PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Models for Time Series and Spatio-Temporal Data: A Survey and Outlook

Ming Jin, Yaxuan Kong, Yuxuan Liang, Chaoli Zhang, Siqiao Xue, Xue Wang, James Zhang, Yi Wang, Haifeng Chen, Xiaoli Li, Vincent S. Tseng, Yu Zheng, Lei Chen, Hui Xiong, Shirui Pan, Qingsong Wen

arXiv:2310.10196v3cs.LGcs.AI

TL;DR

Whether large models can effectively analyze time-series and spatio-temporal data remains an open question. This survey reviews the field through a taxonomy separating LM4TS and LM4STD, and summarizes prominent techniques and their strengths and limitations.

  • Problem

    The survey addresses whether large models can effectively analyze time-series and spatio-temporal data, two broad and intrinsically linked categories.

    Source: §1 · p. 2, time series and spatio-temporal data? · p. 2

  • Method

    The paper surveys large models for time-series and spatio-temporal data using a taxonomy that separates LM4TS from LM4STD.

    Source: §3 · p. 9, §8 · p. 24

  • Results

    The review summarizes prominent techniques in each category and examines their strengths and limitations.

    Source: §8 · p. 24

  • Takeaways & Limitations

    The taxonomy and synthesis provide a structured account of the reviewed models and their strengths and limitations.

    Source: §8 · p. 24

    The theoretical basis for transfer remains limited, with unclear conditions for capturing temporal patterns or failing under distribution shift.

    Source: §7.1 · p. 23

Abstract

from arXiv · show

Temporal data, including time series and spatio-temporal data, are pervasive in real-world applications. Generated in massive volumes by physical and virtual sensors, they record dynamic system behaviors and enable a wide range of downstream tasks. Effectively analyzing such data is crucial to unlocking their rich information content. Recent advances in large language models and other foundation models have accelerated their use in time series and spatio-temporal data mining. These approaches not only improve pattern recognition and reasoning across diverse domains but also support progress toward artificial general intelligence that can understand and process temporal data. In this survey, we present a comprehensive, up-to-date review of large models tailored or adapted for time series and spatio-temporal data along four dimensions: data types, model categories, model scopes, and application areas/tasks. We organize existing work into two main groups: large models for time series analysis (LM4TS) and for spatio-temporal data mining (LM4STD), and further distinguish general-purpose from domain-specific models. We also curate related resources, including datasets, model implementations, and tools, organized by major application areas. Overall, this survey consolidates recent advances and highlights foundations, applications, resources, and open research opportunities in large model-centric temporal data analysis.

1 Introduction … served as precursors to the current large-model paradigm.

The introduction motivates large-model analysis of time series and spatio-temporal data by connecting expanding foundation-model capabilities with temporal data’s broad real-world relevance and distinctive spatial dependencies. It traces prior model limitations and emerging approaches while emphasizing data and integration challenges.

1 Introduction

  • GPT-3 demonstrated few-shot and zero-shot learning capabilities, unlike its 1.5-billion-parameter predecessor GPT-2.The introduction attributes this advance to LLMs’ ability to learn semantic and knowledge representations from large text corpora and reason across tasks.Source: §1 · p. 2
  • LLMs and related foundation models span language, vision, and audio, prompting the question of whether they can analyze temporal data.The introduction notes their success across diverse domains and tasks.Source: §1 · p. 2

time series and spatio-temporal data?

  • Time series and spatio-temporal data are pervasive, but spatio-temporal analysis must capture spatial dependencies alongside temporal dynamics.These data categories also arise from diverse platforms, including sensors.Source: time series and spatio-temporal data? · p. 2

holistic understanding of the inherent dynamics they encapsulate across various systems.

  • Earlier time-series and spatio-temporal methods, from statistical models to RNNs, CNNs, GNNs, and Transformers, were mostly small and task-specific, limiting broad transfer and multi-task reasoning.Self-supervised approaches including TF-C, SimMTM [42], STEP, and AdaMAE [11] improved transferability.Source: holistic understanding of the inherent… · p. 3

served as precursors to the current large-model paradigm.

