Open Dataset · Real-World Events × Time Series

World2TS: Real-World Event Context for Time-Series Forecasting

A human-reviewed dataset that links diverse time series to relevant real-world events, records when the information became available, and preserves the evidence behind every link.

Zhiding Liu^*1 Jiqian Yang^*2 Mingyue Cheng^†1 Zhi Li^†2 Enhong Chen^†1

¹ State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, China

² Shenzhen International Graduate School, Tsinghua University, China

^* Equal contribution · ^† Corresponding authors

Paper · soon Code · GitHub Dataset · Explore Dataset · Hugging Face

01 · Overview

Real-world events linked to relevant time series.

World2TS brings public benchmark collections and newly collected online series into one dataset with shared metadata and validation rules.

Forecasting in the real world is rarely isolated from policy decisions, extreme weather, public-health shifts, operational disruptions, and market news. World2TS links events to the time series they are relevant to, while treating a shared date as an association rather than proof of causation. An LLM proposes and reviews event links, software records and validates the files, and people approve the search plan and final dataset.

01
Relevant event linksEach event is linked only to time series for which it is relevant.
02
Explicit event timingWe record whether information was available before, during, or after the forecast period.
03
Human reviewPeople review the search plan and final event records before release.

World2TS at a glance · real multi-event energy series

See verified events along a real time series.

The 2024 WTI price history moves from right to left. When an event date reaches the center line, its source and summary appear. The markers show association with the series, not proof that an event caused a price movement.

WTI spot price at Cushing · USD/barrel

2024 timeline replay

Jan 02, 2024 0 of 11 events shown

31Datasets

4,179Series

97.1MObservations

1,159Events

56,846Event links

6Domain groups

02 · Framework

LLM proposals, automated searches, and human review.

A real WTI price example shows how World2TS identifies candidate events, searches for evidence, reviews sources, and publishes approved records.

World2TS workflow for finding and reviewing events linked to a time series — **Workflow.** The LLM proposes candidate events and search queries. Software runs each search and records the results. People review the search plan and approve the final event records.

03 · Dataset

One dataset, many sources.

World2TS releases event descriptions and series links for all 31 configurations. Numeric Arrow files are hosted for 26 configurations; the remaining numeric series must be obtained from their original providers under the providers' terms. Browse the data and source notes on Hugging Face ↗

Name	Domain	Frequency	Series	Time span	Events

31 datasets shownNumbers shown here come from the current World2TS dataset summary

Event coverage by dataset

Average number of linked events per series and the percentage of series with at least one event dated within its observation period.

Name	Avg. events / series	Series coverage

04 · Analysis

What kinds of events does World2TS capture?

Dataset summaries and case studies show the range of recorded events and the limits of what can be inferred from their timing.

Common event topics by domain

Each word cloud summarizes common event topics for one of six domains. Environmental topics are grouped with weather, and multi-word terms are kept together.

Domain-specific word clouds for six World2TS domains — **Event topics.** Each word cloud keeps multi-word event phrases together and summarizes one domain independently.

Time-series and event examples

Six examples cover weather, finance, health, transport, crypto, and energy. Highlighted periods show nearby dates, not causal effects.

Six time-series event case studies across World2TS domains — **Examples.** Each panel retains the original units, marks the event date and nearby period, and gives an event description relevant to that series.

Changes near recorded events

Representative yearly series are aligned at event dates and compared with similar periods from the same series that have no recorded event nearby.

Circular event-associated change comparison for energy, crypto, and weather series — **Changes near event dates.** The inner ring follows calendar order. Radial bars show the size of the before-and-after change around each event; the gray band and gold line summarize similar periods without a recorded event. These comparisons describe associations and do not establish causality.

How the comparison is calculated

Same series · same period length · similar comparison dates

Before-and-after change score

C (t; w) = \frac{| m_{after} - m_{before} |}{s_{Δ y}}

Before/after medians in equal windows, normalized by 1.4826 × MAD(Δy).

Event-to-comparison ratio

A = \frac{{median}_{e} [C (t_{e}; w)]}{{median}_{e} [C (t_{c, e}^{*}; w)]}

A > 1: changes are larger near recorded events than in comparable periods without a recorded event.

Same period lengthDaily data: 3 days before and after. Hourly data: 24 hours before and after.
Match the calendarUse the same series, year, quarter, weekday, and hour when applicable.
Match recent variabilityChoose a period away from recorded events whose recent variation is most similar.

05 · Evaluation

Does real-world event context improve forecasting?

The main comparison spans six domains and three forecasting metrics. For each language model, adjacent rows show forecasts with and without World2TS event descriptions.

