Will Federated Learning to be the next key breakthrough of Time-Series Foundation Models?

Author: Guodong Long, Australian Artificial Intelligence Institute, University of Technology Sydney, Australia

Date: Feb 12, 2026

Will Federated Learning Be the Next Breakthrough for Time-Series Foundation Models?

Federated learning has long been viewed primarily as a privacy-preserving technique. However, its true essence lies in something deeper: the intrinsic capability for collaboration among distributed participants. When we shift our perspective from privacy to collaboration, a more ambitious vision of federated learning emerges — and with it, new possibilities for innovation.

Let us momentarily set aside privacy preservation and explore a broader question:
Could federated learning unlock the next generation of time-series foundation models?

If this sparks ideas for papers, grants, or applications, I would be delighted to discuss further.

A Brief History of Time-Series Foundation Models

Foundation models are pre-trained machine learning models designed to capture general knowledge from large-scale datasets and can be fine-tuned for various downstream tasks. Large Language Models (LLMs), such as ChatGPT, are a prime example — trained on internet-scale data to model the structure of human language.

Naturally, researchers have extended this paradigm to other domains. We now see foundation models for:

• Text-to-image generation
• Video generation
• Multi-modal reasoning
• Domain-specific LLMs for health, finance, and law

Time series data — including weather forecasts, stock prices, medical monitoring signals, smart sensor data from aircraft and vehicles, and traffic predictions — represents another highly important domain. As expected, many time-series foundation models (TSFMs) have recently been developed.

Industrial Time-Series Foundation Models

In industry, the most mature time-series foundation models focus on weather forecasting. These systems often integrate multi-modal inputs such as time-series measurements, satellite imagery, and geospatial data.

There are also open-source efforts to build pure time-series weather models, although this area remains largely open for further exploration.

Another highly attractive application is financial market prediction, particularly in digital currencies where price volatility is extreme. In high-frequency trading systems, time-series data is often combined with event-trigger signals derived from large-scale social media and news streams.

However, these systems are typically domain-specific rather than truly general foundation models.

Challenges in Research on Time-Series Foundation Models

The research community, as usual, aims for ambitious goals. One vision is to aggregate all publicly available time-series datasets into a single, universal foundation model — a generative model capable of understanding diverse time-series structures.

This ambition mirrors what LLMs achieved in natural language: ingesting internet-scale corpora and learning the underlying logic of language.

Yet, the analogy breaks down.

Unlike natural language, where diverse datasets still share common linguistic structure, time-series datasets represent fundamentally different physical systems. Weather signals, stock prices, ECG data, and engine sensor outputs obey different generative mechanisms. Heterogeneity across domains is far more severe.

Even when two time-series datasets exhibit similar shape patterns (shapelets), the semantic meaning behind those patterns may differ entirely. As a result, naïve tokenisation strategies inspired by NLP often fail in time-series contexts.

This heterogeneity presents a fundamental bottleneck for building universal time-series foundation models.

A Federated Perspective on Time-Series Foundation Models

Here is where federated learning offers a new perspective.

Rather than forcing all heterogeneous datasets into a single monolithic model, we could:

• Train specialised local models tailored to specific physical systems
• Preserve domain-specific inductive biases
• Then integrate them through collaborative aggregation mechanisms

In other words, instead of a single universal foundation model, we may build a collective intelligence composed of many smaller expert models.

This distributed ensemble could function equivalently to — or perhaps more robustly than — a centralized foundation model.

Our recent work takes a first step in this direction. We propose a federated approach to training time-series models across heterogeneous domains, and this work has been accepted at ICLR 2026 [1].

Looking Forward

Federated learning may evolve from a privacy-preserving framework into a new paradigm for collaborative intelligence construction.

For time-series foundation models, where heterogeneity is intrinsic and unavoidable, this collaborative architecture may represent not merely an alternative — but a necessary breakthrough.

REFERENCES

[1] Shengchao Chen, Guodong Long, Jing Jiang, FeDaL: Federated Dataset Learning for Time Series Foundation Models, ICLR 2026, https://www.arxiv.org/abs/2508.04045