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

Subina Khanal; Seshu Tirupathi; Merim Dzaferagic; Marco Ruffini; Torben Bach Pedersen

cs.LGMar 17, 2026

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

Authors: Subina Khanal, Seshu Tirupathi, Merim Dzaferagic, Marco Ruffini, Torben Bach Pedersen

Organizations: Department of Computer Science Aalborg University · IBM Research, Ireland · Trinity College Dublin

Abstract

Time series foundation models (TSFMs) require diverse, real-world datasets to adapt across varying domains and temporal frequencies. However, current large-scale datasets predominantly focus on low-frequency time series with sampling intervals, i.e., time resolution, in the range of seconds to years, hindering their ability to capture the nuances of high-frequency time series data. To address this limitation, we introduce a novel dataset, msData, that captures millisecond-resolution wireless and traffic conditions from an operational 5G wireless deployment, expanding the scope of TSFMs to incorporate high-frequency data for pre-training. Further, the dataset introduces a new domain, namely, wireless networks, thus complementing existing more general domains like energy and finance. The dataset also provides use cases for short-term forecasting, with prediction horizons spanning from 1 millisecond (1 step) to 96 milliseconds (96 steps). To demonstrate the utility of the dataset, we benchmark traditional machine learning models, transformer-based deep learning models, and TSFMs on forecasting tasks using representative subsets of the data, including a static mobility pattern within YouTube traffic class and a train mobility pattern within Web Browsing traffic class. Across these data distributions, we demonstrate that most TSFM model configurations perform poorly in both zero-shot and fine-tuned settings. Our work underscores the importance of incorporating high-frequency datasets during pre-training and forecasting to enhance architectures, fine-tuning strategies, generalization, and robustness of TSFMs in real-world applications. Code is available at this repository: https://github.com/khanalsubina/msData. Data is available on: Hugging Face.

Figures & tables

Panel of Figure 1 — Figure 1: Comparison of timescales and dataset sizes for standard existing datasets used for pre-training (Table 14 in ( Aksu et al., 2024 ) ) as compared with the new benchmark. The red dot represents the new dataset that is introduced in this paper.

Panel of Figure 2 — Figure 2: Comparison of existing domains for pre-training (Table 14 in ( Aksu et al., 2024 ) ) with the new benchmark. The red bar represents the new dataset that is introduced in this paper.

Panel of Figure 3 — Figure 3: Comparison of prediction lengths of standard test data (Table 2 in ( Aksu et al., 2024 ) ) as compared with the new benchmark. The red bar represents the new dataset that is introduced in this paper.

Panel of Figure 4 — Figure 4: Target variable (Downlink Bitrate; mac_dl_brate): (a) STL decomposition, (b) Rolling mean and standard deviation, (c) Residual Q-Q, (d) Signal-to-Noise Ratio (dB).

Dataset	STL Decomposition
Dataset	Trend	Seasonality	Residuals
Network	Unstable, step-like shifts.	Weak short-term periodic patterns. Hidden by noise.	Sharp spikes. Bursts of noise. Lots of unpredictable variation.
ETTh1	Mostly steady with small rises and falls.	Small, regular repeating pattern.	Tiny random changes.
Electricity	Remains steady throughout.	Strong repeating pattern.	Occasional bursts of noise.
Weather	Almost flat but interrupted by sudden sharp spikes.	No seasonality.	Mostly small, but with rare sudden jumps.
Traffic	Slowly increasing trend over time.	Strong, regular repeating pattern.	Small random changes.

Table 1: Summary of STL Decomposition of all datasets.

Feature	Description
CQI	Channel Quality Indicator
MCS	Modulation and Coding Scheme
pkt ok	Number of packets sent
pkt nok	Number of packets dropped

Table 2: Features used in multivariate setting.

Table 3: Parameters used in model training.

	Univariate		Multivariate
Model	RMSE	MAE	RMSE	MAE
RF	$0.0344\pm 0.0001$	$0.0227\pm 0.0001$	$0.0342\pm 0.0001$	$0.0226\pm 0.0001$
XGB	$0.0354\pm 0.0001$	$0.0232\pm 0.0001$	$0.0354\pm 0.0001$	$0.0231\pm 0.0001$
ARF	$\textbf{0.0270}\pm\textbf{0.0002}$	$0.0189\pm 0.0001$	$\textbf{0.0175}\pm\textbf{0.0007}$	$\textbf{0.0130}\pm\textbf{0.0005}$
Naive	$0.0418\pm 0.0000$	$0.0240\pm 0.0000$	$0.0418\pm 0.0000$	$0.0240\pm 0.0000$
OLR	$0.0551\pm 0.0000$	$0.0308\pm 0.0000$	$0.0555\pm 0.0000$	$0.0310\pm 0.0000$

Table 4: Performance metrics of benchmarked models.

