You are on page 1of 6

Plain

noise data: predict this!


Inthispaper,weexamineacaseinspiredbyanonlinediscussionwithoneoftheusersonoursupport forums:www.spiderfinancial.com/forums/financialtimeseries/forex/.Thecasedealswiththe predictabilityofthedailyclosingEUR/USDexchangerate,givenitspriorhistory. Tostart,weconstructedatradingstrategywherewebuyEURatmarketopen,andclosetheposition (sellingEUR)atmarketclose. TheFXEUR/USDspotmarketisveryefficientastradersconstantlymonitorandexploit(i.e.arbitrage) anyinformationintheratestimeseriesuntilitisuneconomicaltoexploitanyfurther.Asaresult,it comesasnosurprisetousthatthedailyreturnsdatalooksmuchlikeawhitenoisedistribution.This promptsthequestion:whatisnext? AssumingthatourmainobjectiveforforecastinganextdayEUR/USDclosingrateistogenerateprofits, wetookaslightlydifferentapproachanddevisedastrategywithstoplosssettingsthatnotonlylimit thedailyloss,butalsogenerateapositivereturn. Thisapproachhasproventobemorefruitfulandwillpavethewayformanymoreimprovements.

Daily Returns Analysis


Foroursampledata,weusetheFXexchangerateofEUR/USDbetweenMarch2010andApril9th,2012.

Inthispaper,weconsideratradingstrategythatbuysEUR(1EUR=FXrateinUSD)atmarketopen (MOO)andclosestheposition(i.e.sellsEUR)atmarketclose(MOC). Tutorialpredictthis 1 SpiderFinancialCorp,2012

Thedailyreturnstimeseriesshowsasteadymean(proxyby20dayWMA)andarelativelystable volatility(0.51.0%).Examiningthesummarystatistics,theaveragereturnisnotsignificantlydifferent fromzero,withasymmetricaldistributionandnormalliketails.

Fromthetableonthebottomright,wecanseethatdailyreturnsaremorelikeaGaussiandistribution, andthustheypossessnoinformation.

Tutorialpredictthis 2 SpiderFinancialCorp,2012

Discussion
ThedailyreturnsactmorelikeGaussiannoise,sothepastreturnsdonotrevealanyinformationthat canhelpusbetterpredictthefuturepoints.Howaboutnonlinearmodelingtechniques(e.g.neural networks,geneticalgorithms)? Inmyopinion,Iwouldratherexploreandbringnewinformationtotheanalysisthanspendmytime utilizingsophisticatedmodelingtechniques.Inouranalysis,wetrytoavoidtheoverfittingtrapandto buildarobustmodelforreliableforecasting.

Daily HighLow time series


Whilewerestrictedouranalysistomarketopenandclose,whichhappenatprecisetimesintheday,the HIandLOrepresenttheMaxMinratesoftheintradayrates.Furthermore,theyareimpossibleto identifyforcertain,sowecanttradeonthem. Nevertheless,thedailyHIandLOratesbringnewinformationthatisnotalreadyintheopenclosetime series;asaresult,wecanforecasttheirrelativelevelstoacertainextent.

Figure1:scatterplotforopenclosereturnsvs.HILOreturns

Inthefigureabove,pleasenotethatHILOreturnsarerelatedtothevolatilityoftheopenclosereturns.

Tutorialpredictthis

SpiderFinancialCorp,2012

Daily returns with stoploss


Usingthesampledata,weimplementedasellstoplossorderwithalevelsetonadailybasisatmarket open. Definition:Astoplossisanordertobuyorsellanasset(e.g.tradingsecurityorcommodity)oncethe priceoftheassetreachesaspecifiedprice,knownasthestopprice.ASellstoporderisusedhereto limitalossonatradingcommodityweown. UsingthedailyLOrate,weexaminewhetherthestoppricehasbreachedfortheday;ifitbreaches,we setthesalepricetostopprice(noslippageortransactionfeesisassumed),otherwisewecontinueto usethemarketclosepriceforthesaleprice. Inthisexercise,weexaminedthefollowingstoppricetechniques:(1)Fixedpercentagebelowthe marketopenprice(e.g.0.75%),(2)Fixedoffsetbelowthemarketopenprice. 1. Stop price is set at a fixed percentage below market open

Examiningthesummarystatistics,thedataexhibitspositiveskewandamarginallysignificantaverage.

Tutorialpredictthis 4 SpiderFinancialCorp,2012

Tooptimizethepercentagesettingsforthestopprice,weconstructedadatatablewherewevariedthe percentagebetween0.1%and2%.

Plottingthemeanandvolatilityofthetradingstrategy,wepicked0.75%asthebestchoice.Thissetting givesthenewtradingstrategyanannualreturnof11.6%andavolatilityof10.2%. Furthermore,thenewstrategysreturnsactlikewhitenoise(butnotGaussian),sowehavenotleftany informationonthetable. 2. Stop price is set at an offset below market open Inanattempttoimprovethestoppricestrategy,weusedtheEWMAestimateandtheWMAoftheHI LO.Wefollowedthesamesteps:generatestrategyreturns,useadatatabletoexaminedifferent settingsand,finally,pickthebestone.Theresultsarecomparabletothesimpleoneweillustrated above. Fordetailedanalysisandillustration,pleaserefertotheExcelworksheetfile(availableonline)for details.

Conclusion
Sofar,westartedfromaGaussiannoisedatasample,incorporateddailyLOrates,anddevisedanew strategywithstoplossthatnotonlylimitsthedailyloss,butgeneratespositivereturns. Inaddition,wecandeviseanotherstrategythatemploysaselllimitordertolockinprofits.Usingdaily HI,wecanexaminewhetherthelimitpriceisbreachedandconstructnewstrategyreturns. Tutorialpredictthis 5 SpiderFinancialCorp,2012

Howaboutastrategythatemploysboth:stoplossandlimitorder?Unfortunately,thedatasetwehave doesntcontaininformationaboutwhethertheHIoccursbeforeoraftertheLOthatday. ThetakeawayhereisthatwecanstartwithadataseriesthatlooksnomorethanGaussianwhitenoise, butwiththeintroductionofnewinformation(e.g.DailyHIandLO,volume,interestrateparity,G8 summitevents,etc.),wecanderiveanewtimeserieswithmorefavorablecharacteristics.

Tutorialpredictthis

SpiderFinancialCorp,2012

You might also like