  • Limited availability of high-quality, diverse large-scale datasets and distinctive integration challenges continue to hinder temporal-data analysis with large models.The introduction argues that these challenges make a comprehensive review tailored to time-series and spatio-temporal data necessary.Source: served as precursors to the current… · pp. 3–4
  • Pangu-Weather and ClimaX target climate modeling, TFM [223] is a transportation foundation model, and Valley [137], LAVILA, and mPLUG-2 address spatio-temporal video understanding.The video models use large-scale multimodal pre-training to capture complex temporal dynamics.Source: served as precursors to the current… · p. 3
  • Chronos [4], Time-MoE, and MOIRAI aim to provide general-purpose forecasting and representation learning across diverse domains.These are representative time-series foundation models.Source: served as precursors to the current… · p. 3
  • PromptCast, OFA, and Time-LLM adapt LLMs for time-series analysis, while OpenTSLM [98] and MedTsLLM target clinical applications such as medical event prediction.Approaches include prompting, lightweight fine-tuning, and reprogramming time-series inputs with natural-language prompts.Source: served as precursors to the current… · pp. 3–4

to examine the progress in this burgeoning field. · links, methodological differences, and transferable adaptation strategies. · without pre-training or large-model adaptation, and duplicate versions lacking distinct technical contributions.

The paper presents a unified survey of large models for time series and spatio-temporal data, organizing research by data category, model type, scope, application domain, and task. It also compiles resources, reviews the survey’s coverage and selection, and outlines future research directions.

to examine the progress in this burgeoning field.

  • The survey organizes work into LM4TS and LM4STD, then distinguishes LLMs from PFMs and classifies research by model scope, application domain, and task.This multi-faceted taxonomy is intended to provide a coherent roadmap across the field.Source: to examine the progress in this… · p. 4
  • The survey compiles datasets, open-source implementations, and evaluation benchmarks, and summarizes large-model applications across domains.The compilation is positioned as a reference for future research and development.Source: to examine the progress in this… · p. 4
  • It identifies future research opportunities spanning data sources, model architectures, training and inference paradigms, and other prospective directions.The discussion aims to characterize the field’s current state while highlighting avenues for further investigation.Source: to examine the progress in this… · p. 4
  • Unlike reviews that typically center on one modality or model family, this survey brings time series and spatio-temporal data into an integrated temporal-data perspective.It examines LLM repurposing and PFM pre-training across model scope, task type, application domain, and resource availability.Source: to examine the progress in this… · pp. 4–5

links, methodological differences, and transferable adaptation strategies.

  • The survey iteratively searched major scholarly databases and indexing platforms for work on LLM adaptation, PFMs, and large-model temporal-data analysis across four data types.Its scope included time series, spatio-temporal graphs, temporal knowledge graphs, and videos, covering work published or publicly released by submission time.Source: links, methodological differences, and… · p. 5

without pre-training or large-model adaptation, and duplicate versions lacking distinct technical contributions.

  • Candidate studies were expanded through backward and forward snowballing, with selection prioritizing early or influential contributions, peer-reviewed work, and widely cited emerging-topic papers.The survey excluded works with only incidental large-model relevance and conventional small-scale neural methods.Source: links, methodological differences, and… · p. 5
  • The paper proceeds from background and a unified taxonomy to specific approaches, then presents compiled resources and applications before future research directions.The taxonomy provides a broad overview before the paper delves into individual approaches.Source: without pre-training or large-model… · p. 5

this domain. Finally, we conclude our paper in Sec. 8. · 2 Background · with their representative tasks across different domains.

The background establishes the survey’s operational distinction between language-centered LLMs and broader foundation models, and defines the temporal data types and scope it examines.

this domain. Finally, we conclude our paper in Sec. 8.

