TableTime: Reformulating Time Series Classification as Training-Free Table Understanding with Large Language Models

Time series classification has attracted considerable attention in both academia and industry (Ismail Fawaz et al., 2019a; Liu et al., 2024b). Early approaches primarily relied on distance-based methods, such as Dynamic Time Warping (DTW) (Kate, 2016) in combination with K-Nearest Neighbors (K-NN), which are effective in addressing temporal distortions in time series data. To overcome the limitations of these methods, ensemble techniques were introduced to improve classification accuracy. For instance, HIVE-COTE (Lines et al., 2018) is an ensemble learning algorithm that enhances performance by combining multiple feature transformations and classifiers through hierarchical voting, thereby capturing diverse aspects of the data. With the advent of deep learning, more sophisticated models began to surpass traditional methods. Early architectures, such as Fully Convolutional Networks (Wang et al., 2017; Ismail Fawaz et al., 2020) and Recurrent Neural Networks (Hüsken and Stagge, 2003), directly learned hierarchical representations from raw time series data. More recently, models like InceptionTime (Ismail Fawaz et al., 2020) have employed deeper networks with multi-scale convolutions, significantly improving the ability to capture complex patterns across different time scales. Additionally, Transformer-based models (Zhou et al., 2021; Cheng et al., 2023) have emerged as powerful alternatives, excelling at capturing long-range dependencies and global context, further enhancing classification performance.

Given the impressive capabilities of LLMs, researchers in the time series classification community are increasingly exploring their applications in time series analysis (Zhou et al., 2023; Cheng et al., 2024a). LLM-based approaches can be divided into two categories: fine-tuning and generative modeling. Fine-tuning methods, such as Linear Fine-Tuning, combine pre-trained LLMs with time series-specific encoders, leveraging their linguistic capabilities to identify patterns. Generative models, like GPT-based forecasting, predict future time series sequences, while models like TEMPO integrate domain knowledge (e.g., seasonal-trend decomposition) to enhance performance. Despite these advances, challenges remain, including the modality gap between textual and numerical data, which hinders tokenization and semantic understanding (Zhou et al., 2023; Cheng et al., 2024a). Quantization techniques address this but may fail to capture temporal dependencies. Fine-tuning for multivariate or long-horizon time series also incurs high computational costs, and many approaches overlook domain-specific knowledge.

Despite effectiveness, LLMs in time series analysis encounter several significant challenges (Yao et al., 2024; Xing et al., 2024; Tan et al., 2024). First, they struggle to capture temporal dependencies and channel-specific features, which are essential for accurate modeling. Second, a misalignment exists between numerical time series data and LLMs’ semantic space, which complicates processing. Third, fine-tuning these models is computationally expensive, particularly for large-scale applications. Lastly, LLMs fail to fully leverage their reasoning capabilities, limiting their potential. To address these issues, we propose TableTime, a paradigm for time series analysis through table understanding.

Let $\mathbb{D}=\left\{(X^{1},y^{1}),(X^{2},y^{2}),...,(X^{n},y^{n})\right\}$ be a dataset consisting of $n$ pairs $\left(X^{i},y^{i}\right)$, where each $X^{i}\in\mathbb{R}^{t\times m}$ represents a multivariate time series with $t$ time steps and $m$ features, and $y^{i}\in\{c_{1},c_{2},\dots,c_{k}\}$ is the corresponding label. The goal of the classification task is to learn a classifier on $\mathbb{D}$ that maps the input space $X$ to a probability distribution over the class $y$. In the context of TableTime, we propose using prompt engineering based on the characteristics of the dataset and task. Let $P$ denote the prompt, the model’s output can be summarized as follows: $T^{i}=\text{LLM}(P,X^{i})$ where $T^{i}$ is the text generated by the large language model in response to the prompt and the input time series $X^{i}$. To predict the label $y^{i}$ for each instance, we apply a text evaluation process (such as keyword recognition) to $T^{i}$. The predicted label is then extracted as $\hat{y}^{i}$.

Recent advancements in large language models (LLMs) have unveiled a broad range of powerful capabilities, enabling them to address a variety of complex tasks. In this context, we propose leveraging LLMs to enhance multivariate time series classification (MTSC). Several key advantages arise from utilizing LLMs to advance classification techniques. World Knowledge: Pre-trained on vast amounts of textual data, LLMs can integrate general world knowledge into their predictions. This extensive knowledge base enables LLMs to provide contextual insights that are often absent in conventional methods, which typically rely on domain-specific information; Reasoning: LLMs exhibit advanced reasoning and pattern recognition abilities, which can potentially improve classification accuracy by capturing higher-level concepts. In contrast, existing non-LLM methods are predominantly statistical and often lack inherent reasoning capabilities; Training-free Inference: LLMs have demonstrated remarkable training-free inference capabilities, showcasing their potential to generalize across domains without the need for task-specific retraining. In contrast, existing classification methods are often highly domain-specific, which limits their adaptability and generalization potential; Text Generation: LLMs also possess strong text generation abilities, which can be harnessed to generate relevant features or even synthetic data for training purposes, thereby enhancing the robustness of classification models.

