TSFMs

In June 2026, we introduced TS-ICL, a new Time Series Foundation Model (TSFM) developed at EDF Lab and designed to jointly address forecasting and imputation in a zero-shot manner.

• Paper: TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning (Le Naour* et al., 2026)
• Code: EDF-Lab/ts-icl

Joint work with Etienne Le Naour and Adrien Petralia.

Recent Time Series Foundation Models (TSFMs), such as Chronos-2, have significantly advanced zero-shot forecasting performance. However, support for imputation remains limited. At the same time, tabular foundation models adapted to time series, including TabPFN-TS and TabICL-TS, can naturally handle both forecasting and imputation, but their higher inference costs and lack of time-series-specific inductive biases leave room for dedicated TSFMs to address both tasks within a unified framework.

With TS-ICL, we aim to bridge the gap between these two families of methods. TS-ICL is a probabilistic time-continuous Transformer trained through in-context learning on both real-world datasets and synthetic causal priors. By formulating time series tasks as timestamp-aligned regression problems, it can:

• Forecast future values,
• Impute missing observations, a critical capability for sensor monitoring, historical reconstruction, and anomaly investigation,
• Incorporate covariates without task-specific retraining,
• Handle irregular and partially observed time series.

Across a collection of benchmarks, TS-ICL achieves state-of-the-art zero-shot imputation performance while remaining competitive with leading forecasting foundation models. The benefits are particularly pronounced when forecasting from incomplete historical observations.

Past investigations of time series foundation models focused on the crucial yet largely unexplored imputation task.

1. Our proposed architecture: MoTM, Mixture of TimeFlow Models (Le Naour* et al., 2025)
2. A large-scale empirical benchmark showcasing the strong performances of MoTM along with SOTA TABPFN-TS (Le Naour* et al., 2026)