Vitor Cerqueira is currently Post Doctoral fellow at the Faculty of Computer Science of the Dalhousie University, Canada.

You may get more information at Vitor’s Home Page

PhD Thesis

Title: Ensembles for Time Series Forecasting

Supervisor: Luis Torgo; Co-supervisor: Carlos Soares, University of Porto, Portugal

PhD Program Faculty of Engineering/UPorto

Finished Dec/2019

Abstract

A time series represents a collection of data points captured over time. This type of data is actively studied in many domains of application, such as healthcare, finance, energy, or climate. The generalised interest in time series arises from the dynamic characteristics of many real-world phenomena, where events naturally occur and evolve over time.

Uncertainty is a significant issue when analysing time series, which complicates the accurate understanding of their future behaviour. To cope with this problem, organisations engage in forecasting to drive their decision-making process. Forecasting denotes the process of predicting the future behaviour of time series, which allows professionals to anticipate scenarios and take pro-active measures. In this context, the aim of this thesis is to advance the state of the art of the literature in time series forecasting. Particularly, our research goal can be divided into two main parts: (i) forecasting the future numeric values of time series; and (ii) the anticipation of interesting events in time series in a timely manner, a task that is commonly known as activity monitoring. In both parts, we adopt an ensemble learning approach, the field of machine learning that combines different predictive models to address a given predictive task.

The first part is split into two steps. Initially, we study how to estimate the predictive performance of forecasting models. The most appropriate methodology is still an open research question. In this context, we contribute to the literature by presenting an extensive empirical analysis of different methods for estimating the performance of forecasting models. We then develop new methods for time series forecasting. To accomplish this, we leverage the idea that typically, all predictive forecasting models have strengths and limitations throughout a time series, and we adopt an ensemble learning approach to manage these. As such, several forecasting models are created according to different assumptions about the process generating the time series observations. These models are then weighed over time to cope with the dynamics of the data. We contribute to the literature by developing a meta-learning model designed to estimate the weights of each model in the ensemble. The proposed method is based on a regression analysis of the errors of these models. We argue that the developed method can better adapt to changes in the environment relative to the state of the art approaches. We also contribute to the literature by presenting a novel aggregation framework of forecasting models in an ensemble. This aggregation method explores the idea that, like individual models, different subsets of these models show a varying relative performance throughout a time series. Thus, the idea is to apply state of the art methods for the dynamic combination of forecasting models to these subsets, instead of applying them directly to the original set of models. We show the usefulness of the developed methods in an empirical manner, using a large set of time series from different domains of application.

Regarding the second part, we contribute to the literature of activity monitoring by developing a novel method based on layered learning. Layered learning works by dividing a predictive task into different sub-tasks, or layers, which are in principle easier to solve. A predictive model is then applied to each subtask. These models are then combined to make predictions about the original problem. We apply the proposed method to a case study in healthcare, where the objective is to predict impending critical health events, namely hypotension episodes and tachycardia episodes. These represent a significant cause of mortality in intensive care units, and it is essential to anticipate them. Based on the results in this case study, we conclude that the developed method is competitive with state of the art approaches.

More Information

You may get more information at Vitor’s Home Page