TIMEE: End-to-end Time Series Classification via In-Context Learning

Time series classification (TSC) has traditionally relied on a two-stage paradigm: training a feature encoder on a specific dataset and then fitting a classifier on top. This approach is computationally expensive, requires per-dataset training, and prevents the model from using labeled examples to inform its internal representations at inference time. The authors introduce TimEE, an end-to-end foundation model that replaces this pipeline with in-context learning (ICL). By observing a labeled support set at inference, TimEE directly outputs class probabilities in a single forward pass, eliminating the need for weight updates.

Because high-quality labeled time series data is scarce, the authors developed a novel synthetic data generation pipeline based on Vector Autoregressive models with exogenous inputs (VARX). This allows the model to learn from tasks where class identity is tied to meaningful differences in the underlying data-generating process—specifically through structural variations in dependency graphs and signal warping. By training on 7 million synthetic tasks, TimEE learns to classify time series without ever seeing real-world data during its pre-training phase.

Despite its purely synthetic training, TimEE ranks first in ROC AUC and third in accuracy across the 128 datasets in the UCR benchmark. It outperforms both traditional supervised baselines and larger foundation models that rely on real-world pre-training. This result demonstrates that synthetic priors can effectively capture the structural dynamics of time series, providing a scalable and efficient alternative to the standard two-stage classification paradigm.