Costa, Nuno M. C. daNovais, FilipeSantos, LuísFranco, FranciscoShah, VaibhavRodrigues, RicardoFernandes, DuarteGouveia, Emanuel2026-04-232026-04-232026-01-019789897587986e72c9c9a-8349-4920-bfa5-3b4489d954c9http://hdl.handle.net/10400.14/57573<p>Driver-based multivariate time series forecasting aims to predict a target time series-the ”driven”-using multiple influencing variables-the ”drivers”. Traditional linear models often struggle to capture the complex, non-linear relationships in such data. While ensemble methods and meta-learning have been applied, they face limitations like high computational complexity and a narrow focus on model selection. This study addresses these challenges by introducing a new ensemble forecasting approach and extending meta-learning to predict not only the forecasting model but also other critical parameters such as input window size and feature selection methods. We conducted experiments using the publicly available ”Air Quality Monitoring in European Cities” dataset, comparing the proposed ensemble and meta-learning methods across various time series lengths and forecasting horizons (1, 12, and 24 hours). Results demonstrate that the Meta-Learner Forecasting approach outperforms the Ensemble Forecasting approach, especially in smaller datasets and shorter forecasting horizons, achieving improved forecasting accuracy. By extending meta-learning to predict multiple forecasting parameters, this research enhances the versatility and efficiency of multivariate time series forecasting, highlighting the importance of tailoring forecasting parameters to specific data characteristics. The Meta-Learner not only improves accuracy but also reduces computational costs by efficiently narrowing the search space for optimal parameters, making it applicable to more complex forecasting environments.</p>engDriver-based forecastingEnsemble methodsMeta-learningMultivariate forecastingTime series forecastingconference proceedings10.5220/0014634500004058105035532557