  • Fig. 2 presents roadmaps of representative LLMs and other foundation models.The background describes the roadmaps as summarizing model development and main research interests.Source: Figure 2 · p. 6, §2 · p. 6

2 Background

  • The survey reviews large models for temporal-data tasks, including methods where a large model serves as a backbone, interface, or reasoning module rather than a newly trained standalone model.It distinguishes LLMs from PFMs according to their role in existing studies.Source: §2 · p. 6
  • LLMs are language-centered models repurposed for temporal tasks, whereas PFMs use backbones, pre-training data, or objectives designed beyond general language modeling.The survey’s operational distinction places temporal, spatial, relational, visual, and multimodal data within the PFM scope.Source: §2 · p. 6

2.1 Large Language Models

  • Time-series and spatio-temporal analytical tasks can use multimodal repurposing that aligns target-task and pretraining-task modalities, with adaptation by fine-tuning or frozen-LLM reprogramming.OFA and Time-LLM are repurposing examples.Source: §2.1 · p. 7

2.2 Pre-Trained Foundation Models

  • PFMs’ capabilities span modality bridging, reasoning and planning, and interaction, but temporal and spatio-temporal PFMs remain early-stage and far from the latter two capabilities.The survey focuses on temporal-data PFMs, while noting that vision-language models remain the most popular PFM research topic.Source: §2.2 · pp. 7–8

2.3 Time Series and Spatio-Temporal Data

  • Time series are ordered univariate or multivariate observations, while spatio-temporal data combine spatial and temporal dimensions; the survey focuses on STGs, TKGs, and videos.It also notes trajectories and point cloud streaming as other spatio-temporal forms.Source: §2.3 · p. 8

which are showcased in Fig. 1 and summarized below.

The section surveys core downstream tasks across time series, spatio-temporal graphs, temporal knowledge graphs, and video, spanning prediction, classification, completion, and retrieval.

Time Series Tasks

  • Time series analysis covers forecasting, classification, anomaly detection, and imputation; forecasting may be short- or long-term, and anomaly detection distinguishes abnormal from normal series.Imputation fills missing time-series values.Source: which are showcased in Fig. 1 and… · p. 9

Spatio-Temporal Graph Tasks

  • Spatio-temporal graph forecasting predicts future node features from historical attributes and structure; other tasks include traffic forecasting, on-demand services, link prediction, and node/graph classification.Link prediction estimates whether an edge exists from historical information.Source: which are showcased in Fig. 1 and… · p. 9

Temporal Knowledge Graph Tasks

Video Tasks

  • Video tasks include detection, captioning, anticipation, and querying: identifying objects or actions, describing content, predicting forthcoming frames, and retrieving query-relevant segments.These tasks often span multiple modalities and have received considerable attention compared with the previously mentioned data types; the section directs readers to.Source: which are showcased in Fig. 1 and… · p. 9

3 Overview and Categorization … or domain-specific purposes, as outlined in Fig. 3.

The survey classifies temporal foundation-model research by data type, architecture, scope, and application, separating time-series analysis from spatio-temporal mining. It then reviews LLM adaptation and newer multimodal approaches for time series.

3 Overview and Categorization

  • The survey divides large temporal-data models into LM4TS and LM4STD and organizes the literature by data categories, architectures, scopes, and applications or tasks.Fig. 3 and Table 2 summarize the taxonomy and related work.Source: §3 · p. 9
  • For time series, the taxonomy distinguishes repurposed LLMs from purpose-built PFMs, then classifies each as general-purpose or domain-specific.The LLM/PFM distinction is practical: LLM4TS repurposes language-centered models, whereas PFM4TS develops temporal foundation-model backbones.Source: §3 · pp. 9–10

domains, including but not limited to transportation, finance, and healthcare.

