Time series modelling and prediction using fuzzy trend information

Thispaperpresentsanovelapproachto modellingtimeseriesdatasetsusingfuzzy trendinformation. A time series is describedusingnaturallinguistic termssuchas rising more steeplyandfalling lesssteeply. Thesenaturalshape descriptorsenableus to producea glassboxmodelof the series.Thelinguistic shapedescriptorsarerepresentedin our systemby a new featurecalledthe trendfuzzyset. All trend fuzzysetsare derived from a window on the time seriesandassuchdescribethe shapeof the serieswithin that window. Eachwindow canhave membershipin any numberof differenttrendfuzzysets. Predictionusingthese trendfuzzysetsis performedusingtheFril evidentiallogic rule. Examplesof seriespredictionareshown usingsine waveandsunspotime seriesdata.