A dynamic ensemble approach based on trend analysis to COVID-19 incidence forecast

Traditional models have struggled to effectively predict COVID-19 incidence time series. Limitation arises from their difficulty in accurately capturing complex behaviors in COVID-19 data, including potential drifts, leading to underperformance in predictive accuracy. In response to this challenge, this paper introduces a novel approach, Dynamic Ensemble Selection based on Trend Classification (DESTC). DESTC addresses the shortcomings of traditional models by dynamically selecting the best specialist models for predicting different concepts within the COVID-19 series. The four main phases of DESTC involve classifying time series instances into trend categories, training ensemble models, evaluating and ranking models for each trend, and selecting the most appropriate models to forecast new test patterns based on their identified trend. Through experimental simulations involving eight datasets, DESTC demonstrated superior performance compared to seventeen literature single and ensemble approaches. These results highlight DESTC as a robust and competitive solution for forecasting COVID-19 incidence time series.