  • For spatio-temporal data, the survey applies the LLM/PFM distinction and explicitly groups models by domain because the data span varied entities and applications.Model-scope classification depends on modality-specific problem definitions, such as those for spatio-temporal graphs and videos, so it is omitted from Fig. 3.Source: domains, including but not limited to… · p. 10

most prominent domains/modalities, using them as subcategories: spatio-temporal graphs, temporal knowledge graphs,

  • The survey uses spatio-temporal graphs, temporal knowledge graphs, and video as prominent subcategories, with representative tasks as leaf nodes.PFM4STD is more developed than PFM4TS, with current work focused mainly on graphs and video and often adding multimodal bridging.Source: most prominent domains/modalities,… · p. 11

and reasoning.

  • The taxonomy establishes the framework for subsequent reviews of LM4TS and LM4STD, including LLMs and PFMs for graphs, knowledge graphs, and video.The paper frames these reviews as advancing understanding and reasoning about temporal data.Source: and reasoning · p. 11

4 Large Models for Time Series Data

  • The time-series section reviews LLM and pre-trained foundation-model advances, organizing each discussion by whether models serve general or domain-specific purposes.This section is divided into LLM4TS and PFM4TS discussions.Source: §4 · p. 11

4.1 Large Language Models in Time Series

  • The LLM4TS discussion addresses adapting language models to core time-series tasks such as forecasting, imputation, anomaly detection, and classification.It considers general-purpose and domain-specific models separately.Source: §4.1 · p. 11

4.1.1 General Models.

  • General LLM4TS adapts language models through prompting, tokenization, decomposition, or reprogramming; PromptCast uses natural-language inputs and outputs, while LLMTime [52] demonstrates effective zero-shot time-series learning.Source: §4.1.1 · p. 11

decision-making, signaling a shift from pattern matching to general time series reasoning. · contextualize, and interact with complex transportation data. · portfolio construction, and return prediction.

The passages survey large-model approaches across transportation, finance, and healthcare, alongside general-purpose and domain-specific time-series foundation models. They describe systems that combine temporal data with text, structured context, or learned representations for forecasting, interpretation, and reasoning.

4.1.2 Domain-Specific Models.

  • xTP-LLM [55] transforms heterogeneous traffic features—historical flow, temporal indicators, weather, and POI context—into structured textual prompts, enabling an LLM to jointly predict traffic flow and produce natural-language explanations.Source: §4.1.2 · p. 13
  • Traffic-R1 [295] uses traffic-operation logs for routing and congestion reasoning, LEAF incorporates graph topology and road semantics, and LLM-Mob [222] analyzes mobility trajectories and scenarios.Together, these models extend traffic forecasting toward interpreting, contextualizing, and interacting with transportation data.Source: §4.1.2 · p. 13

contextualize, and interact with complex transportation data.

  • M2VN combines price, volume, and news embeddings for volatility prediction, while DINN [70] integrates LLM-based semantics and a pretrained language model in a portfolio forecasting-and-allocation pipeline.The passages report improved equity-market volatility prediction from jointly modeling numerical and textual signals.Source: contextualize, and interact with… · p. 13
  • GPT4FTS [74] uses adaptive patch segmentation and dynamic wavelet-style modules for state-of-the-art stock-return forecasting; Lopez-Lira et al. [134] find robust return predictability from LLM-scored news headlines.The latter result suggests general-purpose LLMs can extract economically meaningful signals from financial text.Source: contextualize, and interact with… · p. 13

portfolio construction, and return prediction.

  • NYUTron, trained on NYU Langone EHRs and fine-tuned for clinical and operational tasks, significantly improves over traditional models and can guide care alongside physicians.The passages frame clinical event forecasting as support for physician and administrator decisions in everyday practice.Source: portfolio construction, and return… · p. 13
  • MedualTime and MedTsLLM connect medical time series with clinical text through dual-adapter and cross-modal learning, while OpenTSLM [98] treats time series as a native modality for natural-language queries and reasoning.These approaches aim to enrich temporal-contextual representations and support interpretation of longitudinal health data.Source: portfolio construction, and return… · pp. 13–14

4.2 Pre-Trained Foundation Models in Time Series

  • PatchTST uses a channel-independent Transformer over sub-series patches to improve long-term forecasting; general PFM4TS trends toward scalable zero-shot models supporting multivariate, covariate-informed inputs, and broader tasks.The survey characterizes this as a rapid convergence toward scalable general-purpose models with strong zero-shot performance.Source: §4.2 · p. 14

