How Foundational are Foundation Models for Time Series Forecasting?
Nouha Karaouli, Denis Coquenet, Elisa Fromont, Martial Mermillod, Marina Reyboz
TL;DR
This work questions the universality of Time Series Foundation Models (TSFMs) for forecasting by showing that zero-shot generalization is highly dependent on alignment between pretraining and target domains. Through a comprehensive evaluation of TSFMs and a compact, scratch-trained model (SAMFormer) on synthetic sine-wave benchmarks and a real-world Elec_Consumption dataset, it demonstrates that large pretrained TSFMs do not consistently exceed small, domain-adapted models in domain-shifted scenarios. The findings highlight that a one-size-fits-all approach may be suboptimal for time series forecasting, and that efficiency and personalization benefits can be achieved with lightweight models when distributions diverge from pretraining. Practically, the paper advocates a deployment strategy that combines TSFMs when pretraining aligns with the task and domain-adapted models for domain mismatch or data scarcity.
Abstract
Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.