Figure 5: Actual v.s. Predicted bitrate values in an univariate setting.

Figure 6: Actual v.s. Predicted bitrate values in a multivariate setting.

Fine-tuning Strategy	RMSE	MAE
Head-only fine-tuning	0.0413	0.0270
Adapter-based fine-tuning	0.0522	0.0334

Table 5: Performance metrics for different fine-tuning strategies for TTM.

Temporal Resolution	Prediction Horizon	ARF		TTM Zero-shot		TTM Fine-tuning		Chronos Zero-shot
Temporal Resolution	Prediction Horizon	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE
100 ms	96	0.0457	0.0262	0.0765	0.0434	0.0743	0.0421	0.0622	0.0338
200 ms	48	0.0471	0.0267	0.0870	0.0499	0.0880	0.0496	0.0740	0.0389
500 ms	20	0.0398	0.0218	0.0855	0.0490	0.0894	0.0542	0.0711	0.0372
1000 ms	10	0.0297	0.0176	0.0856	0.0500	0.0856	0.0500	0.0580	0.0326
2000 ms	5	0.0289	0.0169	0.0880	0.0527	0.0915	0.0584	0.0671	0.0354

Table 6: Performance metrics of benchmarked models with the increasing temporal resolution.

	Univariate		Multivariate
Model	RMSE	MAE	RMSE	MAE
RF	0.0751	0.0542	0.0740	0.0536
XGB	0.0774	0.0536	0.0758	0.0533
ARF	0.0605	0.0366	0.0459	0.0230
Naive	0.0727	0.0476	0.0727	0.0476
OLR	0.0819	0.0569	0.0817	0.0567

Table 7: Performance metrics of benchmarked models on the filtered subset defined by the train mobility pattern within the Web Browsing traffic class.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 7 — Figure 7: STL decomposition of time series: (a) Network, (b) ETTh1, (c) Electricity, (d) Weather, (e) Traffic.

Panel of Figure 8 — Figure 8: Rolling mean and standard deviation of time series: (a) Network, (b) ETTh1, (c) Electricity, (d) Weather, (e) Traffic.

Panel of Figure 9 — Figure 9: Autocorrelation of time series: (a) Network, (b) ETTh1, (c) Electricity, (d) Weather, (e) Traffic.

Panel of Figure 10 — Figure 10: Autocorrelation of time series: (a) Network, (b) ETTh1, (c) Electricity, (d) Weather, (e) Traffic.

Dataset	No. of features
Network	47
ETTh1	8
Electricity	322
Weather	22
Traffic	863

Appendix

Table 8: Number of features in different datasets.

Feature	Description
CQI	Channel Quality Indicator
MCS	Modulation and Coding Scheme
pkt ok	Number of packets sent
pkt nok	Number of packets dropped
id_ue	Number of ue’s connected in the BS
pusch_sinr	Noise ratio interference in the Physical Uplink Shared Channel

Appendix

Table 9: Features used in multivariate setting.

Model	Multivariate
Model	RMSE	MAE
XGB	0.0347	0.0234
ARF	0.0273	0.0155
Naive	0.0417	0.0239
TTM (Zero-shot)	0.0359	0.0230
TTM (Fine-tuning)	0.0358	0.0228

Appendix

Table 10: Performance of benchmarked models using ten features.

Figure 11: Actual v.s. Predicted bitrate values using ten features.

	Univariate		Multivariate
Model	RMSE	MAE	RMSE	MAE
XGB	0.1440	0.1087	0.1440	0.1087
ARF	0.1728	0.1125	0.0968	0.0634
Naive	0.1309	0.0932	0.1309	0.0932
TTM (Zero-shot)	0.1279	0.0922	0.1279	0.0922

Appendix

Table 11: Performance metrics of benchmarked models the filtered subset defined by the train mobility pattern within the Dos-Hulk-C traffic class.

