Dataset Card for TimesX

Dataset Summary

TimesX is a context-enriched, multimodal time-series forecasting benchmark introduced in "Rethinking Multimodal Time-Series Forecasting Evaluation" (Liu, Zhou, Sen, Prakash, Das; ICML 2026). Each example is one rolling-window "prediction origin" on a real-world numeric series, paired with the textual context a genuine forecaster would have at that point in time: a domain background, a scenario description, upcoming calendar effects, related-indicator statistics, and time-stamped news events retrieved by an automated data pipeline. The benchmark spans three domains -- commodity prices, currency exchange rates, and Google Search trends -- chosen so that no single numeric-only or text-only method dominates.

This mirror flattens the upstream benchmark's per-variable JSON files (one file per series, disagreeing on whether value is int- or float-valued) into a single typed table, so the whole benchmark loads with one load_dataset call instead of one call per variable file.

• Paper: https://huggingface.co/papers/2607.06973
• Homepage: https://haoxin1998.github.io/TimesX-project/
• Source repository: https://github.com/haoxin1998/TimesX-project

Licensing Information

The license for this dataset is not yet confirmed. The upstream repository has no LICENSE file, GitHub's own license detector reports none, and the paper's data-availability and impact statements do not address licensing terms for the released data (the CC BY 4.0 badge on the arXiv page covers the paper text, not the dataset). This mirror is published for research reference only, pending confirmation from the original authors -- do not redistribute further, or use in a product, until that is resolved. If you are one of the authors and can confirm terms, please open a discussion on this dataset or contact the uploader.

Dataset Structure

Data Instances

One row is one rolling-window prediction origin for one variable: a fixed-length history, a fixed-length future horizon to forecast, and the textual context available as of that origin's timestamp.

Data Fields

Field	Type	Description
domain	string	Top-level category, e.g. "CommodityPrice", "Currency", "SearchTrend".
subdomain	string	Second-level category, e.g. "EnergyAndFuels", "CropsAndStaples".
variable	string	Machine-readable series identifier, e.g. "bitumen_cny_t".
variable_name	string	Human-readable series name, e.g. "bitumen CNY T".
idx	int64	Rolling-window index of this prediction origin within its variable (0 = earliest).
date	string	Calendar date of this prediction origin.
freq	string	Sampling frequency, e.g. "1D", "1W".
background	string	Domain description for the series.
scenario	string	Prediction-target period and time-stamped related news events, numbered as they should be cited.
holiday_info	string	Upcoming calendar effects relevant to the forecast window.
covariates_info	string	Summary statistics of related indicator series over the recent past.
past_timestamps	list[string]	Timestamps of the historical window, oldest first.
past_values	list[float64]	Observed values at past_timestamps.
future_timestamps	list[string]	Timestamps to forecast, in order.
future_values	list[float64]	Ground-truth values at future_timestamps (the forecasting target).

Data Splits

train and test, split per variable: for every variable, the last 8 rolling windows (by idx, i.e. the most recent prediction origins) go to test and the rest to train. This is a walk-forward, per-variable holdout, not a random split -- samples of the same variable are overlapping windows over one time series, so a random split would leak near-duplicate windows (and future values already seen in an earlier window's history) across the boundary.

This does not reproduce the original paper's official split. The paper's protocol trains on a 2018-2022 split and evaluates on a disjoint 2023-2025 period; that training-period data is not present in the upstream repository this mirror was built from (its date coverage is 2022-05 to 2026-03 at the time of this mirror, consistent with the project page's "Dataset updated to 2026.03" -- the repository appears to be a rolling window of recent history, not the original release's full range). The per-variable holdout above is a substitute that avoids leakage but is not comparable to results reported against the paper's own test split.

The google-research/google-research monorepo (Apache 2.0) includes what appears to be the official data-generation pipeline for TimesX's textual context, under TimesX/dataset_agent, with batch configs targeting the 2018-2019 period by name (configs/Batch1/US-1.yaml: project_name: SearchTrend-Event-1112-Currency-2018-2019, time_range: {start: 2018-01-01, end: 2019-12-31}). Regenerating the missing training-period data through that pipeline would need live Google Custom Search and Gemini API access and has not been attempted for this mirror.

Dataset Creation

Series and their textual context were assembled by TimesX's automated data-generation pipeline (see the paper, Section 3 and Appendix E), which retrieves and fact-checks time-stamped news events against source pages before attaching them to a prediction window. This mirror only reshapes the upstream JSON files into a single typed table (uniform float64 values, one row per sample); no values were recomputed or edited.

Citation

@inproceedings{
liu2026rethinking,
title={Rethinking Multimodal Time-Series Forecasting Evaluation},
author={Haoxin Liu and Yichen Zhou and Rajat Sen and B. Aditya Prakash and Abhimanyu Das},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=auG2Cr76ed}
}