufacturing,R.Roy,T.FuruhashiandP.K.ChawdhryEds.Springer,Lon-don.

25.Kasabov,N.(1999)Evolvingfuzzyneuralnetworksforadaptive,on-lineintelligent

agentsandsystems.In:O.Kaynak,S.TosunogluandM.AngEds.RecentAdvancesinMechatronics.Springer,Berlin.

26.Kasabov,N.Fedrizzi,M.(1999)Fuzzyneuralnetworksandevolvingconnectionistsys-temsforintelligentdecisionmaking.Proc.ofthe8thInternationalFuzzySystemsAsso-ciationWorldCongress,Taiwan,17-20.

27.Kasabov,N.,Kozma,R.(1999)Multi-scaleanalysisoftimeseriesbasedonneuro-fuzzy-chaosmethodologyappliedtofinancialdata.InRefenes,A.,Burges,A.andMoody,B.Eds.ComputationalFinance1997.KluwerAcademic.

28.Kasabov,N.,KimJ.S.et.al.(1997)FuNN/2-afuzzyneuralnetworkarchitecturefor

adaptivelearningandknowledgeacquisition.InformationSciences,101(3-4),155-175.29.Kasabov,N.,Kozma,R.(1997)Neuro-fuzzy-chaosengineeringforbuildingintelligent

adaptiveinformationsystems.In:IntelligentSystems:FuzzyLogic,NeuralNetworksandGeneticAlgorithms.DaRuanEds.KluwerAcademicBoston/London/Dordrecht,213-237.

30.Kasabov,N.,Israel,S.,Woodford,B.(1999)Methodologyandevolvingconnectionistar-chitectureforimagepatternrecognition.In:Pal,GhoshandKunduEds.SoftComputingandImageProcessing,Physica-Verlag(Springer)Heidelberg,accepted.

31.Kasabov,N.,Song,Q.(1999)Dynamic,evolvingfuzzyneuralnetworkswith‘m-out-of-nactivationnodesforon-lineadaptivesystems.TechnicalReportTR99-04,DepartmentofInformationScience,UniversityofOtago.

32.Kasabov,N.,WattsM.(1997)Geneticalgorithmsforstructuraloptimisation,dynamic

adaptationandautomateddesignoffuzzyneuralnetworks.In:Proc.oftheInter.Conf.onNeuralNetworks(ICNN97),IEEEPress,Houston.

33.Kasabov,N.,Watts,M.(1999)Spatial-temporalevolvingfuzzyneuralnetworksSTE-FuNNsandapplicationsforadaptivephonemerecognition,TechnicalReportTR99-03,DepartmentofInformationScience,UniversityofOtago.

34.Kasabov,N.,Woodford,B.(1999)Ruleinsertionandruleextractionfromevolvingfuzzy

neuralnetworks:algorithmsandapplicationsforbuildingadaptive,intelligentexpertsystems.Proc.ofInt.Conf.FUZZ-IEEE,Seoul.

35.Kawahara,S.,Saito,T.(1996)Onanoveladaptiveself-organisingnetwork.Cellular

NeuralNetworksandTheirApplications,41-46.

36.Kohonen,T.(1990)TheSelf-OrganizingMap.ProceedingsoftheIEEE,78,1464-1497.37.Kohonen,T.(1997)Self-OrganizingMaps,secondedition.Springer.

38.Kozma,R.,Kasabov,N.(1998)Chaosandfractalanalysisofirregulartimeseriesembed-dedintoconnectioniststructure.In:Brain-likeComputingandIntelligentInformationSystems.S.Amari,N.KasabovEds.SpringerSingapore,213-237.

39.Kozma,R.,Kasabov,N.(1999)Genericneuro-fuzzy-chaosmethodologiesandtech-niquesforintelligenttime-seriesanalysis.In:SoftComputinginFinancialEngineering,R.Ribeiro,R.Yager,H.J.Zimmermann,J.Kacprzyk(Eds.),Physica-VerlagHeidelberg.40.Krogh,A.,Hertz,J.A.(1992)Asimpleweightdecaycanimprovegeneralisation.Ad-vancesinNeuralInformationProcessingSystems4,951-957.

41.LeCun,Y.,DenkerJ.S.,Solla,S.A.(1990)OptimalBrainDamage.InTouretzky,D.S.

Ed.AdvancesinNeuralInformationProcessingSystems,2,598-605.42.Lin,C.T.,LeeC.S.G.(1996)NeuroFuzzySystems.PrenticeHall.43.Mitchell,M.T.(1997)MachineLearning,MacGraw-Hill.

44.Moody,J.,Darken,C.(1989)Fastlearninginnetworksoflocally-tunedprocessingunits.

NeuralComputation,1,281-294.

45.Persaud,A.(1998)Globalforeignexchangeresearch.In:EventRiskIndicatorHand-book,JPMorgan,London,44-171.

46.Saad,D.Ed.(1999)On-lineLearninginNeuralNetworks.CambridgeUniversityPress.47.Sankar,A.,Mammone,R.J.(1993)Growingandpruningneuraltreenetworks.IEEE

Trans.Comput.,42(3),291-299.

48.Swope,J.A.,Kasabov,N.,Williams,M.(1999)Neuro-fuzzymodellingofheartrate

signalsandapplicationstodiagnostics.In:SzepanjukEd.,FuzzySystemsinMedicine,PhysicaVerlag,accepted.

49.Trippi,R.,TurbanE.Eds.(1993)NeuralNetworksinFinanceandInvesting.IrwinPro-fessionalPublications,NewYork.

50.Vaga,T.(1990)Thecoherentmarkethypothesis.FinancialAnalystsJournal,November-December,36-49.

51.Watts,M.,Kasabov,N.(1998)Geneticalgorithmsforthedesignoffuzzyneuralnet-works.In:Proc.ofICONIP’98,Kitakyushu,Japan,21-23.

52.Watts,M.,Kasabov,N.(1999)Neuro-genetictoolsandtechniques.In:Neuro-Fuzzy

TechniquesforIntelligentInformationSystems,N.KasabovandR.Kozma,Eds.,PhysicaVerlagHeidelberg,97-110.

53.Woldrige,M.,Jennings,N.(1995)Intelligentagents:Theoryandpractice.TheKnowl-edgeEngineeringReview,10.

54.Zirilli,J.(1997)FinancialPredictionUsingNeuralNetworks.ThomsonComputerPress,

London.