An overview of the TableTime is shown in Figure 2. TableTime introduces an innovative approach to multivariate time series classification (MTSC) by leveraging large language models (LLMs). To enhance reasoning capabilities, we first employ neighbor retrieval to identify relevant neighbors, improving the understanding of LLMs. These raw numerical time series are then converted into tabular format, preserving both temporal and channel information. To guide the LLM’s reasoning process, we design a comprehensive prompt that includes contextual text, neighbor knowledge, and task decomposition. In the final step, TableTime applies zero-shot reasoning to classify the data based on the provided information, without requiring task-specific retraining. Compared to other LLM-based models, our approach preserves both temporal and channel-specific information, retaining more of the original data. Additionally, by reformulating the time series as text format, our method aligns with the LLMs’ semantic space, facilitating better integration. Furthermore, TableTime does not require any training, enabling stronger generalization capabilities. These features demonstrate the powerful capabilities of TableTime.

A primary challenge of LLMs lies in their inability to effectively classify unseen samples due to the lack of task-specific examples, resulting in uncertainty and difficulty in capturing complex data patterns. To address this, we introduce the neighborhood-assisted in-context reasoning mechanism. By retrieving relevant neighbors from the training data, we provide essential contextual guidance to LLMs. These neighbors, including both positive and negative samples that share key features or contrasting patterns with the test sample, serve as reliable references to guide the generation of more accurate predictions.

Specifically, we employ two neighbor retrieval strategies: (1) positive samples only; (2) contrast enhancement. For the former, we apply similarity measurement algorithms to identify the $k$-nearest neighbors from the training dataset for each test sample. The retrieval process can be formalized as follows:

where $X^{test}$ denotes test sample; $\mathbb{D}$ denotes the training dataset; $f$ denotes the similarity measuring algorithm, e.g., euclidean distance and $k$ denotes the number of the neighbors.

To further assist LLMs in classification, we introduce contrast enhancement mechanism that incorporates negative samples into the model’s reasoning process. We begin by clustering the training dataset using algorithms such as K-means. Then, we select negative samples from clusters that do not contain the test sample, ensuring that these samples are sufficiently dissimilar. The process of K-means clustering can be described by the following formula:

where $C_{k}$ denotes the $k^{th}$ cluster of training dataset $\mathbb{D}$, $\mu_{k}$ denotes the centroid of the $k^{th}$ cluster, $K$ denotes the numbers the clusters. By introducing these negative examples, TableTime leverages contrastive learning to help the LLMs better differentiate between classes and refine its decision boundaries.

Contextual text information and neighborhood serve complementary roles in enhancing LLM reasoning. Contextual neighbors provides task-specific, data-driven context by identifying samples from the training data, directly guiding classification decisions through pattern and feature alignment. In contrast, contextual text description leverages the broad, pre-trained semantic understanding of LLMs to establish task rules and logical frameworks, offering a warm up for LLMs. By combining the precision of neighbor-based contextual information with the generalization capabilities of domain context, TableTime achieves robust and accurate classification.

Ensemble methods have demonstrated their effectiveness in time series classification. Similarly, self consistency (Wang et al., 2022) leverages multiple outputs from one LLM, selecting the most consistent response to ensure coherence and accuracy. In the meanwhile, due to the inherent randomness in LLMs’ responses, directly applying them to classification tasks remains challenging.

Building on the aforementioned considerations, we introduce a multi-path ensemble enhancement strategy. Multi-path inference utilizes multiple distinct reasoning pathways to generate a diverse set of outputs, thereby capturing a wider range of feature representations. By aggregating these results, the approach mitigates the potential biases inherent in any single inference path, improving both the robustness and accuracy of predictions. This makes multi-path inference particularly well-suited for our task, as it effectively handles the complexities of multivariate time series data. The multi-path ensemble enhancement can be formalized as follows:

where $M$ denotes the number of models, $f_{i}$ denotes the $\text{i}^{th}$ classifier, $y_{\text{final}}$ denotes the final classification result and $\mathbb{I}$ denotes the indicative function. Each classifier is an LLM. We can use different parameters to integrate, or integrate results from different LLMs. By generating multiple results using varied models and aggregating them through voting mechanism, we ensure that the final classification reflects the model’s most confident and consistent reasoning.

In the following, we summarize the characteristics of TableTime and discuss its relations to LLM-based and Dist-based models.

