TimeNet: An Extensible Unified Data Infrastructure for Next-Generation Temporal Foundation Models

Martin Maritsch; Timo Stoffregen; Thomas Kaar; Behsad Riemer; Maxwell A. Xu; Max Rosenblattl; Juncheng Liu; Nicolas Zumarraga

cs.AIOct 3, 2026

TimeNet: An Extensible Unified Data Infrastructure for Next-Generation Temporal Foundation Models

Authors: Martin Maritsch, Timo Stoffregen, Thomas Kaar, Behsad Riemer, Maxwell A. Xu, Max Rosenblattl, Juncheng Liu, Nicolas Zumarraga, Yu Yvonne Wu, Denys Herasymuk, Sparsh Rastogi, Hyungjun Yoon, Bosong Huang, Arvind Pillai, Dmytro Lopushanskyy, Tony Chen, Robin Deuber, Yichen Liu, Shvat Messica, Dan Li, Jian Lou, Yuwei Zhang, Jaeho Kim, Renée Rosillo Garcia, Fan Wu, Robert Müller, Elgar Fleisch, Flora D. Salim, Dimitris Spathis, Yuzhe Yang, Aaqib Saeed, Daniel McDuff, Ming Jin, Markus Kreft, Kevin O'Sullivan, Robert Jakob, Azul Garza, Paul Schmiedmayer, Patrick Langer

Organizations: Aionic Labs, Switzerland · Agentic Systems Lab, ETH Zurich, Switzerland · Stanford University, USA · Technical University of Munich, Germany · Google · National University of Singapore, Singapore · Dartmouth College, USA · New York University, USA · UNSW Sydney, Australia · KAIST, South Korea · Griffith University, Australia · Columbia University, USA · Harvard University, USA · Sun Yat-sen University, China · University of Cambridge, UK · University of Virginia, USA · TimeCopilot, USA · Centre for Digital Health Interventions, St. Gallen, Switzerland · University of California, Los Angeles, USA · Eindhoven University of Technology, The Netherlands

Abstract

Temporal Foundation Models (TFMs) aim to generalize across domains, datasets, and tasks. Yet, their development remains constrained by fragmented, task-specific data formats, annotations, and processing pipelines. We introduce TimeNet, an open-source data standard and scalable infrastructure that decouples temporal data from task definitions and represents signals, metadata, annotations, and supervision in a shared, extensible data model. TimeNet supports multimodal signals with regular, irregular, or ordinal time axes and expresses different task families (including classification, forecasting, temporal localization, question answering, generation, and editing) as reusable views over the same recordings. This shared representation enables heterogeneous time-series datasets to be combined for large-scale model training across domains, modalities, and tasks. We demonstrate TimeNet by transcoding datasets with 1.5M task instances spanning diverse domains, modalities, temporal scales, and forms of supervision, while retaining practical I/O performance relative to native formats. TimeNet enables an existing TFN training pipeline to support joint training on a configurable number of heterogeneous datasets through configuration changes alone. We show this capability by training TFM across multiple datasets and tasks, obtaining a 14% F1 score improvement compared with models trained on individual datasets. These results show that TimeNet provides the data and systems foundation needed to move beyond task- and dataset-specific TFMs toward models that can learn jointly across heterogeneous domains, modalities, temporal scales, and forms of supervision from a common data model.

Figures & tables

Figure 1: An Overview of TimeNet. We introduce a unified data standard, open software infrastructure and an extensive framework for training and evaluation. Multiple tasks can be defined over the same underlying data and across datasets, which can be combined seamlessly during training.

Figure 2: TimeNet processing pipeline. TimeNet allows for flexible reuse and recombination of heterogeneous data. Once ingested into TimeNet using a specialized connector, a dataset can be used for any model that operates over applicable Tasks.

Figure 3: TimeNet data model. TimeNet represents each dataset as Records with recursive hierarchies of Sources and Signals. A Source describes where data originated and can contain child Sources. Signals describe value streams and link them to stored values and metadata. Tasks reference one or more Records, and each Record can participate in multiple Tasks. Datasets, Tasks, Records, Sources, and Signals can all have annotations. DuckDB stores the hierarchy, annotations, and many-to-many relationships. Underlying Signal values are stored separately in efficient, sharded binary files.

