ETD

• ablation study
• adaptability
• foundation models
• generalization
• time series forecasting
• transformer-based models

This thesis investigates the adaptability of foundation models for time series forecasting through a fair comparison of state-of-the-art approaches, aiming to evaluate their real-world applicability.
It focuses on three main models, Chronos Bolt, Moirai, and TimesFM, representative of recent advances in large-scale pretraining for time series forecasting.
The analysis leverages the GIFT-Eval and Chronos benchmarks to ensure consistent zero-shot evaluation across domains and temporal granularities.
The experimental assessment evaluates their performance, efficiency, and generalization, highlighting trade-offs between accuracy and computational cost.
Finally, it examines their finetuning capabilities to assess adaptability and the creation of domain-specific experts.