Relation to LLM-based Models. LLMs have shown great superiority in sequence modeling tasks. However, the huge number of parameters in LLMs makes training very difficult. Therefore, how to efficiently use LLMs has become an urgent problem. Existing methods learn embeddings for time series in LLMs’ latent space from scratch or map from external models to align with LLMs. Although effective, they cannot represent the original time series in a lossless manner. In contrast, TableTime can encode the raw time series fully and achieve zero-shot classification.

Relation to Traditional Models. Distance-based methods are traditional but effective and highly interpretable. However, distance-based methods are sensitive to noise and outliers and have difficulty identifying local patterns. They also lack the ability to effectively represent the features of the sequence and cannot fully explore the potential structural information in the sequence. In TableTime, we retain the advantages of distance-based methods and leverage LLMs to achieve understanding of raw time series. Compared with ensemble learning methods, TableTime obtains the final result by integrating the generated results of the same LLM under different parameter settings. Instead of using multiple models, we introduce multi-path ensemble strategy, which can enhance the robustness of the model and make the final result more accurate.

Table 2 summarizes the classification accuracy of all compared methods, and Figure 4 reports the critical difference diagram as presented in (Demšar, 2006). Compared to other baseline models, the experimental results demonstrate that our proposed TableTime achieves competitive performance and notable advantages on several datasets. For each dataset, the performance of TableTime is either the most accurate one or very close to the best one. These existing proposed models typically cannot always achieve the most distinct results. One may wonder whether the TableTime can be effective enough. However, the experimental results are largely consistent with previous empirical studies (Bagnall et al., 2017; Ruiz et al., 2021b), i.e., one single model cannot always achieve superior performances in all scenarios.

In particular, we observe that TableTime could surpass other baselines by a large margin in datasets like AF and SWJ, in which the size of the training and the testing dataset are both small. This also indirectly confirms the powerful zero-shot reasoning ability of LLMs, since deep learning-based methods are difficult to optimize due to insufficient data. Moreover, our method consistently outperforms Time-LLM and Time-FFM, further demonstrating that TableTime effectively leverages the reasoning capabilities of LLMs to achieve superior classification performance. We also observed that TableTime lags behind the optimal methods on the BL and UWG datasets. We speculate that this discrepancy arises from the large scale of these datasets, which allows deep learning methods to capture more comprehensive and intricate features.

The multi-path ensemble module aims to enhance classification robustness and accuracy by aggregating predictions from diverse inference paths generated through varying the temperature of the LLM. The ensemble method addresses the variability and uncertainty in single-path outputs by combining multiple predictions, effectively mitigating errors and improving generalization. By leveraging the diversity among predictions, the module ensures more reliable and robust classification outcomes.

To generate diverse paths, the module varies the LLM’s temperature parameter, exploring a range from deterministic outputs at low temperatures to more diverse predictions at higher temperatures. These outputs are aggregated using techniques such as majority voting, where the final prediction reflects the most consistent result across paths. This approach incorporates uncertainty analysis to capture diverse perspectives, ensuring consistent and accurate reasoning even in complex or ambiguous tasks.

To evaluate the impact of problem decomposition, we conduct ablation experiments by removing this component from the prompt.

Problem decomposition systematically breaks down complex tasks, such as time series classification, into smaller, manageable steps. This ensures that the model effectively processes temporal consistency, evaluates channel-specific features, and integrates relevant information for decision making, thereby reducing the risk of misallocating attention to irrelevant patterns.

As shown in Table 3, results indicate that removing this mechanism leads to noticeable drops in both accuracy and consistency, particularly on datasets with high-dimensional and complex temporal patterns. The structured reasoning enabled by problem decomposition not only enhances interpretability but also strengthens the model’s ability to generalize, demonstrating its critical role in achieving robust and reliable performance.

This investigation aims to reveal the relationship between LLMs’ predictions and the labels of its nearest neighbors. Understanding this relationship is essential because it provides insights into how much the model relies on nearest neighbors for accurate classification and how it performs independently when there is disagreement. Such an analysis helps to assess the model’s robustness and decision-making ability, especially in cases where the reference (nearest neighbor) label might not align with the LLMs’ prediction.

In our study, we divide the model’s classification results into two groups: those that matched the nearest neighbor’s label and those that did not. For each group, we analyzed the proportion of correct and incorrect classification results.

As shown in Figure 7, the results highlight that our model effectively leverages the reasoning capabilities of LLMs. Even when inconsistent neighbors are retrieved, the classification accuracy remains above 50% (60.0% for FM, 58.3% for SRS2), demonstrating the robustness of our proposed TableTime. This indicates that, while consistent neighbors enhance performance, our model can still achieve reliable predictions through LLM reasoning, even in the presence of noisy or incorrect neighbor samples.