A System Forecasting Currency Exchange Rates

ASYSTEMTOFORECASTCURRENCYEXCHANGERATESBYKRISHNANP.SANKARANAThesisPresentedtoTheHongKongUniversityofScienceandTechnologyInPartialFulllmentoftheRequirementsfortheDegreeofMasterofPhilosophyinComputerScienceHongKongJanuary1996CopyrightcbyKrishnanP.Sankaran1996AuthorisationIherebydeclarethatIamthesoleauthorofthethesis.IauthorisetheHongKongUniversityofScienceandTechnologytolendthisthesistootherinstitutionsorindividualsforthepurposeofscholarlyresearch.IfurtherauthorisetheUniversityofScienceandTechnologytoreproducethethesisbyphotocopyingorbyothermeans,intotalorinpart,attherequestofotherinstitutionsorindividualsforthepurposeofscholarlyresearch.iiASYSTEMTOFORECASTCURRENCYEXCHANGERATESBYKRISHNANP.SANKARANAPPROVED:DR.B.WUTHRICH,SUPERVISORPROF.ROLANDT.CHIN,ACTINGHEADOFDEPARTMENTDepartmentofComputerScience22January1996iiiACKNOWLEDGEMENTSWearedeeplygratefultoUnionBankofSwitzerlandbranchHongKong,andespeciallytotheheadoftreasurymanagementoce,SeahHawayKeng,andhisteam.FinallywewouldliketothankinformationtechnologymanagerP.Hugliwhotriggeredthisfascinatingresearchproject.ivTABLEOFCONTENTSTitlePageiAuthorisationPageiiSignaturePageiiiAcknowledgementsivTableofContentsvListofFiguresviAbstractvii1Introduction12NovelForecastingTechniques42.1Overview42.2DerivedFunctionDenitions72.3Thresholds102.4GlobalOptimizations113TheSystem164Experiments204.1DataCollection204.2ForecastingBidRateClosing204.3ForecastingVolatility224.4ComparisonwithotherTechniques235Conclusions28References29vLISTOFFIGURES1DataSheet142StartupScreen173LearnFunctionwithtimepoints174TexteditortopopulateREPLACEstatements185ForREADFILE186ForStoredfunctions19viASYSTEMTOFORECASTCURRENCYEXCHANGERATESBYKRISHNANP.SANKARANAThesisPresentedtoTheHongKongUniversityofScienceandTechnologyInPartialFulllmentoftheRequirementsfortheDegreeofMasterofPhilosophyinComputerScienceHongKongJanuary1996ABSTRACTNoveltechniquestoforecastdailycurrencyexchangeratesandasystemthatincorporatesthesetechniquesarepresented.Forecastingsaregeneratedfromautomaticallygeneratedprobabilisticdecisionrules,thresholdvaluesandglobaloptimizationtechniquesformultiplepredictelearning.Asextensiveexperimen-tationsreveal,thesystem’spredictionaccuracycomparesfavourablywiththoseofactualcurrencytraders.Itisalsoshownthatthepresentedtechniquesoutper-formstatisticaltechniquesinthisspecicapplicationdomain.viiCHAPTER1INTRODUCTIONThesedaysnancialmarketsarebecomingincreasinglyglobalandinterconnected.BankerscanalmostsimultaneouslybuycurrenciesorstocksinHongKongandselltheminSingapore.Thisglobalizationleadsmoreandmoretotheneedofautomatingandcomputerizingthebankingandsecuritiesindustry.Inthisthesiswelookatoneaspectofthiscomputerizationandautomationofworldnancialmarkets.Wesuggestnoveltechniquesforforcastingcurrencyexchangerates.Currencyexchangetradersaroundtheworldhavetoinstantlymaketheirdecisions,tobuyorsell,duringtradinghours.Sotheyareheavilyinterestedinpotientialsystemsthatcansupporttheirdecisions.Suchasystemcould,forinstance,predictwhether,fromthecurrenttimepoint,aparticularcurrencywillappreciateordepreciatewithinthenexthour.Eventhoughthisisintra-dayforecastingofcurrencies,thetraderswouldalreadybesupportedwhentheyknowwhetherattheopeningofHongKong’smarkettheexchangerateoftheYenagainsttheUSdollar,asquotedonWallStreetovernight,willriseorfallduringthetradinghoursinHongKong.Theapplicationwediscussheremakessuchdailypredictions.Thatis,thetaskistopredictat9amlocaltimeinHongKongwhetherthebidrateoftheYen,DEMandSFRwillbeaboutthesame,higherorloweragainsttheUSdollarat4pmHongKongtimewhenthelocalmarketcloses.Butbeforedescribingourdecision-supportsystemandtheexperimentalresultsweachieved,wereviewsomeworkdonepreviouslytoproduceshort-termforecastsofcurrencyexchangerates.[11]discussesempiricaltestsontheabilityofforeignexchangemarketpartic-ipantstoforecastthefuturevalueoftheAustraliandollarforone-andfour-weekhorizons.TheAustraliandollarisknowntobeoneofthemostvolatilecurrencies1intheworld.ThestudyisbasedonsurveydatapublishedbyTheAustraliannewspaper.Theybasicallyconcludethattheexchangerateisarandomwalk.Sothecurrentvalueisthebestpredictionforthefuturevalue.[11]hasalsotriedtocombinetheforecastingsofthirtydierenttraderssoastomakeanoverallandmorereliableprediction.Butasitturnedout,thiscombinationalsocouldnotimprovethepredictionaccuracy.[13]isanicetextbookgivinganoverviewonrecentpopulartechnicalanaly-sismethods.Technicalanalysisisconcernedwithforecastingreversalsintrendsfromcharts.Suchmethodsincludethelocalizationofaheadandshoulderpat-terninordertodetectanexchangeratefall,theuseofpeak-and-troughanalysistopredictbearandbullmarkets,andsoon.Themaindierencetoourapproachisthattechnicalanalysisisonlybasedonquantiableinformation.QuantiableinformationisforinstancethecurveshowinghowtheUSdollarperformedoverthelasttwoyearsagainsttheJapaneseYen.Wewillalsorelypartlyonquan-tiableinformation,butmainlytrytomakeforecastingsfromunmeasurableinputprovidedbydomainexperts,thetraders.Anunmeasurablefactorisforinstancemarketanticipationorgeneralpoliticalnews.Themostinnovat