Source-linked AI summary
Large Language Models for Time Series: A Survey
Xiyuan Zhang, Ranak Roy Chowdhury, Rajesh K. Gupta, Jingbo Shang
TL;DR
LLMs are being explored for time series analysis, but their textual training creates a gap in interpreting numerical data. This survey organizes knowledge-transfer methods, datasets, comparisons, and research directions for applying LLMs to time series.
Problem
LLMs’ textual training leaves a gap in understanding how they can effectively interpret numerical time series, despite growing interest in applying them beyond language and vision.
Method
The survey develops a five-category taxonomy, reviews representative methods and multimodal datasets, compares approaches, and outlines guidelines based on data, model, efficiency, and optimization.
Results
The survey systematically categorizes knowledge transfer from LLMs to numerical time series as prompting, quantization, aligning, vision bridging, and integration with analytical tools.
Takeaways & Limitations
The survey identifies multimodal and multitask analysis, customization, privacy preservation, and broader research directions as opportunities for expanding LLM-empowered time series analysis.
Takeaways & Limitations
Existing work mostly addresses single modalities and single tasks, without simultaneously supporting multimodal and multitask time series analysis.
Abstract
from arXiv · showhide
Large Language Models (LLMs) have seen significant use in domains such as natural language processing and computer vision. Going beyond text, image and graphics, LLMs present a significant potential for analysis of time series data, benefiting domains such as climate, IoT, healthcare, traffic, audio and finance. This survey paper provides an in-depth exploration and a detailed taxonomy of the various methodologies employed to harness the power of LLMs for time series analysis. We address the inherent challenge of bridging the gap between LLMs' original text data training and the numerical nature of time series data, and explore strategies for transferring and distilling knowledge from LLMs to numerical time series analysis. We detail various methodologies, including (1) direct prompting of LLMs, (2) time series quantization, (3) aligning techniques, (4) utilization of the vision modality as a bridging mechanism, and (5) the combination of LLMs with tools. Additionally, this survey offers a comprehensive overview of the existing multimodal time series and text datasets and delves into the challenges and future opportunities of this emerging field. We maintain an up-to-date Github repository which includes all the papers and datasets discussed in the survey.
1 Introduction
Time series analysis supports diverse domains and tasks, while LLM research motivates extending language-model capabilities to numerical sequences. The survey organizes approaches around bridging this modality gap and a five-category taxonomy.
- Time series analysis supports climate modeling, traffic management, healthcare monitoring, finance analytics, classification, forecasting, anomaly detection, and imputation.
- LLM successes in language and vision have motivated researchers to explore their potential for time series analysis.
- The survey emphasizes bridging LLMs’ discrete-text training with the continuous numerical nature of time series.
- Its taxonomy covers prompting, quantization, aligning, vision as a bridge, and tool integration, each targeting a stage of a typical LLM pipeline.
2 Background and Problem Formulation
The survey formulates time series tasks in terms of variable inputs, outputs, and model components. Inputs may combine time series and text, outputs may be time series, text, or numbers, and language, time series, and vision models are optimized through a loss function.
- Inputs combine time series x_s ∈ R^T×c with optional text x_t, where T is sequence length and c is the number of features.
- Outputs y may be generated time series, text for report generation, or numbers for classification and regression.
- The formulation represents language, time series, and vision models with f_θ, g_ϕ, and h_ψ, with f_θ typically initialized from pretrained LLMs.
- Parameters θ, ϕ, and ψ are optimized through a loss function L.
3 Taxonomy
The survey organizes LLM-based time series methods into five taxonomy groups, each targeting a different stage of an NLP-style pipeline. The groups include prompting, quantization, aligning, vision-based bridging, and tool integration, with prompting, quantization, and aligning represented in the supplied passages.
- The taxonomy categorizes methods into five groups and summarizes each group's formulations, representative works, advantages, and limitations.The five groups are prompting, time series quantization, aligning, vision-modality bridging, and LLM integration with tools.
- Prompting: Prompting treats time series as raw text and directly supplies them to existing LLMs through task-specific prompts.PromptCast converts numerical series into text prompts for sentence-to-sentence forecasting; related methods address mobility, energy, health, tabular classification, and anomaly detection.
- Time Series Quantization: Quantization converts numerical time series into discrete representations or predefined text categories before using language models.Approaches include VQ-VAE codebook indices, K-Means indices, frequency units, discrete bins, and text categories such as directional price-change labels.
- Time Series Quantization: Index-based quantization uses nearest codebook or cluster indices as tokens, with VQ-VAE and K-Means forming two illustrated subtypes.The encoded representation is mapped to nearest codewords, and the resulting indices are passed to language models; representative applications include ECG, EEG, forecasting, classification, translation, and audio.
- The taxonomy also includes methods that extend LLM applications through related modalities and analytical tools.Audio, speech, and music systems use dedicated encoders and feed their embeddings to LLMs, while the broader taxonomy includes vision bridging and tool integration.