Panel of Figure 12 — Figure 12: Actual v.s. Predicted bitrate values for the filtered subset: (a) Univariate, (b) Multivariate.

Mobility Pattern	Multivariate
	Traffic Class : YouTube				Traffic Class : Portscan
	Model
	ARF		TTM (Zero-shot)		ARF		TTM (Zero-shot)
	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE
Static	0.0166	0.0123	0.0359	0.0230	0.0858	0.0585	0.1192	0.0904
Pedestrian	0.0457	0.0262	0.0765	0.0434	0.0819	0.0555	0.1127	0.0860

Appendix

Table 12: Performance metrics of benchmarked models across additional filtered subsets with different mobility patterns and traffic classes.

Learning Rate (LR)	RMSE	MAE	Fine-tune Percent	RMSE	MAE	No. of Epochs	RMSE	MAE
0.01	0.0390	0.0249	10	0.0358	0.0227	50	0.0359	0.0227
0.001	0.0387	0.0247	15	0.0365	0.0226	80	0.0359	0.0227
0.00001	0.0359	0.0227	25	0.0366	0.0227	100	0.0359	0.0227
0.000001	0.0359	0.0229	30	0.0367	0.0226

Appendix

Table 13: Hyper-parameter tuning of the TTM model in the multivariate setting.

Context Length	RMSE	MAE	Batch Size	RMSE	MAE
15	0.0350	0.0231	16	0.0330	0.0227
25	0.0324	0.0217	32	0.0314	0.0218
35	0.0327	0.0221	128	0.0332	0.0221

Appendix

Table 14: Hyper-parameter tuning of Lag-Llama model.

Case	$d_{\text{model}}$	$d_{ff}$	Heads	Layers	RMSE	MAE
1	64	128	2	3	0.0321	0.0206
1	64	128	4	3	0.0319	0.0205
2	64	128	8	2	0.0320	0.0207
2	64	128	8	4	0.0323	0.0207
3	128	256	8	3	0.0324	0.0208
4	256	512			0.0324	0.0209

Appendix

Table 15: PatchTST hyper-parameter tuning configurations in the multivariate setting. Case 1 varies the number of attention heads, Case 2 varies the number of encoder layers, Cases 3-5 examine the effect of increasing model width, i.e.,

(d_{\text{model}},d_{ff})\in\{(128,256),(256,512),(512,1024)\}

Case	$d_{\text{model}}$	$d_{ff}$	Heads	Layers	RMSE	MAE
1	64	128	2	3	0.0326	0.0205
1	64	128	4	3	0.0319	0.0206
2	64	128	8	2	0.0321	0.0206
2	64	128	8	4	0.0324	0.0209
3	128	256	8	3	0.0321	0.0207
4	256	512			0.0318	0.0207

Appendix

Table 16: iTransformer hyper-parameter tuning configurations in the multivariate setting. Case 1 varies the number of attention heads, Case 2 varies the number of encoder layers, Cases 3-5 examine the effect of increasing model width, i.e.,

(d_{\text{model}},d_{ff})\in\{(128,256),(256,512),(512,1024)\}

n_estimators	max_depth	LR = 0.1	LR = 0.3	LR = 0.5
50	2	0.0342 / 0.0232	0.0346 / 0.0231	0.0350 / 0.0233
	4	0.0339 / 0.0225	0.0347 / 0.0228	0.0356 / 0.0233
	6	0.0341 / 0.0225	0.0352 / 0.0230	0.0366 / 0.0238
	8	0.0341 / 0.0225	0.0356 / 0.0234	0.0381 / 0.0248
100	2	0.0341 / 0.0228	0.0346 / 0.0230	0.0350 / 0.0232
	4	0.0341 / 0.0225	0.0350 / 0.0229	0.0374 / 0.0244

Appendix

Table 17: Hyper-parameter optimization results for XGBoost in the multivariate setting. Each cell reports RMSE / MAE.

n_estimators	max_depth	max_features = sqrt	max_features = 1
50	None	0.0342 / 0.0226	0.0342 / 0.0228
	2	0.0355 / 0.0259	0.0403 / 0.0309
	4	0.0347 / 0.0244	0.0355 / 0.0257
100	None	0.0339 / 0.0223	0.0342 / 0.0226
	2	0.0355 / 0.0259	0.0408 / 0.0314
	4	0.0346 / 0.0244	0.0358 / 0.0261

Appendix

Table 18: Hyperparameter optimization results for Random Forest in the multivariate setting. Each cell reports RMSE / MAE.

