Research On Forecasting Models And Methods Based On The Fuzzy Time Series

With the development of the science and society, people have to meet greaterinformation which leads the situation of business or science to be more complex thanever before. To make decision before what happens, more and more people becomeinterest in forecast. Time series forecast, one of the results of this demand. However,traditional time series forecasting models are incapable of handling linguistic valueand the series with small amount. To solve this problem, some researchers proposedthe concept of fuzzy time series forecasting model based on the fuzzy set theory andthis model’s performance is so excellent when tackle with time series consist ofsamples whose value is linguistic value than more and more researcher are willing todo some work in this domain.However, most of existing fuzzy time series forecasting models’ performance interms of forecasting accuracy and scalability are still limited result from that mostresearchers do not take full consideration of the presentation of the data and ignoresome important situation, such as multi-attribute time series, corresponding externaltime series, ignoring the trend factor in time series and subjective method used topartition the discourse universe. Besides, uncertain data which is popular in manyapplication domains is taken no consideration.In this paper, we propose two novel fuzzy time series forecasting models to solvethe problems above. One is proposed based on the existing fuzzy time series modeldesigned for the forecasting of certain data, the existing models are improved byintroducing the fuzzy cluster method into the process of partitioning the discourseuniverse and taking full use of the trend factor extracted from the time series to helpus adjust the order of the fuzzy relation dynamically, the new model not only havehigher forecasting accuracy but also have wider adaptive scale. The other oneintroduce the concept of uncertainty into the time series, and then we have a timeseries which consists of uncertain data. This model’ training and forecasting methodsare then designed based on this kind of time series. For the reason that works in thisdomain are rare, we don’t have a suitable index to evaluate the model. For the sake ofproviding a baseline to other researchers, we also propose three alternative indexes toevaluate the performance of the model. Numbers of time series are selected as the final experiment datasets. The result ofthe experiment demonstrates that the first algorithm which applies the fuzzy clusteringmethod and takes the trend factor into consideration has better performance in termsof forecasting accuracy and scalibity compared with existing model. Besides, thesecond model’s experiment result shows that this model can handle the time seriesconsists of uncertain data effectively. Generally speaking, both of the modelsproposed in this paper could improve the performance of the fuzzy time seriesforecasting models.