mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale

Xiaona Zhou, Constantin Brif, Ismini Lourentzou

🎉 Accepted to TMLR 2026

📄 Paper · 📝 arXiv · 🌐 Project Page · 💻 Code

---

Overview

mTSBench is the largest and most diverse benchmark for multivariate time series anomaly detection (MTS-AD) and model selection. It contains 344 labeled multivariate time series from 19 datasets across 12 application domains, and evaluates 24 anomaly detectors and 3 unsupervised model selection methods under a unified evaluation suite.

No single anomaly detector performs consistently well across datasets, and current unsupervised model selection methods remain far from optimal.

---

Key Idea

Multivariate time series anomaly detection is challenging because of high-dimensional dependencies, cross-correlations between time-dependent variables, and the scarcity of labeled anomalies. Because no single detector dominates across datasets, choosing the right detector for a new time series is a problem in its own right.

mTSBench is the first benchmark to integrate unsupervised model selection methods and evaluate them alongside anomaly detectors under consistent settings, with a unified evaluation suite of point-based and ranking-based metrics.

---

Highlights

• Largest MTS-AD benchmark: 344 labeled multivariate time series from 19 datasets across 12 application domains, including healthcare, cybersecurity, industrial monitoring, spacecraft telemetry, and IT infrastructure.

• 24 anomaly detectors: Spanning reconstruction-, prediction-, statistics-, and LLM-based approaches, including the only two publicly available LLM-based methods for multivariate time series.

• 3 unsupervised model selection methods: MetaOD, FMMS, and Orthus.

• Unified evaluation suite: Point-based and ranking-based metrics for anomaly detection and model selection, including VUS-PR, VUS-ROC, AUC-PR, AUC-ROC, AUC-PTRT, and Precision@k, Recall@k, and NDCG@k.

• Key finding: Even the strongest selectors remain far below the near-optimal baseline, and a single detector, PCA, surpasses all three selectors on 9 of 13 metrics.

---

Dataset Overview

Dataset	Config	Domain	#TS	#Dim	Length	#AnomPts	#AnomSeqs	License
CIC-IDS-2017	cicids	Cybersecurity	5	73	> 100K	0–8656	0–2546	Citation Required
CalIt2	CalIt2	Smart Building	1	3	> 5K	0	21	CC BY 4.0
CreditCard	creditcard	Finance / Fraud Detection	1	30	> 100K	219	10	Citation Required
Daphnet	Daphnet	Healthcare	26	10	> 50K	0	1–16	CC BY 4.0
Exathlon	Exathlon	IT Infrastructure	30	21	> 50K	0–4	0–6	Apache 2.0
GECCO	GECCO	Industrial Process	1	10	> 50K	0	37	Citation Required
GHL	GHL	Industrial Process	14	17	> 100K	0	1–4	Contact Authors
Genesis	Genesis	Industrial Process	1	19	> 5K	0	2	CC BY-NC-SA 4.0
GutenTAG	GutenTAG	Synthetic Benchmark	30	21	> 10K	0	1–3	MIT
MITDB	MITDB	Healthcare	47	3	> 500K	0	1–720	ODC-By v1.0
MSL	MSL	Spacecraft Telemetry	26	56	> 5K	0	1–3	BSD 3-Clause
Metro	metro	Transportation	1	6	> 10K	20	5	CC BY 4.0
OPPORTUNITY	OPPORTUNITY	Human Activity Recognition	13	33	> 25K	0	1	CC BY 4.0
Occupancy	room-occupancy	Smart Building	2	6	> 5K	1–3	9–13	CC BY 4.0
PSM	PSM	IT Infrastructure	1	27	> 50K	0	39	CC BY 4.0
SMAP	SMAP	Spacecraft Telemetry	48	26	> 5K	0	1–3	BSD 3-Clause
SMD	SMD	IT Infrastructure	18	39	> 10K	0	4–24	MIT
SVDB	SVDB	Healthcare	78	3	> 100K	0	2–678	ODC-By v1.0
SWAN-SF	swan	Astrophysics	1	39	> 50K	5233	1382	MIT

#AnomPts and #AnomSeqs are the numbers of anomalous points and anomalous sequences per time series. Licenses are those of the original datasets.

Each folder corresponds to one dataset and contains *_train.csv, *_val.csv, and *_test.csv files. Each CSV contains a timestamp column, dataset-specific feature columns, and a binary is_anomaly label. The *_val.csv files are held-out validation time series used to train the model selection methods. See data_summary.csv for per-file statistics.

---

Usage

This repository uses Git LFS for the CSV files:

git lfs install
git clone https://huggingface.co/datasets/PLAN-Lab/mTSBench

Or load one of the 19 configurations with 🤗 Datasets (one configuration per dataset, so files with different schemas are not combined):

from datasets import load_dataset

calit2 = load_dataset("PLAN-Lab/mTSBench", "CalIt2")

df_train = calit2["train"].to_pandas()
df_test = calit2["test"].to_pandas()

---

Resources

• Paper: mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale (TMLR 2026)
• arXiv: arxiv.org/abs/2506.21550
• Project Page: plan-lab.github.io/projects/mtsbench
• Code: github.com/PLAN-Lab/mTSBench

---

Citation

If you find mTSBench useful in your research, please consider citing our work:

@article{zhou2026mtsbench,
  title={mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale},
  author={Zhou, Xiaona and Brif, Constantin and Lourentzou, Ismini},
  journal={Transactions on Machine Learning Research},
  year={2026}
}