Improving time series forecasting accuracy by ensemble methods and time series features
Wideroos, Kimmo (2021)
2021
Master's Programme in Computational Big Data Analytics
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2021-12-10
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202112099077
https://urn.fi/URN:NBN:fi:tuni-202112099077
Tiivistelmä
Ensemble methods can improve prediction accuracy of machine learning models, but applying ensemble methods in time series analysis is not feasible due to high dimensionality of time series. Augmenting time series data with features that are extracted from the original time series is an approach that can make traditional machine learning methods applicable also for analysing time series.
Echo state network is a type of recurrent neural network with a special property that only its output layer is trained. In this research the weights of the echo state network output layer are used as time series features. These novel features are compared to more conventional set of time series features extracted using Tsfresh software library.
Several input feature combinations based on features derived from an echo state network and features extracted using Tsfresh library are evaluated in the experiments. Main differences in the input combinations are whether there are no time series features at all involved or there is a certain combination of features included. Additionally, different feature selection methods area applied. All analysis are conducted using a 5-fold cross-validation with validation and test set. Data set that is used in the research is from M4 time series analysis competition. The best submitted methods from the M4 competition are taken as the base learners for the ensembles.
It is shown that the time series features derived from an echo state network can improve the performance of ensembles for time series prediction. However, it is also shown that the conventional time series features outperform the echo state network features with a high margin. It is to be shown in a future research whether some nonlinear dimensionality reduction methods could improve the features derived from echo state network weights.
Echo state network is a type of recurrent neural network with a special property that only its output layer is trained. In this research the weights of the echo state network output layer are used as time series features. These novel features are compared to more conventional set of time series features extracted using Tsfresh software library.
Several input feature combinations based on features derived from an echo state network and features extracted using Tsfresh library are evaluated in the experiments. Main differences in the input combinations are whether there are no time series features at all involved or there is a certain combination of features included. Additionally, different feature selection methods area applied. All analysis are conducted using a 5-fold cross-validation with validation and test set. Data set that is used in the research is from M4 time series analysis competition. The best submitted methods from the M4 competition are taken as the base learners for the ensembles.
It is shown that the time series features derived from an echo state network can improve the performance of ensembles for time series prediction. However, it is also shown that the conventional time series features outperform the echo state network features with a high margin. It is to be shown in a future research whether some nonlinear dimensionality reduction methods could improve the features derived from echo state network weights.