Teams link https://teams.microsoft.com/meet/39392823379690?p=Id7Us5YJWXbgc85tsO
Add to Google Calendar
Abstract
Despite several advances in time series forecasting algorithms, current evaluation practices show several limitations. Evaluation approaches often fail to provide reliable estimates of performance, undermining empirical studies and proposed methodologies. This talk explores four critical gaps between current evaluation methods and practical forecasting needs, specifically: 1) myopic evaluations that average forecasting accuracy across all instances; 2) the mismatch between forecasting accuracy and practical utility; 3) dataset selection bias; and 4) ignoring temporal constraints during evaluation.
Short BioVitor Cerqueira is a Researcher at the Faculty of Engineering, University of Porto, and member of the Laboratory for Artificial Intelligence and Computer Science (LIACC) and the Center for Responsible AI. Since 2015, he has specialized in time series analysis and automated machine learning, publishing in top-tier venues and developing innovative approaches for real-time anomaly detection, meta-learning for forecasting, and concept drift detection.
His research focuses on developing and benchmarking methods for time-evolving data, with applications spanning aquaculture, energy, bioinformatics, and industry. Notable contributions include award-winning work on mixture of experts in time series forecasting (ECML’17 best student paper) and influential research on forecast evaluation. He is the leading author of a book on deep learning for time series forecasting.
More information at News U. Porto