Dynamic ensemble selection forecasting system based on trend classification

Abstract

Dynamic Ensemble Selection systems (DES) have been proposed as an useful alternative for modeling and forecasting time series. The basic idea is to assess the performance of single models and select the best ones for predicting a new test instance. One of the most common selection strategies involves constructing regions of competence (RoC). In this case, based on a new test instance to be predicted, one evaluates which instances from the training and/or validation set are most similar using a similarity metric. However, the absence of similar pat- terns between the test and training/validation sets compromises the quality of the RoC and adversely affects the predictive capabilities of these systems. Besides, the choice of which similarity measure to adopt is a complex and ongoing research problem. Consequently, the fol- lowing question arose: “How to conduct the selection phase considering structural changes in terms of trend in the time series, without relying on similarity measures?”. This thesis proposes a new DES approach, Dynamic Ensemble Selection based on Trend Classification (DESTC), which uses trend analysis to select the models to be combined. Trend is the prevailing direc- tion or pattern in data observed over time. DESTC consists of two main phases: the training phase (a), in which a pool of models is evaluated to determine the best ones for each trend class, and the testing phase (b), in which each new instance has its trend assessed, and the top-performing models are selected for prediction. To evaluate the predictive performance of DESTC, two experiments were conducted. In Experiment A, the proposed approach was ap- plied to COVID-19 incidence time series data from eight countries and compared with single and ensemble-based algorithms from the literature. The proposed approach achieved superior forecasting performance and lower computational cost. In Experiment B, DESTC was further evaluated on time series exhibiting distinct characteristics from various phenomena. The results demonstrated that DESTC is a competitive alternative to other Multiple Predictor Systems (MPS). The main limitation of the proposed method is that DESTC tends to have lower predictive performance when the time series lacks a clear trend cycle pattern, making model selection based on trend classification impractical. Moreover, the results presented and dis- cussed in both experiments demonstrate that the proposed method, DESTC, is a competitive alternative to other MPSs found in the literature.