Imputation by decomposition and by time series nature : novel imputation methods for missing data in time series

• Ano de defesa: 2021
• Autor(a) principal: Silvana Mara Ribeiro
• Orientador(a): Cristiano Leite de Castro
• Banca de defesa: Luis Antonio Aguirre, Frederico Gadelha Guimarães
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica
• Departamento: ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA
• País: Brasil
• Palavras-chave em Português: Missing data; Time series; Imputation methods; Decomposition; Pattern
• Link de acesso: http://hdl.handle.net/1843/46099; https://orcid.org/0000-0002-6754-4374

Dealing with missingness in time series data is a very important, but oftentimes overlooked, step in data analysis. In this dissertation, the pattern of time series data and missingness mechanisms are described to help identify which imputation method should be used to impute missing data, along with a review of imputation methods and how they work. Recommended methods from literature are used to impute synthetic data of different pattern and the results are discussed. In this dissertation, two new methods to impute missing time steps are presented and compared to other classical imputation methods, as well as state-of-the-art methods. The first imputation method presented is Imputation by Pattern. This method is based on the premise that imputing the data using the literature- recommended methods will achieve the best results. Heuristics are proposed to separate the time series by pattern. The second imputation method presented is Imputation by Decomposition. This method consists in decomposing the time series in its components and then imputing them using the literature-recommended methods. The combination of these methods and the Kalman filter are also tested. The discussed imputation methods are used to impute a financial indexes and instability trackers data set, a COVID-19 data set and a deng data set and then predictions are made and the results are presented. The Imputation by Pattern method combined with the Kalman filter achieved consistently satisfactory results, although it did not always achieve the best results. The Imputation by Decomposition method achieved good results, specially when some time was spent investigating which variation worked better with each data set. Overall, both imputation method achieved similar, and in some cases, better results than the classical imputation methods.