Task family & construction	Signature $\phi$	Defining relation	Derived tasks	TimeNet realization
Attribute inference $f(X,C)\to y$	$(\textsc{Value},\Delta_{\mathrm{in}},1)$	Infer a property of the recorded object or a span of it.	Classification, regression, state and severity estimation, captioning	classification , scalar prediction , answer
Temporal localization $f(X,q)\to\{s_{i}\}$ , $s_{i}\subseteq\mathcal{T}$	$(\textsc{Span},\Delta_{\mathrm{in}},1)$	Identify where or when a property or event occurs in the recorded extent.	Event and anomaly localization, change-point detection, segmentation	temporal localization
Future-directed inference $f(X_{\leq t},C)\to Y_{>t}$	$(\ast,\{\textsc{Fut},\textsc{Past}\},1)$	Infer a target whose support lies beyond the observed context.	Forecasting, event and future-state prediction, time-to-event, onset prediction	forecasting ; other types $+\textsc{Shift}$
Relational inference $f(X_{i},X_{j},C)\to r$	$(\ast,\ast,{\geq}2)$	Infer a relation between two or more temporal objects or spans.	Ordering, comparison, alignment, similarity, retrieval	correspondence ; any type $+\textsc{Pair}$
Temporal construction $f(C)\to X^{\prime}$	$(\textsc{Signal},\textsc{Null},1)$	Construct a new temporal object from conditioning alone.	Conditional and unconditional generation	generation
Temporal transformation $f(X,C)\to X^{\prime}$	$(\textsc{Signal},\Delta_{\mathrm{in}},1)$	Produce a temporal object by transforming or completing a recorded one.	Editing, denoising, reconstruction, imputation, virtual sensing	editing ; editing $+\textsc{Mask}$

Table 1: Compositional taxonomy of temporal tasks. Each family is a region of the signature space

\phi=(\kappa,\Delta,\alpha)

(Definitions 1 – 2 ); the six regions partition all well-posed signatures (Appendix B.3 ).

\ast

: any value;

\Delta_{\mathrm{in}}=\{\textsc{Obs},\textsc{Gap},\textsc{Chan}\}

, i.e., the target lies within the recorded extent, whether observed, masked, or in an unmeasured channel. TimeNet realizations in monospace ;

+\textsc{Shift}

etc. denote constructors (Definition 3 ). For future-directed inference, “other types” are classification , scalar prediction , and temporal localization .

Figure 4: TimeNet task taxonomy. TimeNet has five semantic task families, each containing built-in task types that can be further specialized or expanded. Tasks may be expressed in natural language.

Figure 5: Data collation in TimeNet. The structure of TimeNet allows for a common mapping from Tasks to prompts that can be used during training.

	Sleep-EDF		ECG-QA		HEARTS		ARFBench		Average
Training	F1	Acc.	F1	Acc.	F1	Acc.	F1	Acc.	F1	Acc.
OpenTSLM Single	37.49	77.34	31.19	67.19	15.03	35.71	32.94	33.33	29.16	53.39
OpenTSLM Pooled	41.80	80.47	34.21	73.44	19.70	38.72	37.41	32.82	33.28	56.36
SLIP Single	31.78	87.16	33.93	40.01	9.98	31.07	20.51	26.92	24.05	46.29
SLIP Pooled	42.27	87.87	30.00	31.32	24.78	45.61	20.38	29.75	29.36	48.64

Table 2: OpenTSLM and SLIP performance when trained on each dataset individually versus jointly.

Dataset	Representation	Consumer	Canonical item count [items]	Time to first item [s]	Full-read time [s]	Sequential throughput [items/s]	Decoded payload throughput [MB/s]	Transferred HTTP payload [GB]	HTTP request count [requests]	Peak resident set size [MiB]
HEARTS	Original	pandas	1,005	0.275	93.50	10.7	9.93	1.264	1,008	397.3
	Original	PyTorch	1,005	0.247	95.30	10.5	9.74	1.264	1,008	395.5
	TimeF	pandas	1,005	0.574	18.43	54.5	50.38	0.331	220	879.1
	TimeF	PyTorch	1,005	0.488	17.35	57.9	53.50	0.331	220	888.2
ARFBench	Original	pandas	750	0.479	64.70	11.6	4.42	0.227	1,010	860.5
	Original	PyTorch	750	0.366	59.51	12.6	4.70	0.227	1,010	866.7

Table 3: Results of I/O performance assessment across 6 datasets.

Contributing institutions. Numbers follow the affiliation list on the first page.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: TimeNet API: loading signals and defining spans, attaching annotations, and deriving tasks.

Dataset	Size	Released format	Domains	Task types	Overlap
Sleep-EDF Exp.	8.1 GB	EDF, EDF+ (int16), XLS	sleep, EEG, EOG, EMG, respiration	scalar prediction, temporal localization	SLIP
ECG-QA CoT	1.89 GB	WFDB (int16), CSV	cardiology, ECG	question answering	SLIP
HEARTS	1.32 GB	pickle (pandas DataFrame, float64)	metabolic, respiration, COVID cough	scalar prediction, forecasting, temporal localization	Capture-24
ARFBench	103 MB	Parquet (datetime64, float64, string, int64), CSV	infrastructure, networking, database, security		none
VerbalTS	800 MB	NPY (float64, UTF-32, int64), JSON	synthetic, energy, traffic, motion, weather	series generation	ETTm1, Weather
SLIP	6.14 GB	Parquet (nested float32, string)	health, web, nature, energy, IoT, transport, activity, stress	question answering	Sleep-EDF, PTB-XL, Capture-24

