Time series forecasting Lab

Major Competencies

The Time Series Forecasting Lab is a research team at the University of Reykjavik working on foundational models for time series forecasting, imputation, classification and anomaly detection. We train large models on a variety of time series data to infer broader temporal patterns across multiple domains. This eclectic approach allows us to further fine-tune our models for downstream usage. Our work encompasses the entire pipeline, from architecture building to training orchestration.

‍

High-performance Compute for architecture validation:

We employheavily modified transformer architecture and test the various additionsthrough focused ablations. Many of the innovations worked at the lab areinspired from the flourishing literature of large language models. HPC allowsus to validate and stress test new components.

‍

Large-scale pretraining on a diverse corpus:

Our pre-training methods rely on a cross-domain, various corpus of datasets composed of many time series from different fields: seismology, energy consumption, neuroscience,.. At this scale, high-performance compute, monitoring and fast disk speeds are key factors to successfully train a checkpoint.

‍

Multi-task and cross-learning:

We work on the 4 main tasks of timeseries analysis: forecasting, imputation, classification and anomaly detection.This transversal approach allows our model to cross-learn through differenttasks, and refine the signal representation within the model.

‍