自适应混合高阶神经网络在汇率预测中的发展与性能评价(IJISA-V9-N8-8)

I.J.IntelligentSystemsandApplications,2017,8,71-85PublishedOnlineAugust2017inMECS()DOI:10.5815/ijisa.2017.08.08Copyright©2017MECSI.J.IntelligentSystemsandApplications,2017,8,71-85DevelopmentandPerformanceEvaluationofAdaptiveHybridHigherOrderNeuralNetworksforExchangeRatePredictionSaratChandraNayakKommuriPratapReddyInstituteoftechnology,DepartmentofComputerScience&Engineering,Ghatkesar,R.R.Dist.-500088,Hyderabad,IndiaE-mail:[e-mail]:28February2017;Accepted:10April2017;Published:08August2017Abstract—HigherOrderNeuralNetworks(HONN)arecharacterizedwithfastlearningabilities,strongerapproximation,greaterstoragecapacity,higherfaulttolerancecapabilityandpowerfulmappingofsinglelayertrainableweights.Sincehigherordertermsareintroduced,theyprovidenonlineardecisionboundaries,henceofferingbetterclassificationcapabilityascomparedtolinearneuron.Nature-inspiredoptimizationalgorithmsarecapableofsearchingbetterthangradientdescent-basedsearchtechniques.ThispaperdevelopssomehybridmodelsbyconsideringfourHONNssuchasPi-Sigma,Sigma-Pi,JordanPi-SigmaneuralnetworkandFunctionallinkartificialneuralnetworkasthebasemodel.TheoptimalparametersoftheseneuralnetsaresearchedbyaParticleswarmoptimization,andaGeneticAlgorithm.Themodelsareemployedtocapturetheextremevolatility,nonlinearityanduncertaintyassociatedwithstockdata.Performanceofthesehybridmodelsisevaluatedthroughpredictionofone-step-aheadexchangeratesofsomerealstockmarket.TheefficiencyofthemodelsiscomparedwiththatofaRadialbasisfunctionalneuralnetwork,amultilayerperceptron,andamultilinearregressionmethodandestablishedtheirsuperiority.Friedman‟stestandNemenyipost-hoctestareconductedforstatisticalsignificanceoftheresults.IndexTerms—HigherOrderNeuralNetwork,JordanPi-SigmaNeuralNetwork,RadialBasisFunction,Pi-SigmaNeuralNetwork,FunctionalLinkArtificialNeuralNetwork,GeneticAlgorithm,ParticleSwarmOptimization,ExchangeRatePrediction.I.INTRODUCTIONTremendousimprovementincomputationalintelligencecapabilitiessincelastfewdecadeshasbeenenhancedthemodelingofcomplex,dynamicandmultivariatenonlinearsystems.SoftcomputingmethodologieswhichincludeArtificialNeuralNetwork(ANN),EvolutionaryAlgorithms(EA),GeneticAlgorithms(GA),andfuzzysystemshavebeenappliedsuccessfullytotheareadataclassification,financialforecasting,creditscoring,portfoliomanagement,risklevelevaluationandarefoundtobeproducingbetterperformance.ArtificialNeuralNetworkhastheanalogywiththethinkingcapacityofhumanbrainandthusmimickingit[1,2].TheANNcanimitatetheprocessofhumanbehaviorandsolvenonlinearcomplexproblems.Thesepropertieshavemadeitpopularandarecommonlyusedincalculatingandpredictingcomplicatedsystems.Multilayerperceptron(MLP)isafeedforwarderrorbackpropagationANNwhichcontainsatleastonehiddenlayerofneurons.Ithasbeenestablishedasmostfrequentusedneuraltoolforsolvingmanyrealapplications.TheabilityofMLPtoperformcomplexnonlinearmappingsandtolerancetonoiseinfinancialtimeserieshasbeenwellestablished.However,sufferingfromslowconvergence,stickingtolocalminimaarethetwowellknownlacunasassociatedwithit.Inordertoovercomethelocalminima,morenumberofnodescanbeaddedtothehiddenlayers.Multiplehiddenlayersandmorenumberofneuronsineachlayeralsoaddmorecomputationalcomplexitytothenetwork.Also,variousfeedforwardandmultilayerneuralnetworksarefoundtobecharacterizedwithseveraldrawbackssuchaspoorgeneralization,nonlinearinput-outputmappingcapabilityaswellasslowrateoflearningcapacity.HigherOrderNeuralNetworks(HONN)havefastlearningproperties,strongerapproximation,greaterstoragecapacity,higherfaulttolerancecapabilityandpowerfulmappingsinglelayertrainableweights[3].HigherordertermsinHONNcanincreasetheinformationcapacityofthenetwork.Thisrepresentationalpowerofhigherordertermscanhelpsolvingcomplexnonlinearproblemswithsmallnetworksaswellasmaintainingfastconvergencecapabilities.Withtheincreaseinorderofthenetwork,theremayexponentialgrowthintunableweightsinHONNandhencemorecomputationtime.However,thereisaspecialtypeofHONNcalledasPi-Sigmaneuralnetwork(PSNN)usinglessnumberofweightshasbeenintroducedbyShinandGhoshin1992[4].GhazalidevelopedaHONNbasedfinancialforecastingmodelwhichperformedsuperiorascomparedtoconventionalmultilayerneuralnetworkbasedmodels72DevelopmentandPerformanceEvaluationofAdaptiveHybridHigherOrderNeuralNetworksforExchangeRatePredictionCopyright©2017MECSI.J.IntelligentSystemsandApplications,2017,8,71-85[5].Knowlesetal.in2009usedHONNswithBayesianconfidencemeasureforpredictionofEUR/USDexchangeratesandobservedthatthesimulationresultsaremuchbetterthanmultilayerapproachbasedmodels[6].However,thereisexponentialincreaseintherequiredhigherordertermswhichmayaffectthenetworkperformance.ButthePSNNdevelopedbyShinandGhoshin1992isabletoavoidsuchincreaseinnumberofweightvectorsalongwiththeprocessingunits[7].SomerecentapplicationsofPSNNarefoundin[8-10].ThePSNNhasbeensuccessfullyemployedsolvingseveraldifficultproblemsincludingpolynomialfactorization[11],zeroingpolynomials[12],classification[7,13],timeseriesforecasting[14,15].YongNieandWeiDengpr