Main forecasting results

Forecasting errors are reported by domain. For each language model, the Numerical-only row and the W / Context row show what changes when event descriptions are added. Lower is better for every metric.

Forecasting errors by model, input setting, domain, and metric.
Model	Input	Crypto			Transport			Energy			Weather			Health			Finance
Model	Input	MAE	RMSE	sMAPE	MAE	RMSE	sMAPE	MAE	RMSE	sMAPE	MAE	RMSE	sMAPE	MAE	RMSE	sMAPE	MAE	RMSE	sMAPE
Seasonal-7	Numerical only	0.2986	0.3967	12.41%	0.5148	1.2166	53.88%	0.4122	0.6260	68.47%	0.4437	0.5232	99.05%	0.0742	0.2255	13.35%	0.2954	0.3440	15.66%
PatchTST	Numerical only	0.2745	0.3584	11.26%	0.2942	0.5890	45.31%	0.1725	0.2462	37.05%	0.2986	0.3774	77.57%	0.3231	0.5076	54.66%	0.3844	0.4568	19.50%
Chronos2	Numerical only	0.2486	0.3304	10.26%	0.4103	0.9391	49.17%	0.2901	0.4413	53.89%	0.2981	0.3888	71.74%	0.1231	0.2848	20.72%	0.2176	0.2858	11.13%
GPT	Numerical only	0.2620	0.3521	11.03%	0.5908	1.2851	62.57%	0.3147	0.4926	55.95%	0.4467	0.5424	88.44%	0.1082	0.2976	17.72%	0.1886	0.2376	10.17%
GPT	W / Context	0.2519	0.3304	10.50%	0.5412	1.1967	58.22%	0.2586	0.3842	49.52%	0.3557	0.4411	79.66%	0.0957	0.2837	16.16%	0.1761	0.2241	9.47%
DeepSeek	Numerical only	0.2650	0.3595	10.96%	0.6022	1.2814	65.22%	0.3145	0.5067	53.98%	0.4222	0.5194	89.24%	0.0744	0.2463	12.35%	0.1856	0.2398	10.00%
DeepSeek	W / Context	0.2487	0.3345	10.28%	0.5212	1.1546	58.79%	0.2515	0.3808	48.63%	0.3668	0.4719	79.00%	0.0680	0.2218	12.32%	0.1865	0.2419	9.96%
CastFSR-Zero	Numerical only	0.2501	0.3249	10.41%	0.4464	1.0241	51.94%	0.2628	0.4099	48.92%	0.2796	0.3672	66.90%	0.1267	0.3255	20.94%	0.2081	0.2628	10.80%
CastFSR-Zero	W / Context	0.2461	0.3206	10.21%	0.4388	0.9987	53.91%	0.2415	0.3651	47.96%	0.2817	0.3644	69.57%	0.0904	0.2669	15.96%	0.2006	0.2602	10.31%

↓ Lower is better for MAE, RMSE and sMAPEBold marks the lowest error in each domain and metric

Forecast examples

Forecast examples across six domains

Six examples compare the observed future values with forecasts based on numerical history alone and forecasts that also receive World2TS event descriptions.

Forecast comparison for weather, finance, health, transport, crypto, and energy time series — **Forecast comparison.** Blue shows the observed future values, orange the numerical-only forecast, and green the forecast with World2TS event descriptions. In every panel, the forecast begins immediately after the displayed history.

Reasoning comparison

How event context changes LLM reasoning

Both examples start from the same numerical history and show how an added event description changes the model’s interpretation and forecast. It helps in the health example but hurts in the energy example.

Health · context explains a change in the reporting schedule

Positive Case

Health reasoning trace comparing numerical-only and context-aware forecasts — **Positive Case.** With context, the model links the floor-and-spike pattern after October 20 to the CDC switch to weekly reporting, repeats the first complete week after the change, and retains the Friday spike.

Energy · context overstates lingering storm effects

Negative Case

Each example follows the steps present in the model output. We do not force both reasoning traces into the same template.

06 · Citation

Cite World2TS.

Author information is included below; venue and publication identifiers will be updated after finalization.

@misc{world2ts2026,
  title  = {World2TS: Real-World Event Context for Time-Series Forecasting},
  author = {Liu, Zhiding and Yang, Jiqian and Cheng, Mingyue and Li, Zhi and Chen, Enhong},
  year   = {2026},
  note   = {Dataset and project page}
}

Released files

Minimal by design.

The public release contains dataset descriptions, event records, and links between time series and relevant events. Numeric Arrow files are included where redistribution is supported.

dataset_info.jsonunified_events.jsonlunified_event_series.jsonl