4.2.2 Domain-Specific Models.

  • SSPT tailors stock-series pretraining with a dual-layer Transformer and auxiliary stock-code, sector, and moving-average prediction objectives.The objectives are designed to capture both global stock-series information and sector-related structure.Source: §4.2.2 · p. 14

temporal patterns and stock-specific structure, improving robustness in non-stationary markets. DELPHYNE [40] · negative transfer and achieving strong performance on diverse downstream forecasting tasks.

The passages describe finance and healthcare foundation models that tailor pretraining to domain-specific time series and support forecasting or clinical transfer.

temporal patterns and stock-specific structure, improving robustness in non-stationary markets. DELPHYNE [40]

  • DELPHYNE [40] targets financial applications with pre-training on a mixture of general and finance-specific time series.The passage says this reduces something, but does not specify what.Source: temporal patterns and stock-specific… · p. 15

negative transfer and achieving strong performance on diverse downstream forecasting tasks.

purely general-purpose time series models. … adapting LLMs to large dynamic networks.

The survey organizes spatio-temporal large models around spatio-temporal graphs, temporal knowledge graphs, and videos. Within graph modeling, it reviews LLM-based approaches and pretrained foundation models spanning general-purpose and domain-specific applications.

5 Large Models for Spatio-Temporal Data

  • The survey covers spatio-temporal graphs, temporal knowledge graphs, and videos, categories with broad real-world applications.It frames these as three major categories of large models for spatio-temporal data.Source: §5 · p. 15

5.1 Spatio-Temporal Graphs

  • STGNNs model spatial correlations with graph neural networks and temporal dependencies with models such as RNNs or CNNs; LLMs and PFMs add textual context and multimodal fusion.These models can also generate human-interpretable explanations.Source: §5.1 · p. 15

streamlining both training and inference processes.

  • Recent graph-coupled LLM forecasting approaches aim to streamline training and inference while combining language semantics with graph topology.The reviewed work describes this as a shift from purely GNN-based designs toward LLM-centric, graph-aware prediction; practical limits include graph construction quality, spatial-dependency stability, and cost.Source: §5.1.1 · pp. 15–16

5.1.2 Pre-Trained Foundation Models in Spatio-Temporal Graphs.

  • UMSST uses holistic graph representations and multi-subgraph contrastive pretraining, while UniSTD unifies multiple spatio-temporal domains with a Transformer and rank-adaptive mixture of experts.The survey characterizes pretrained models as transferring learned representations to improve STG learning and generalization across tasks and domains.Source: §5.1.2 · p. 16
  • Pangu-Weather, trained on decades of reanalysis data, achieves fast, accurate medium-range global weather forecasts; FourCastNet-3 [16] combines a sphere-aware architecture and probabilistic ensembles for calibrated long-range forecasts and substantial computational-efficiency gains.Source: §5.1.2 · p. 16
  • TrafficBERT predicts overall traffic flow and pretrains across roads to improve generalization, unlike conventional approaches trained separately for each road.It uses a bidirectional Transformer structure inspired by BERT [39].Source: §5.1.2 · p. 16
  • The Transportation Foundation Model (TFM) [223] combines traffic simulation with graph structures and dynamic graph generation to model participant interactions and address structural complexity and accuracy challenges.The survey describes this as a data-driven, model-free simulation approach.Source: §5.1.2 · pp. 16–17

with real-world data. … graph dynamics with continuous temporal changes.

The section describes temporal knowledge graphs as timestamped facts whose evolving structure supports forecasting and completion, and surveys video models that combine visual and textual information for understanding temporal scenes.

