Towards Pragmatic Time Series Intelligence

This thesis aims to democratize time series intelligence by making advanced modeling capabilities accessible to users without specialized machine learning knowledge. We pursue this goal through three complementary contributions that build foundation models, improve our understanding of them, and address challenges emerging in their practical use.

We start by introducing MOMENT, the first family of open source time series foundation models capable of performing well on a variety of tasks on data from diverse domains with minimal supervision. We extend these models to handle long multivariate contexts and integrate multimodal data, enabling their application to complex real-world scenarios where traditional approaches often fall short.

Next, we examine what these foundation models learn by investigating their compositional reasoning abilities, representation structures, and encoded concepts. We identify practical insights that improve both our understanding of the models and their performance.

Then, we tackle deployment challenges by developing methods to learn from distributed unlabeled data, assess label quality, and select robust models when labeled data is scarce. We conclude this thesis by exploring how Large Language Model agents can automate the time series intelligence engineering process, using open-source tools and tools developed in this thesis.

We demonstrate the utility of our methods in clinical settings, where time series data is plentiful and where modeling it can be impactful. We conclude that specialized foundation models, combined with practical tools supporting their real-world deployment, can substantially advance time series intelligence and yield practical solutions of societal importance.