Advances in forecasting model selection based on meta-learning

Abstract

Forecasting is one of the main tasks of time series analysis. Selecting forecasting models is a crucial decision-making step in many real-world applications. To obtain forecasts that best explain the variables of interest in the future, it is necessary to select the model that best fits a given time series. However, the selection task can be quite computationally expensive due to many available techniques and the possible limitation of highly specialized knowledge. This proposal aims at a hybrid forecasting approach, using meta-learning to select models for components of a time series derived from a decomposition. Meta-learning comes with the proposition of using machine learning techniques to extract knowledge from past tasks and make past tasks faster, more efficient, and better performance. Allied to this, methods with hybrid forecast have shown promising results in the literature. As a result, it is expected to obtain new hybrid models with good performance and a better understanding of the meta-learning process. (AU)