5.2 Temporal Knowledge Graphs

  • TKGs extend knowledge-graph triples (s, p, o) to timestamped facts (s, p, o, t), capturing temporal dependencies alongside structure to explain how entities generate facts over time.Knowledge graphs represent facts as subject, predicate, and object triples; TKGs add a timestamp as knowledge evolves.Source: §5.2 · p. 17
  • Chain-of-History reasoning incrementally adds higher-order historical facts to LLM prompts, providing compact temporal context for event forecasting.This supplies historical context for event prediction.Source: §5.2 · p. 17
  • LLM-DA adapts temporal logic rules as events arrive for interpretable, up-to-date reasoning without repeated fine-tuning, while LATE combines structural, temporal, and textual signals through multi-view prompting.LATE is described as improving factual consistency across diverse temporal settings.Source: §5.2 · p. 17
  • G2S trains on anonymized temporal structures before adding scenario-specific entities and relations, while AnRe retrieves and replays similar historical event chains without training [9].Source: §5.2 · p. 17

can exploit both semantic understanding and structural evolution for forecasting.

  • Luo et al. formulate temporal link prediction as event generation along historical chains with parameter-efficient fine-tuning, while TeRDy decomposes relation dynamics into long-term trends and short-term fluctuations [136].The Chain-of-History extension supports interpolation and extrapolation of temporal facts.Source: can exploit both semantic… · p. 17

components and injecting them into relation–timestamp embeddings, thereby providing a strong, general-purpose

  • A structure-aware approach integrates LLMs with graph representations to handle emerging entities and relations, using textual semantics and temporal graph context to infer missing links in evolving TKGs.Source: components and injecting them into… · p. 17
  • LLM-based TKG methods progress from historical prompting and retrieval to rule adaptation and structure-aware completion, but remain constrained by incomplete facts, sparse timestamps, and alignment of symbolic information.Source: components and injecting them into… · p. 17

5.3 Videos

  • Video is typically represented as a sequence of images or frames that collectively convey motion and temporal changes; conventional approaches for video understanding primarily involve 2D CNNs and 3D CNNs.2D CNNs process each video frame separately and aggregate along the time axis at the top of the network; 3D CNNs learn spatio-temporal video representations via 3D convolutions, aggregating spatial and temporal features.Source: §5.3 · p. 18

The emergence of large multimodal models marks a shift in video under- … relevant information, and maintaining reasoning capability across extended temporal horizons.

The video-LLM discussion traces a shift from multimodal video interpretation toward long-context reasoning, generative temporal tasks, and domain-tailored systems.

The emergence of large multimodal models marks a shift in video under-

  • Video-ChatGPT supports video conversation with a video-adapted visual encoder and LLMs, while Video-LLaMA uses Q-Former to integrate visual and audio information.Video-LLaMA aligns encoder outputs with the LLM embedding space to support understanding of both visual and auditory content in videos.Source: The emergence of large multimodal… · p. 18
  • Video-LLaMA uses video/image-caption pairs and quality visual-instruction data for alignment, while Video-LLaMA-3 [275] advances multimodal alignment and temporal reasoning.The latter enables more coherent video understanding.Source: The emergence of large multimodal… · p. 18

long-form and dynamic video content.

  • CaKE-LM uses LLM world knowledge for zero-shot video QA, MovieChat handles videos exceeding 10,000 frames with sparse memory, and MA-LMM [60] retains long-term context.MovieChat replaces dense token processing with selective memory representations; MA-LMM uses a memory-augmented mechanism.Source: long-form and dynamic video content · p. 18
  • Apollo studies scaling and unified pretraining for video-task generalization, video-SALMONN-o1 combines cognitive reasoning with audio-visual fusion, and ProgGen [197] models video prediction and dynamic events.video-SALMONN-o1 targets video question answering and inference, while ProgGen applies programmatic reasoning.Source: long-form and dynamic video content · pp. 18–19

dataset, aided by ChatGPT for conversation design.

choosing query-relevant frames during inference, thus boosting performance without additional model training.

relevant information, and maintaining reasoning capability across extended temporal horizons.

