One Model to Forecast Them All: Towards Universal Temporal Representations

Project ID: 2531bd1648

(You will need this ID for your application)

Research Theme: Artificial Intelligence and Robotics

Research Area(s): Artificial Intelligence Technologies
Artificial Intelligence and Robotics
Natural Language Processing

UCL Lead department: Computer Science

Department Website

Lead Supervisor: Vasileios Lampos

Project Summary:

Why this research is important

Time series forecasting is a crucial tool for effective decision-making. From predicting extreme weather events to anticipating financial market shifts or tracking the spread of disease, reliable forecasts can save lives, protect communities, and guide policy. However, today’s state-of-the-art forecasting models often fail to work well outside the specific domain they were trained on. Recent advances in “foundation models”—large-scale systems inspired by breakthroughs in natural language processing—offer a new way forward. These models have shown the ability to generalise across tasks by learning from massive, diverse data sets. Applying this idea to time series forecasting could transform our ability to model dynamic processes across domains, enabling forecasts that are both accurate and adaptable.

What you will be doing

The successful applicant will explore how to design and train foundation models for time series forecasting. This will involve (i) developing backbone architectures that can capture long-term patterns and relationships across different datasets, drawing inspiration from cutting-edge deep learning approaches, (ii) building diverse collections of time series for large-scale pre-training, and designing robust benchmarks that reflect real-world challenges such as irregular sampling and shifting distributions, (iii) advancing methods for quantifying uncertainty, ensuring forecasts are not only accurate but also reliable, and (iv) investigating methods such as transfer learning, in-context forecasting, and few-shot learning to adapt models quickly to new problems. Applications will focus on high-impact areas such as meteorology, climate science, finance, and health, while also testing whether universal “temporal representations” can transfer across very different domains.

Who we are looking for

We welcome applicants with strong quantitative skills and an interest in machine learning, statistics, or related fields. During this project, the PhD student will develop expertise in deep learning, probabilistic modelling and large-scale pre-training, while contributing to the next generation of forecasting systems.