Early time-series classification (ETSC) aims to make accurate predictions from partially observed time series as early as possible. Although various stopping mechanisms and feature learning strategies have been developed for ETSC, most existing methods assume access to sufficient labeled training data, which may be unrealistic in applications with limited annotation. Under limited supervision, learning an additional sample-level stopping module and extracting effective classification features can both become challenging. In this paper, we propose FETERS, a few-shot ETSC framework that selects a dataset-level stopping ratio through class-wise leave-one-out (LOO) evaluation on the support set and uses a penalty-based reward function to manage the accuracy-earliness trade-off, thereby avoiding the need to train an additional stopping module. FETERS further combines Rocket-based features with frozen Chronos representations for classification. Extensive experiments on 69 public datasets spanning 14 domains show that FETERS achieves state-of-the-art (SOTA) performance in the 5-shot setting, with the highest average harmonic mean (HM) and the best HM on 38 datasets, while outperforming the current SOTA method on 44 datasets. FETERS also remains competitive in the full-shot setting, demonstrating its effectiveness in managing the accuracy-earliness trade-off.
Figures & tables
Figure 1: Illustration of the accuracy–earliness trade-off in ETSC using a sample from the UCR Archive ( Dau et al. 2019 ) . Shorter observed prefixes yield earlier predictions and greater earliness.
Figure 2: Overview of the FETERS framework. The figure illustrates the training workflow, where the red path first performs ratio selection and the green path then trains the shared pipeline using the selected ratio r∗ . During inference, test samples are truncated using r∗ and processed by the trained model following the same shared pipeline.
Method
5-shot
Full-shot
Acc. ↑
Ear. ↓
Wins
Acc. ↑
Ear. ↓
Wins
EARLIEST
0.289
0.018
0
0.341
0.063
1
TEASER
0.599
0.305
26
0.671
0.212
13
UTSC
0.611
0.296
15
0.701
0.239
18
CALIMERA
0.536
0.113
4
0.702
0.171
22
FETERS (Ours)
0.628
0.213
26
0.725
0.197
22
Table 1: Average accuracy, earliness, and number of datasets with the highest accuracy in the 5-shot and full-shot settings.
Dataset
5-shot HM
Full-shot HM
EARLIEST
TEASER
UTSC
CALIMERA
FETERS
EARLIEST
TEASER
UTSC
CALIMERA
FETERS
FiftyWords
0.069
0.555
0.532
0.476
0.571
0.241
0.661
0.667
0.668
0.671
Adiac
0.128
0.724
0.728
0.747
0.743
0.064
0.745
0.753
0.771
0.748
Beef
0.360
0.657
0.627
0.787
0.873
0.331
0.709
0.664
0.805
0.925
CBF
0.545
0.231
0.788
0.811
0.808
0.495
0.645
0.809
0.826
0.814
ChlorineConc.
0.471
0.329
0.340
0.508
0.499
0.699
0.688
0.728
0.718
0.737
Table 2: Per-dataset HM comparison under the 5-shot and full-shot settings. Bold indicates the best result for each dataset and setting. Wins count the best results across datasets. AVG Mean reports the average HM over all datasets.
Figure 3: Critical difference diagram of HM in 5-shot setting.
Figure 4: Comparison of mean earliness and earliness change between the 5-shot and full-shot settings under the same parameter configuration.
Setting
Variant
Acc.
Ear.
HM
W / T / L
W / T / L
W / T / L
5-shot
w/o Chr.
28 / 4 / 37
33 / 8 / 28
19 / 14 / 36
w/o Roc.
11 / 0 / 58
19 / 1 / 49
7 / 0 / 62
Full-shot
w/o Chr.
26 / 9 / 34
8 / 46 / 15
24 / 10 / 35
w/o Roc.
7 / 3 / 59
18 / 18 / 33
5 / 3 / 61
Table 3: Ablation study of the hybrid feature representation. W/T/L denotes the number of datasets on which the corresponding ablated variant outperforms, ties with, or underperforms the complete FETERS model.
Setting
Penalty
Acc. ↑
Ear. ↓
HM ↑
5-shot
p=0.6
0.598
0.165
0.677
p=1.0
0.618
0.195
0.680
p=1.5
0.633
0.226
0.678
p=2.0
0.644
0.253
0.672
w/o penalty
0.721
0.594
0.436
Full-shot
p=0.6
0.688
0.151
0.747
Table 4: Ablation and sensitivity analysis of the penalty-based reward function under different values of p .
Setting
Ratio
Acc. ↑
Ear. ↓
HM ↑
5-shot
r=0.05
0.464
0.050
0.586
r=0.10
0.509
0.098
0.616
r=0.15
0.544
0.147
0.631
r=0.20
0.577
0.198
0.642
r=0.25
0.606
0.249
0.645
r=0.30
0.631
0.298
0.643
Table 5: Comparison between FETERS and globally fixed ratios from 0.05 to 0.30, focusing on the early-decision region where the accuracy–earliness trade-off is most relevant.
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.
Franco Martino O'Rourke, Ana Trisovic, Dimitris Bertsimas
In this paper, we propose the first VLMagentic reasoning framework for few-shot multimodal Time Series Classification (MarsTSC), which introduces a self-evolving knowledge bank as a dynamic context iteratively refined via reflective agentic reasoning. The framework comprises three collaborative roles: i) Generator conducts reliable classification via reasoning; ii) Reflector diagnoses the root causes of reasoning errors to yield discriminative insights targeting the temporal features overlooked by Generator; iii) Modifier applies verified updates to the knowledge bank to prevent context collapse. We further introduce a test-time update strategy to enable cautious, continuous knowledge bank refinement to mitigate few-shot bias and distribution shift. Extensive experiments across 12 mainstream time series benchmarks demonstrate that MarsTSC delivers substantial and consistent performance gains across 6 VLM backbones, outperforming both classical and foundation model-based time series baselines under few-shot conditions, while producing interpretable rationales that ground each classification decision in human-readable feature evidence.
Lin Li, Jiawei Huang, Qihao Quan +7
Sun Yat-sen University Zhuhai, Guangdong, China · Xiaomi Corporation Beijing, China · University of Science and Technology of China Hefei, Anhui, China +1
Developing foundation models for time series classification is of high practical relevance, as such models can serve as universal feature extractors for diverse downstream tasks. Although early models such as Mantis have shown the promise of this approach, a substantial performance gap remained between frozen and fine-tuned encoders. In this work, we introduce methods that significantly strengthen zero-shot feature extraction for time series. First, we introduce Mantis+, a variant of Mantis pre-trained entirely on synthetic time series. Second, through controlled ablation studies, we refine the architecture and obtain MantisV2, an improved and more lightweight encoder. Third, we propose an enhanced test-time methodology that leverages intermediate-layer representations and refines output-token aggregation. In addition, we show that performance can be further improved via self-ensembling and cross-model embedding fusion. Extensive experiments on UCR, UEA, Human Activity Recognition (HAR) benchmarks, and EEG datasets show that MantisV2 and Mantis+ consistently outperform prior time series foundation models, achieving state-of-the-art zero-shot performance.