Scalable Models For Probabilistic Forecasting With Fuzzy Time Series

Abstract

In the field of time series forecasting, the most known methods are based on pointforecasting. However, this kind of forecasting has a serious drawback: it does not quantifythe uncertainties inherent to natural and social processes neither other uncertaintiescaused by the data gathering and processing. Because this in last years the interval andprobabilistic forecasting methods have been gaining more attention of researches, speciallyon environmental and economical sciences. But these techniques also have their own issuesdue to the methods being black-boxes and requiring stochastic simulations and ensemblesof multiple forecasting methods which are computationally expensive.On the other hand, the data volume (number of instances) and dimensionality (numberof variables) have reached magnitudes even greater, due to the commoditizing of thecapturing and storing computational devices, in a phenomenon known as Big Data. Suchfactors impact directly on the model’s training and updating costs, and for time serieswith Big Data characteristics, the scalability became a decisive factor in the choosing ofpredictive methods.In this context the Fuzzy Time Series (FTS) methods emerge, which have been growing inrecent years due to their accurate results, easiness of implementation, low computationalcost and model explainability. The Fuzzy Time Series methods have been applied toforecast electric load, market assets, economical indicators, tourism demand etc. But thereis a lack on FTS literature regarding interval and probabilistic forecasting.This thesis proposes new scalable Fuzzy Time Series methods and discusses its applicationto point, interval and probabilistic forecasting of mono and multivariate time series, for oneto many steps ahead. The parameters and hyper-parameters are discussed and fine tunningalternatives are presented. Finally the proposed methods are compared with the mainFuzzy Time Series techniques and other literature approaches using environmental andstock market data. The proposed methods obtained promising results on point, intervaland probabilistic forecasting and presented low computational cost, making it useful for awide range of applications.