  • TRIVIA [164] uses automatic captioning as weak supervision for traffic knowledge, while MatchVision supports soccer event recognition, commentary generation, and foul detection.MatchVision is a large-scale foundation model trained on SoccerReplay-1988.Source: relevant information, and maintaining… · p. 19

video understanding. … spatio-temporal data, which are listed in Tab. 3.

The survey describes video foundation models, shared adaptation challenges and methods across time-series and spatio-temporal data, and resources for applications. It highlights how techniques and trade-offs in one area can inform the other.

5.3.2 Pre-Trained Foundation Models for Video Data.

  • Video pre-training uses contrastive or generative strategies [61], with models such as OmniVL unifying image-language and video-language modeling across visual-only, alignment, and multimodal tasks.mPLUG-2 shares universal modules while separating modality-specific ones to support flexible unimodal and cross-modal tasks across text, image, and video.Source: §5.3.2 · p. 19
  • Video models also pursue modality interaction and scale: mPLUG-Owl2 strengthens text-image-video collaboration, while InternVideo combines masked video modeling, contrastive learning, and next-token prediction.These approaches target cross-modal alignment and reasoning, long-form video comprehension, and multimodal generalization; MSTA [25] adds lightweight adapters and spatio-temporal consistency constraints for efficient transfer.Source: §5.3.2 · pp. 19–20

5.4 Methodological Convergence and Cross-Domain Insights

  • LM4TS and LM4STD both adapt large models to preserve temporal order, context, and domain semantics, while converting observations into tokens suited to each data type.LM4TS uses patches, symbols, summaries, or prompt-aligned embeddings [4] [51]; LM4STD adds topology, entity relations, trajectories, or video frames for STGs, TKGs, and videos [121].Source: §5.4 · p. 20
  • Both fields manage long-range dependencies within limited context budgets, using patching, decomposition, or lightweight adaptation for LM4TS and summarization, pooling, memory, or frame selection for LM4STD.The latter methods address expansion over nodes, entities, regions, or frames.Source: §5.4 · p. 20
  • Cross-domain transfer works both ways: spatio-temporal pooling, retrieval, video memory, and frame selection can inspire structured representations and compression for time series, while LM4TS prompt- and adapter-based transfer can reduce costs.LM4TS mainly uses prompts, calendar features, domain descriptions, or historical retrieval, whereas LM4STD can additionally exploit road networks and entity relations.Source: §5.4 · p. 20

6 Resources and Applications

  • The survey summarizes datasets, models, and tools across applications involving time-series and spatio-temporal data.This section frames shared application resources as a survey contribution.Source: §6 · p. 20

6.1 Traffic Application

  • Traffic resources include speed and flow datasets, 10,080 annotated SUTD-TrafficQA videos with 62,535 question-answer pairs, and tools such as the SUMO simulator.METR-LA covers LA County loop-detector speeds; PEMS-BAY and PEMS04 provide California traffic speed and flow data, while SUMO is open-source.Source: §6.1 · p. 20

real-world human mobility data. … documentation, in the field of neural event sequence modeling.

The survey catalogs datasets, models, and tools across temporal-data applications, from healthcare and weather to video, event sequences, and other domains. The supplied passages describe representative resources and their uses.

real-world human mobility data.

  • Table 3 summarizes common dataset resources across different applications.The supplied excerpt does not include the table’s entries.Source: Table 3 · p. 21

6.2 Healthcare Application

  • Healthcare resources include PTB-XL, NYU datasets, and the UF Health clinical corpus.PTB-XL contains 21,837 10-second, 12-lead ECGs from 18,885 patients; NYU datasets include ten years of unlabelled inpatient notes and task-specific fine-tuning datasets.Source: §6.2 · p. 21

6.3 Weather Application

  • Weather resources include NASA datasets [28], CMIP6, and forecasting models Pangu-Weather and GraphCast.The NASA datasets track hourly changes in 12 weather parameters, while CMIP6 evaluates global climate models; the passage describes Pangu-Weather as rapid and precise and notes GraphCast incorporates GNNs.Source: §6.3 · p. 21

