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

Jul 16, 2026

Upload by Miles

Time series classification (TSC) is often solved with a two-stage pipeline that trains a feature encoder separately from the classification objective, fits a new classifier per dataset, and cannot exploit label structure at inference. TIMEE introduces a 4.5M-parameter foundation model that performs end-to-end TSC via in-context learning: given labeled support and a query series, it outputs class distributions in a single forward pass without per-dataset training. Meta-training uses only synthetic labeled tasks with structured distributional shifts, achieving top ROC AUC on the UCR benchmark and strong accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages31

Reading Time1 h 18 min

Abstract

Introduction

arXiv :2607 .07500v 1 [ cs .LG] 8 Jul 2026 TIMEE: End-to-end Time Series Classification via In-Context Learning Jaris Küken 1,2,* Shi Bin Hoo 1,* Martin Mráz 1 Frank Hutter 3,4,1 Lennart Purucker 3,1 1 University of Freiburg, 2 Zuse School ELIZA Darmstadt, 3 Prior Labs, 4 ELLIS Institute Tübingen * Equal contribution {kuekenj,[hoos}@cs.uni-freiburg.de](hoos}@cs.uni-freiburg.de) Abstract Time series classification (TSC) is dominated by a two-stage paradigm: train a feature encoder—either from scratch on the target dataset or via pretraining on large corpora—and then fit a task-specific classifier on top. While effective, this decoupling optimizes representation learning independently of the classification objective, requires per-dataset training, and prevents the model from exploiting label information during inference. We introduce TIMEE1 , a 4 .5M-parameter foundation model for end-to-end TSC via in-context learning. Given a labeled support set and a query time series, TIMEE directly outputs a predicted class distribution in a single forward pass with no per-dataset training required. Following the prior-data fitted network (PFN) framework, TIMEE is meta-trained exclusively on synthetic TSC tasks, where each task contains time series with distinct class identities arising from structured distributional shifts in the generative process. Despite seeing no real time series during pre-training, TIMEE ranks first in ROC AUC (and third on accuracy) on the UCR benchmark among all compared methods, which include both foundation models and supervised deep learning baselines. To our knowledge, TIMEE is the first purely synthetic-pretrained model to reach state-of-the-art performance on the UCR benchmark. These results establish end-to-end ICL with synthetic priors as a compelling, largely unexplored direction for TSC, with scaling, prior design, and richer generation mechanisms as natural avenues for improvement. Code is publicly available at [https://github.com/automl/timee](https://github.com/automl/timee). 1 Introduction The dominant approach to time series classification (TSC) decomposes the problem into two stages: first, a feature encoder maps each time series to a fixed-dimensional representation; second, atask-specific classifier is trained on these representations to produce predictions. Despite the surfacelevel differences of existing methods, this two-stage design is universal. Classical methods such as MiniRocket [Dempster et al., 2021] and Hydra [Dempster et al., 2023], per-dataset deep learning methods such as TS2Vec [Yue et al., 2022], and recent foundation models such as Chronos-2 [Ansari et al., 2025], TiRex [Auer et al., 2025b], and MantisV2 [Feofanov et al., 2026] all instantiate it. We argue that this is not merely an implementation convention, but a structural bottleneck. The encoder is optimized independently of the classification objective, and a new classifier must be fitted for every new dataset. Most critically, at inference time, the encoder cannot directly access the labeled training samples and therefore cannot produce representations tailored to the dataset’s class structure. 1 TIMEE pronounced as "Timmy" Preprint. T EE InceptionTime MiniRocket MantisV2 Chronos-2 TiRex TS2Vec Moment NuTime Catch22 Hydra TICT DTW-1NN 0 2 4 6 8 10 12 14 Mean Rank over ROCAUC (95% CI) 0 2 4 6 8 10 12 14 Mean Rank over Accuracy (95% CI) MiniRocket Hydra T EE InceptionTime MantisV2 TS2Vec TiRex Chronos-2 Moment NuTime Catch22 DTW-1NN TICT | | Statistical | Deep Learning | FMs | | --- | --- | --- | --- | | | | | | | | Tuned Untuned | | --- | --- | | | | Figure 1: We compare mean rank across all 128 UCR benchmark datasets. TIMEE achieves the lowest mean rank on ROC AUC (left), outperforming all compared methods including task-specific and foundation model baselines, while ranking third on accuracy (right) . DTW-1NN is excluded

FAQs

What problem does TIMEE address in existing time series classification pipelines?

TIMEE targets limitations of the common two-stage paradigm, where representation learning is optimized independently of classification, requires per-dataset training, and cannot tailor representations to the dataset’s class structure during inference.

How does TIMEE make predictions without per-dataset training?

At inference, TIMEE observes a labeled support set together with a query time series and directly outputs a predicted class distribution in a single forward pass, without updating model weights for each dataset.

How is TIMEE trained if it sees no real time series during pre-training?

TIMEE is meta-trained exclusively on synthetic TSC tasks. Each task contains time series whose class identities come from structured distributional shifts in a generative process, enabling label-meaningful synthetic prior learning.

Jul 16, 2026

Upload by Miles