- Aligning: Aligning methods train time-series encoders to connect time-series embeddings with language-model semantic spaces or feed embeddings into LLM backbones.The supplied examples use contrastive or other similarity losses, while backbone methods retain or use pretrained language-model components for downstream time-series or text outputs.
4 Comparison within the Taxonomy
The survey compares taxonomy categories using data availability, model scale, efficiency, and optimization considerations. Prompting suits zero-shot settings, whereas quantization or aligning is preferable when training data is available, with different computational and optimization trade-offs.
- Data: With no training data and a zero-shot objective, prompting is preferable because it uses pretrained language-model capabilities without fine-tuning.The survey cautions that converting numbers to strings can reduce the semantic value intrinsic to numerical data.
- Data: When adequate training data is available, quantization or aligning methods become more advantageous than direct prompting.The survey identifies these as the most extensively studied categories in the existing literature.
- Model: Prompting and tool integration commonly use billion-parameter off-the-shelf models, while quantization and aligning span million- to billion-parameter models.Model scale varies with application requirements and available computational resources.
- Efficiency: Prompting is inefficient for high-precision, multivariate, and long-term numerical forecasting because it creates long inputs or generated sequences.Multivariate inputs may require separate univariate representations, while long-horizon predictions increase sequence-generation demands.
- Optimization: Quantization may require two-stage training, whereas aligning can support end-to-end training but may achieve different optimization behavior.Tool integration increases optimization complexity across the LLM and other components, so existing works commonly use off-the-shelf LLMs without further fine-tuning.
5 Multimodal Datasets
Multimodal time-series and text datasets support LLM-based analysis across IoT, finance, and healthcare. They pair sensor or biomedical signals with labels, descriptions, reports, metadata, or other natural-language information and support both time-series and language tasks.
- Internet of Things (IoT): Ego4D provides 3,670 hours of daily-life activity data, including IMU measurements and 3.85 million temporally aligned textual descriptions.Its scenarios span household, outdoor, workplace, and leisure activities.
- Internet of Things (IoT): Ego-Exo4D adds expert commentary, participant narrate-and-act descriptions, and atomic action descriptions paired with activity data.DeepSQA supports querying raw sensory data about human activities using natural language.
- Finance: PIXIU contains 136K financial multimodal instruction-tuning samples spanning nine datasets and language-understanding and prediction tasks.MoAT pairs timestep-level textual information, including relevant news articles, with financial time series covering fuel, metal, stock, and bitcoin domains.
- Healthcare: Zuco datasets pair eye-tracking and EEG recordings with natural-reading and annotation conditions, while PTB-XL provides expert-annotated ECG metadata and reports.PTB-XL includes diagnostic statements, diagnosis likelihoods, and signal-specific properties.
- These datasets support classification and forecasting alongside captioning, question answering, and translation.Reported evaluation metrics include accuracy and macro-F1 for classification, MSE, MAE, RMSE, and MAPE for forecasting, and BLEU, ROUGE, METEOR, and EM for language tasks.
6 Challenges and Future Directions
The survey identifies theoretical, multimodal, computational, statistical, customization, and privacy challenges that shape future LLM-based time series analysis.
- Theoretical Understanding: Empirical studies show LLMs can support non-language tasks and learn robust linear representations, but their interpretation of numerical time series remains theoretically underexplained.Transformer universality results and GPT4TS provide preliminary theoretical analysis, while the broader theory of numerical time series understanding remains incomplete.
- Multimodal and Multitask Analysis: Most existing systems address one modality and one task at a time rather than simultaneous multimodal and multitask analysis.The survey points to unified multimodal architectures in vision and audio as examples motivating more comprehensive time series foundation models.
- Efficient Algorithms: Multivariate or long-history time series increase computational complexity, while large patches can obscure semantic information and reduce performance.The survey therefore calls for more efficient algorithms for large-scale analysis and user interaction.
- Statistical Knowledge: Combining statistical knowledge, such as seasonal-trend decomposition and frequency-domain representations, with LLMs may improve time series analysis.TEMPO uses decomposed components as semantic inductive biases, while FreqTST tokenizes frequency units with weights for forecasting.
- Customization and Privacy: Future systems may benefit from user-specific customization built on global models, while privacy remains important for privately collected clinical and smart-home data.FedAlign is described as an initial federated-learning approach that aligns client latent spaces through natural-language class names.
7 Conclusion
The survey systematically organizes how knowledge from LLMs can be transferred to numerical time series analysis and reviews representative data, methods, challenges, and future directions.
- Conclusion: The survey presents five categories for transferring LLM knowledge to numerical time series: prompting, quantization, aligning, vision-based bridging, and tool integration.For each category, it discusses mathematical formulations, representative works, and advantages and limitations.
- Conclusion: It also reviews multimodal text and time series datasets across healthcare, IoT, finance, and audio domains.The paper concludes by outlining challenges and emerging research directions for LLM-empowered time series analysis.