DaSSWeb | Truths and myths in multivariate time series and panel data
Detalhes
Speaker Vitor Miguel Ribeiro FEP-U.Porto Abstract In the current period marked by the massive penetration of machine learning models trying to fulfill prediction and classification tasks in multivariate time series and
Detalhes
Speaker
Vitor Miguel Ribeiro
FEP-U.Porto
Abstract
In the current period marked by the massive penetration of machine learning models trying to fulfill prediction and classification tasks in multivariate time series and panel data problems, there is a risk associated with the inability to distinguish between innovation and imitation. On top of this stylized fact, a weakness associated with the development of artificial intelligence network architectures is the observation of an opaque communication language that is only perceptible by computer science specialists. This seminar explains in detail the importance of mathematical, statistical and econometric foundations that allowed the creation of two innovations – roll padding and input set splitting with subsequent use of convolutional layers inside an attention block – inherent to the development of a new deep learning architecture. An academic study and a real-life example highlight their importance.
Short Bio
Vitor Miguel Ribeiro is Assistant Professor in the Econometrics Group of the Department of Economics of the School of Economics and Management of the University of Porto, Portugal. He is also Associate Researcher at the Center of Economics and Finance of the University of Porto.