No Thumbnail Available
Files
Bailly_24641600_2022.pdf
Open access - Adobe PDF
- 4.38 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Neural networks are commonly used nowadays for multiple tasks, among which is nonlinear regression. In this thesis, we apply it to nonlinear autoregressive time series forecasting and investigate the ARNN(p) model. We study its performances for one-step ahead forecasting (benchmarked against linear AR(p) fit) and provide a few guidelines for optimization and model selection. In particular, we extend some recent results on L2 penalization for neural networks to the ARNN(p) model and find out that it leads both theoretically and empirically to sparse networks, and has a few advantages over regular L1 penalization.