Appendix

Table 4: The datasets, as released. Task types name the tags of TimeF’s task registry.

	single object ( $\alpha=1$ )					relational
	Obs	Gap	Fut	Chan	Null	$\alpha\geq 2$
Value	[A] classification, scalar, answer	[A] value over masked region	[F] event prediction	[A] cross-channel inference	—	[R] comparison, ordering
Span	[L] localization, segmentation	[L] localization in masked region	[F] onset prediction	[L] cross-channel localization	—	[R] alignment
Signal	[T] editing, denoising	[T] imputation	[F] forecasting	[T] virtual sensing	[C] generation	[R] signal relation
Index	—	—	—	—	—	[R] retrieval

Appendix

Table 5: Signature grid for temporal tasks. Rows denote output kind

\kappa

and columns denote support relation

\Delta

for single-object tasks. The final column contains relational tasks with

\alpha\geq 2

. [A], [L], [F], [R], [C], and [T] denote attribute inference, temporal localization, future-directed inference, relational inference, temporal construction, and temporal transformation, respectively.

Type	Kind $\kappa$	Default $\mu$
$\textsc{Label}_{K}$ , $\textsc{MultiLabel}_{K}$	Value	accuracy, macro-F1
$\mathbb{R}^{d}$	Value	MAE, MAPE
Text	Value	exact match; model judge †
$\textsc{Spans}(\mathcal{T})$	Span	event-F1, IoU
$\textsc{Signal}(\mathcal{T}^{\prime},c)$	Signal	MASE, DTW, task-specific metrics
$\textsc{Index}_{N}$ , $\textsc{Ranking}_{N}$	Index	accuracy, MRR, Recall@ $k$

Appendix

Table 6: Representative output types. † Requires model or human judgment and is therefore not directly usable as a verifiable reward.

Time-Series Foundation Models

Time-series foundation models are pretrained sequence models designed to transfer across forecasting, classification, anomaly detection, and other temporal tasks.

See all 285 papers

TempusBench: An Evaluation Framework for Time-Series Forecasting
Apr 13, 2026Denizalp Goktas, Gerardo Riaño-Briceño, Omkar Tekawade +14Time Series Forecasting Forecasting Benchmarks
Simulacrum New York City, NY, USA
msData: A Millisecond-Resolution Network Dataset for Advancing Time Series Foundation Models
Mar 17, 2026Subina Khanal, Seshu Tirupathi, Merim Dzaferagic +2Wireless Local Area Networks Time-Series Foundation Models
Department of Computer Science Aalborg University · IBM Research, Ireland · Trinity College Dublin
Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding
May 18, 2026Paul Quinlan, Jeremy Levasseur, Qingguo Li +1Cross-Modal Learning Multivariate Time Series Forecasting
1InertialAI · Department of Electrical and Computer Engineering, Queen’s University · Department of Mechanical and Materials Engineering, Queen’s University

Apr 13, 2026J/K move · Enter open · S save

TempusBench: An Evaluation Framework for Time-Series Forecasting

Denizalp Goktas, Gerardo Riaño-Briceño, Omkar Tekawade, Alif Abdullah+13

Simulacrum New York City, NY, USA

Time Series Forecasting Forecasting Benchmarks Time-Series Foundation Models Time Series Modeling

Open paper AskPDF arXiv

Foundation models have transformed natural language processing and computer vision, and a rapidly growing literature on time-series foundation models (TSFMs) seeks to replicate this success in forecasting. While recent open-source models demonstrate the promise of TSFMs, the field lacks a comprehensive and community-accepted model evaluation framework. We see at least four major issues impeding progress on the development of such a framework. First, existing evaluation frameworks comprise benchmark forecasting tasks derived from often outdated datasets (e.g., M3), many of which lack clear metadata and overlap with the corpora used to pre-train TSFMs. Second, these frameworks evaluate models along a narrowly defined set of benchmark forecasting tasks, such as forecast horizon length or domain, but overlook core statistical properties such as non-stationarity and seasonality. Third, domain-specific models (e.g., XGBoost) are often compared unfairly, as existing frameworks do not enforce a systematic and consistent hyperparameter tuning convention for all models. Fourth, visualization tools for interpreting comparative performance are lacking. To address these issues, we introduce TempusBench, an open-source evaluation framework for TSFMs. TempusBench consists of 1) new datasets which are not included in existing TSFM pretraining corpora, 2) a set of novel benchmark tasks that go beyond existing ones, 3) a model evaluation pipeline with a standardized hyperparameter tuning protocol, and 4) a tensorboard-based visualization interface. We provide access to our code on GitHub: https://github.com/Smlcrm/TempusBench and maintain a live leaderboard at https://smlcrm.com/tempusbench.

Abstract

Time-Series Foundation Models

Explore similar work

TempusBench: An Evaluation Framework for Time-Series Forecasting

msData: A Millisecond-Resolution Network Dataset for Advancing Time Series Foundation Models

Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding