RocketPFN: Accurate Time Series Classification via In-Context Learning

2026-06-19

AI Summary

This work addresses the high computational and data demands of time series classification (TSC) by proposing an efficient, high-accuracy method that requires no task-specific training. It introduces a novel two-stage in-context classification pipeline that uniquely combines training-free random convolutional feature extraction (Rocket) with a pretrained tabular foundation model (TabPFN v2.5). As the first zero-shot TSC foundation model, it achieves an average accuracy of 0.900 on the UCR/UEA benchmark—on par with the current state-of-the-art method HC2—and substantially outperforms other TSC foundation models. With a median inference time of only 30 seconds per fold, this approach establishes a new paradigm and strong baseline for efficient zero-shot time series classification.

Abstract

We introduce RocketPFN, a training-free pipeline for time series classification that combines random convolutional feature extraction (Rocket) with in-context classification via a pretrained tabular foundation model (TabPFN v2.5). On 92 UCR datasets (30-resample protocol), RocketPFN matches HC2, the strongest published method on the archive, in mean accuracy (both 0.900, Wilcoxon p=0.50), with no training on the target data and a median inference time of 30 seconds per fold. It also significantly outperforms every individual classifier in the HC2 ensemble. On UEA (20 datasets) the difference is likewise not statistically significant. A separate comparison concerns TSC foundation models: when paired with the same downstream classifier, MOMENT, Mantis, and MantisV2 are all significantly outperformed by RocketPFN using fewer extracted features and no learned parameters (p<0.001 in each case). This holds even when the encoders were pretrained on corpora that include the UCR training samples. We propose this two-stage pipeline as a reference point for evaluating zero-shot TSC foundation models.

Problem

• time series classification
• in-context learning
• zero-shot learning
• foundation models
• training-free

Innovation

• in-context learning
• training-free
• time series classification
• foundation model
• random convolutional features