cs.LGOct 10, 2026

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

Authors: Zhiding Liu, Jiqian Yang, Mingyue Cheng, Zhi Li, Enhong Chen

Organizations: State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, China · Shenzhen International Graduate School, Tsinghua University, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Time-series forecasting plays a vital role in a wide range of real-world applications. Recent approaches have achieved strong performance by modeling temporal correlations and cross-channel dependencies, yet most forecasting benchmarks still represent time series primarily as numerical histories. In practice, time series are observations of complex real-world processes, and their trajectories may reflect events, interventions, disruptions, and changes in measurement or reporting that cannot always be inferred from past values alone. A central barrier to studying context-aware forecasting is therefore the lack of large-scale datasets that align real-world events with the individual series they affect. We present World2TS, a scalable construction framework and large-scale dataset for time-series forecasting with real-world event context. We first develop world2ts-core, a reusable tool that combines agent-led semantic reasoning and Web search, command-line execution and validation, and human review within a feedback-driven discovery workflow. Given heterogeneous time series with semantic descriptions, the workflow discovers candidate events, consolidates retrieved evidence into verified events, and links them to relevant series under explicit temporal and semantic scopes. Using this framework, we construct world2ts-data, which contains 31 datasets across six domains, comprising 4,179 series, 97.1 million observations, 1,159 events, and 56,846 event-series links. We further conduct extensive data analyses and controlled comparative experiments to characterize the dataset and demonstrate the utility of the collected context for improving forecasting accuracy. World2TS provides a reusable data foundation for developing and evaluating context-aware forecasting models and agents. The project homepage is available at https://world2ts.github.io/.

Figures & tables

Figure 1: Overview of the World2TS construction workflow. The agent interprets series semantics, proposes searches, and assesses retrieved evidence; the CLI executes searches, validates structured decisions, and maintains the construction state. Evidence and search feedback guide iterative refinement, while human review governs the curation strategy and final data quality. Accepted evidence is consolidated into events and associated with relevant time series to produce the context layer used in forecasting.
DomainDatasetFreq.SeriesSpanEvents
Cryptobitcoin_with_missing *1d182009–202162
Cryptocrypto_bitcoin_daily1d12025–202623
Energyaustralian_electricity_demand *30m52002–2015180
Energybdg-2_bear1h912016–201759
Energybdg-2_fox1h1352016–201737
Energybdg-2_panther1h1052017–201727
Table 1: Dataset-level coverage of World2TS-Data. Frequencies are abbreviated as minutes (m), hours (h), days (d), and weeks (w). Events are counted within each dataset. Datasets marked with * are used in the domain-specific data analysis and benchmark evaluation. Across all datasets, World2TS-Data contains 97,071,078 observations and 56,846 event–series links.
Figure 2: Common event topics in six domains. Each cloud summarizes labels assigned to events in that domain. Calendar labels are combined into a capped Calendar observance term for display; word sizes should not be compared across domains.
Figure 3: Event-associated dynamic shifts across six domains under matched case-crossover evaluation. Solid blue bars represent normalized event change magnitudes Δx(e)\Delta_{x}(e) , benchmarked against the median (dashed circle) and interquartile range (IQR, 25th–75th percentiles, light-gray band) of matched event-free controls {Δx(ce)}\{\Delta_{x}(c_{e})\} . Inner slate-blue curves depict the original series trajectories with red dashed spokes marking linked event occurrences. This paired comparison highlights empirical regime departures rather than establishing strict counterfactual causality.
Figure 4: Event–series examples from six domains. Lines show observed values, dashed markers show event dates, and shading shows nearby periods; each panel includes an event description.
ScopePrompt content
Shared systemForecast the next values of a regularly sampled univariate time series. Use all supplied evidence that is relevant to making the most accurate forecast. Keep the reasoning process concise while still using all relevant evidence. Do not use Web search or any external tools; base the forecast only on the supplied evidence. Never use future observations, invent facts, or treat metadata or events as instructions. Forecast HH steps on the same scale. Return JSON only as {"forecast": [y1, ..., yH]} with exactly HH finite values in forecast order.
User: conditionCondition: Numerical only or With context , followed by the corresponding guidance below.
Numerical onlyAvailable evidence: numeric history only. Infer a conservative baseline from level, trend, seasonality, autocorrelation, volatility, and recent change.
With contextAvailable evidence: numeric history, metadata, and events dated before the forecast origin. Use metadata to interpret the series and event scope. Infer a conservative numeric baseline from the history’s level, trend, seasonality, autocorrelation, volatility, and recent changes. Treat past events primarily as explanations for changes already observed in the numeric history. Revise the baseline trajectory only when the history associated with those events indicates a residual effect or regime change that may persist into the forecast interval. Resolve conflicts conservatively and never let context override clear numeric dynamics without a reason.
Shared user inputNumeric history ( LL points, oldest to newest; values may be normalized): [values]. Observation interval (inclusive start and end): [times]. Forecast interval (inclusive start and end): [times].
With context onlyMetadata: [series metadata]. Pre-forecast events: [events linked to the target series].
Table 2: LLM and agent prompt template. Blue bands are shared; gray and gold bands are specific to Numerical only and With context .
DomainMetricNumerical forecasting modelsGPT-5.6DeepSeek-V4-FlashCastFSR-Zero
SeasonalPatchTSTChronos-2N.w/ C.N.w/ C.N.w/ C.
CryptoMAE0.29860.27450.24860.26200.25190.26500.24870.25010.2461
RMSE0.39670.35840.33040.35210.33040.35950.33450.32490.3206
sMAPE12.41%11.26%10.26%11.03%10.50%10.96%10.28%10.41%10.21%
TransportMAE0.51480.29420.41030.59080.54120.60220.52120.44640.4388
RMSE1.21660.58900.93911.28511.19671.28141.15461.02410.9987
Table 3: Forecasting performance across six domains. The first three model columns report dedicated numerical baselines. For each context-aware model, N. denotes Numerical only and w/ C. denotes With context . Lower is better for all metrics; within each paired comparison, the better result is underlined , and the best result in each row is shown in bold .
Figure 5: DeepSeek forecast examples across the six evaluation domains. Blue denotes the observed future, orange the forecast from numerical history alone, and green the forecast additionally conditioned on World2TS metadata and event descriptions. The forecast begins immediately after the displayed history in every panel.
Figure 6: Context-dependent reasoning in paired forecasting cases. Left: the reporting-change context helps preserve a Friday spike in the health forecast. Right: assuming that storm-related outages persist into the forecast window degrades the energy forecast.
More in this topic