Model	Prediction Horizon = 192		Prediction Horizon = 336		Prediction Horizon = 720
Model	RMSE	MAE	RMSE	MAE	RMSE	MAE
XGB	0.0399	0.0268	0.0445	0.0313	0.0579	0.0436
ARF	0.0204	0.0134	0.0187	0.0140	0.0188	0.0138
Naive	0.0533	0.0296	0.0623	0.0366	0.0742	0.0506
OLR	0.0622	0.0353	0.0681	0.0398	0.0774	0.0488
PatchTST	0.0357	0.0228	0.0400	0.02584	0.0494	0.0337

Appendix

Table 19: Performance comparison of selected benchmarked models across different prediction horizons in the multivariate setting.

Wireless Local Area Networks

See all 18 papers

Oct 3, 2026

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

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.

Martin Maritsch, Timo Stoffregen, Thomas Kaar +36

Aionic Labs, Switzerland · Agentic Systems Lab, ETH Zurich, Switzerland · Stanford University, USA +17

Time-Series Foundation Models

Ask PDF

Apr 12, 2026

WaveMoE: A Wavelet-Enhanced Mixture-of-Experts Foundation Model for Time Series Forecasting

Time series foundation models (TSFMs) have recently achieved remarkable success in universal forecasting by leveraging large-scale pretraining on diverse time series data. Complementing this progress, incorporating frequency-domain information yields promising performance in enhancing the modeling of complex temporal patterns, such as periodicity and localized high-frequency dynamics, which are prevalent in real-world time series. To advance this direction, we propose a new perspective that integrates explicit frequency-domain representations into scalable foundation models, and introduce WaveMoE, a wavelet-enhanced mixture-of-experts foundation model for time series forecasting. WaveMoE adopts a dual-path architecture that jointly processes time series tokens and wavelet tokens aligned along a unified temporal axis, and coordinates them through a shared expert routing mechanism that enables consistent expert specialization while efficiently scaling model capacity. Preliminary experimental results on 16 diverse benchmark datasets indicate that WaveMoE has the potential to further improve forecasting performance by incorporating wavelet-domain corpora.

Shunyu Wu, Jiawei Huang, Weibin Feng +6

Sun Yat-sen University · Xiaomi Corporation · National University of Singapore

Time Series Forecasting Time-Series Foundation Models Frequency-Domain Time Series Modeling Mixture-of-Experts Models Frequency-Domain Feature Learning

Ask PDF

Mar 5, 2026

Timer-S1: A Billion-Scale Time Series Foundation Model with Serial Scaling

We introduce Timer-S1, a strong Mixture-of-Experts (MoE) time series foundation model with 8.3B total parameters, 0.75B activated parameters for each token, and a context length of 11.5K. To overcome the scalability bottleneck in existing pre-trained time series foundation models, we perform Serial Scaling in three dimensions: model architecture, dataset, and training pipeline. Timer-S1 integrates sparse TimeMoE blocks and generic TimeSTP blocks for Serial-Token Prediction (STP), a generic training objective that adheres to the serial nature of forecasting. The proposed paradigm introduces serial computations to improve long-term predictions while avoiding costly rolling-style inference and pronounced error accumulation in the standard next-token prediction. Pursuing a high-quality and unbiased training dataset, we curate TimeBench, a corpus with one trillion time points, and apply meticulous data augmentation to mitigate predictive bias. We further pioneer a post-training stage, including continued pre-training and long-context extension, to enhance short-term and long-context performance. Evaluated on the large-scale GIFT-Eval leaderboard, Timer-S1 achieves state-of-the-art forecasting performance, attaining the best MASE and CRPS scores as a pre-trained model. Timer-S1 is released to facilitate further research.

Yong Liu, Xingjian Su, Shiyu Wang +7

Tsinghua University · ByteDance

Time Series Forecasting Time-Series Foundation Models Long-Term Time Series Forecasting Sequence Modeling Mixture-of-Experts Models

Ask PDF

Abstract

Wireless Local Area Networks

Explore similar work

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

WaveMoE: A Wavelet-Enhanced Mixture-of-Experts Foundation Model for Time Series Forecasting

Timer-S1: A Billion-Scale Time Series Foundation Model with Serial Scaling