6.4 Finance Application

  • Finance datasets support corporate-event detection and text-based stock prediction, stock-price modeling, employment analysis, and tweet-linked stock prediction.Examples include EDT, NASDAQ-100, Finance (Employment) [49], and StockNet; WeaverBird [261] and FinGPT [262] are cited as financial LLMs.Source: §6.4 · p. 22

6.5 Video Application

  • Video resources span question answering, captioning, and other vision-language tasks, with datasets including TGIF-QA, WebVid [10], and MSR-VTT.TGIF-QA has 165K GIF-based QA pairs, WebVid contains 10 million annotated video clips, and MSR-VTT includes 10,000 clips across 20 categories.Source: §6.5 · p. 22

6.6 Event Prediction Application

  • Event-sequence datasets cover e-commerce reviews, retweets, question-answer activity, and taxi pickups, while Tick and EasyTPP provide modeling resources.Tick focuses on statistical learning for classical TPPs; EasyTPP is described as an open-source repository with data, models, and evaluation programs.Source: §6.6 · p. 22

6.7 Other Applications

  • Other application benchmarks include ETT for transformer-temperature forecasting, Alibaba Cluster Trace [33] for cloud workloads, TSSB for segmentation, and UCR [37] for classification.The passage says these datasets span temporal patterns, sampling frequencies, and settings, supporting evaluation of temporal-model generalization; tools include OpenSTL, Merlion, darts, and PyTorch Geometric Temporal.Source: §6.7 · p. 22

7 Outlook and Future Opportunities … 8 Conclusion

The survey identifies six future directions for more powerful, transparent, and reliable temporal-data models and concludes with a taxonomy-based synthesis of existing work, its strengths and limitations, and research opportunities.

7 Outlook and Future Opportunities

  • Six future directions target more powerful, transparent, and reliable large models for temporal data.Source: §7 · p. 23

7.1 Theoretical Analysis of Large Models

  • Theoretical work should explain when language representations, tokenization, and attention capture temporal patterns or fail under distribution shift, precision loss, and modality mismatch.It should characterize language–temporal sequence similarities and differences, representation-transfer conditions, and task-specific limits in forecasting and anomaly detection.Source: §7.1 · p. 23

7.2 Development of Multimodal Models

  • Multimodal models should jointly represent temporal dynamics and modality-specific information, align signals at different temporal resolutions, and account for edge-cloud deployment, latency, privacy, and robustness.Source: §7.2 · p. 23

7.3 Continuous Learning and Adaptation

  • Large temporal models should adapt to non-stationary patterns without catastrophic forgetting; source-free adapter mapping for lithium-ion battery health suggests privacy-preserving cross-domain transfer.Online learning, concept-drift adaptation, and evolving-pattern modeling remain under-explored in large temporal models.Source: §7.3 · p. 23

7.4 Interpretability and Explainability

  • Interpretability research should explain what large models learn from temporal data and how they predict, providing rationales and enabling counterfactual, causal, root-cause, and scenario analyses.Intervention planning is also a proposed use of explicit temporal reasoning.Source: §7.4 · pp. 23–24

7.5 Privacy and Adversarial Attacks on Large Models

  • Because models trained or fine-tuned on sensitive temporal data may memorize private details, differential privacy and federated learning are proposed to protect data while enabling analysis.Source: §7.5 · p. 24

7.6 LLM Agents and Decision Making Process in Time Series Systems

  • LLM agents face unsafe or overconfident decisions under non-stationarity, regime shifts, rare events, and noisy, biased, or adversarial trajectories, motivating robust, uncertainty-aware policies.Proposed safeguards include out-of-distribution detection and abstention, risk-sensitive policies, safe exploration, and human- or rule-in-the-loop oversight.Source: §7.6 · p. 24

8 Conclusion

  • The survey offers an extensive, up-to-date review and a novel taxonomy of large models for time-series and spatio-temporal analysis, summarizing techniques, strengths, limitations, and research avenues.Source: §8 · p. 24

Related papers

Loading 2310.10196v3…