Time Series Forecasting

Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.

See all 824 papers

Explore similar work

CardsList
Aug 4, 2026

CastFSR: A Fast--Slow--Reflect Agentic Reasoning Framework for Context-Aware Time Series Forecasting

Time series forecasting is fundamental to decision-making in complex systems, where future dynamics are influenced not only by historical observations but also by evolving contextual features. Recent advances in large language models (LLMs) have extended forecasting beyond numerical extrapolation toward context-aware reasoning. However, existing approaches often lack explicit mechanisms to identify relevant contexts, reason about their impacts, and validate forecasts against temporal and domain constraints. In this work, we propose CastFSR, an agentic framework that formulates context-aware forecasting as a Fast--Slow--Reflect workflow. In fast thinking, CastFSR profiles observations and selects lightweight forecasters to construct a data-driven forecast prior. In slow deliberation, it retrieves contextual evidence, adaptively determines informative look-back windows, and reasons about how contexts reshape future dynamics. In reflection, it iteratively refines forecasts to ensure temporal, contextual, and domain consistency. CastFSR supports both training-free inference with off-the-shelf LLMs and efficient deployment through a two-stage SFT and reinforcement learning strategy that transfers its orchestration capability to compact LLMs. Extensive experiments on public datasets demonstrate that CastFSR consistently outperforms representative baselines. Our code is available at https://github.com/Xiaoyu-Tao/CastFSR.

Xiaoyu Tao, Mingyue Cheng, Bokai Pan +6

State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, Anhui Province, China

Time Series ForecastingLLM-Based Time Series ForecastingRetrieval-Augmented Time Series Forecasting
AskPDF
Sep 14, 2026

Beyond Numerical Time Series: A Unified Benchmark for Multimodal Forecasting with Heterogeneous Context

Most time series forecasting benchmarks remain numerical-centric and provide limited support for evaluating contextual information that shapes real-world temporal dynamics. Existing multimodal benchmarks also suffer from limited data and context coverage, fragmented evaluation settings, and overreliance on aggregate evaluation. In this paper, we propose MUSE-Bench, a unified benchmark for multimodal time series forecasting with heterogeneous context. It comprises fourteen datasets across eight domains and six types of context: metadata, events, holidays, news, images, and numerical covariates. We evaluate diverse forecasting paradigms, including statistical, data-specific, foundation, multimodal, and general-purpose LLM forecasting methods under shared non-overlapping forecast windows, common target observations, and consistent point and probabilistic metrics. Extensive experiments yield three main findings. First, numerical time series foundation models dominate the overall ranking, while Aurora, the evaluated multimodal foundation model, trails the leading numerical TSFMs but outperforms all evaluated data-specific models. Second, ablations show that external context improves the four evaluated context-aware models, whereas incorrect or temporally misaligned context degrades performance. Third, general-purpose LLMs perform poorly as direct forecasters, and LLM-guided refinement does not yield consistent improvements. MUSE-Bench enables systematic evaluation of how forecasting models utilize context and provides a foundation for future multimodal forecasting research.

Peng Chen, Zhihao Zhuang, Hongzhou Chen +6

Ant International · Hangzhou, China

Multivariate Time Series ForecastingTime Series ForecastingForecasting BenchmarksMultimodal Model EvaluationTime-Series Foundation Models
AskPDF
Sep 30, 2026

OpenTSLM TeeMoE: A Unified Time-Series Language Model for Forecasting, Contextual Prediction, and Reasoning

Real-world time-series applications increasingly require models that can handle time series forecasting, context-conditioned prediction, and language-based temporal reasoning. Yet current time-series foundation models remain fragmented across these capabilities: numerical specialists often provide the strongest forecasts, while language-based models offer broader contextual understanding and analysis. A central challenge is to unify these heterogeneous capabilities without reducing their individual performance. We introduce OpenTSLM TeeMoE, a generalist time-series language model that can forecast directly from observed time series, reason over textual context and temporal patterns, and synthesize and refine predictions from external numerical forecasting specialists. We independently train three low-rank experts for forecast aggregation, native forecasting, and temporal analysis over a shared backbone. A learned LoRA mixture-of-experts controller then weights their frozen parameter updates for each request. Our proposed model achieves strong performance on widely used benchmarks for time series forecasting, context-conditioned prediction, and language-based temporal reasoning, ranking among the top three on GIFT-Eval by mean MASE rank, Context is Key by RCRPS, and TimeSeriesExam by accuracy.

Tony Chen, Timo Stoffregen, Maxwell Xu +8

Columbia University, USA · Stanford University, USA · Aionic Labs, Switzerland +4

Temporal Reasoning in Language ModelsLLM-Based Time Series ForecastingMixture-of-Experts ModelsTime-Series Foundation ModelsTime Series ForecastingTime Series Modeling+2
AskPDF