TSFMI-Synthetic

Synthetic time-series datasets with mathematically exact ground-truth labels for the TSFMI baseline-controlled probing protocol.

Companion data for the NeurIPS 2026 Evaluations & Datasets Track submission "TSFMI: A Baseline-Controlled Evaluation Protocol for Time-Series Foundation Model Representations." The code (anonymous) lives at https://anonymous.4open.science/r/TSFMI.

Why this dataset exists

Probing time-series foundation models (TSFMs) is hard because high probe accuracy may reflect probe capacity rather than encoded knowledge. TSFMI addresses this with explicit non-model controls (hand-crafted features, raw signal, random projection, ROCKET) evaluated under the same canonical 60/20/20 split, the same sklearn estimator, the same 5 seeds, and the same bootstrap CI as the model probe. The synthetic datasets in this repository provide six standard temporal properties plus five hard variants under which TSFM representation claims can be evaluated against the controls.

Configurations (11 tasks)

Config	Type	Classes / Range	Description
trend	classification	3 (up / down / flat)	linear slope ± noise
seasonality	regression	period ∈ {8, 16, 32, 64}	sinusoid + noise; label = period
frequency	classification	8 frequency bins	discrete sinusoidal frequency bands
stationarity	classification	2 (stationary / non-stationary)	Gaussian noise vs. random walk
anomaly	classification	2 (normal / anomaly)	single ±5σ point spike — kurtosis is a sufficient statistic
change_point	classification	2 (none / has-CP)	mean+variance shift at midpoint
*_hard (5)	classification	varies	structured background + subtle anomalies / weak slopes / overlapping harmonics / variance drift / mild CP

Each config has 1000 sequences of length 512, partitioned 60/20/20 (train/val/test = 600/200/200). All data is generated deterministically with seed=42 (data) and split_seed=0 (partition), matching the canonical artefacts in the paper.

Quick start

from datasets import load_dataset

ds = load_dataset("EvalData/TSFMI", "anomaly")
print(ds)
# DatasetDict({
#   train: 600, validation: 200, test: 200
# })
print(ds["train"][0]["sequence"][:8], ds["train"][0]["label"])

Each row contains:

• sequence: list[float] of length 512 (the time-series window)
• label: int (classification) or float (seasonality regression)
• seed: int (data-generation seed; always 42 here)
• split_seed: int (split seed used to partition this row; 0 for the canonical artefacts)

Reproduction recipe

The same 60/20/20 + 5-seed + bootstrap protocol used in the paper is implemented in the GitHub repository:

curl -L -o TSFMI.zip "https://anonymous.4open.science/api/repo/TSFMI/zip"
unzip TSFMI.zip && cd TSFMI
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt && pip install -e ".[dev]"

# CPU smoke test (<10 min) — reproduces the headline HC anomaly = 0.858 cell
make smoke

# Full canonical pipeline (~48 A100-hours)
make extract-representations
make reproduce-all

Headline result reproducible from this dataset

Under the canonical 60/20/20 protocol with sklearn LogisticRegression and a bootstrap 95% CI over 5 seeds:

• An 8-D hand-crafted feature vector reaches 0.858 test accuracy on the canonical anomaly configuration.
• A single kurtosis feature alone reaches 0.859.
• A single max-absolute-magnitude feature alone reaches 0.907.
• The best of seven pre-trained TSFMs reaches 0.753 (TimesFM); all other TSFMs (MOMENT, Chronos, PatchTST, GPT4TS, Timer, Moirai) score 0.50–0.73.

This inversion is what motivates the baseline-controlled discipline of TSFMI and is documented in §3.3 and Appendix A.8 of the paper.

License

MIT. The synthetic data is fully procedurally generated; there is no human-derived or scraped content. Real-world datasets used elsewhere by the TSFMI evaluation pipeline (ETTh1, Weather, Electricity, Traffic, Exchange Rate, UCR) are not redistributed here and remain under their respective original licenses; see the GitHub LICENSE file for details.

Citation

@misc{tsfmi2026,
  title = {{TSFMI}: A Baseline-Controlled Evaluation Protocol for Time-Series Foundation Model Representations},
  author = {Anonymous Authors},
  howpublished = {Anonymous submission to the NeurIPS 2026 Evaluations \& Datasets Track},
  year = {2026},
  url = {https://anonymous.4open.science/r/TSFMI}
}

Responsible AI notes

• Data collection: synthetic procedural generation, no human subjects, no scraping.
• Limitations: each generator instantiates one statistically simple test bed; the canonical anomaly task is kurtosis-trivial by design (realistic_anomaly variant in the GitHub repo bounds this).
• Biases: none — purely synthetic with no demographic content.
• Personal/sensitive info: none.
• Use cases: probing TSFM internal representations, calibrating new probing protocols against simple non-model baselines.
• Misuse cases: not intended as a production anomaly detector; do not interpret the canonical anomaly task as evidence that any model "detects anomalies" in any operational sense.
• Synthetic data indicator: 100% synthetic.

The full Croissant 1.0 + RAI metadata is provided as CROISSANT.json in this repository.