Rethinking Multimodal Time-Series Forecasting Evaluation

# Rethinking Multimodal Time-Series Forecasting Evaluation
Source: [https://arxiv.org/html/2607.06973](https://arxiv.org/html/2607.06973)
\\pdftrailerid

redacted\\correspondingauthorHaoxin Liu[<hliu763@gatech\.edu\>](https://arxiv.org/html/2607.06973v1/mailto:[email protected])\. This work was done while Haoxin Liu was a Student Researcher at Google Research\. This work is accepted by The International Conference on Machine Learning \(ICML\) 2026\.

Yichen ZhouGoogle ResearchRajat SenGoogle ResearchB\. Aditya PrakashGeorgia Institute of TechnologyAbhimanyu DasGoogle Research

###### Abstract

We introduce a new context\-enriched, multimodal time series forecasting benchmark, TimesX\. TimesX contains a wide selection of high\-quality real\-world time series with diverse domains and textual contexts obtained from an automated data generation pipeline, which helps address three main issues of existing multimodal forecasting benchmarks: \(1\) poor generalization due to the small scale and synthetic nature of benchmark data, \(2\) very limited types of textual contexts in the benchmarks, and \(3\) an inability to mitigate data leakage in evaluation\. We conduct a thorough empirical study of zero\-shot multimodal forecasting approaches on TimesX\. Our results suggest that many approaches that perform well on existing benchmarks may fail on TimesX\. In contrast, simple ensemble methods that leverage rich textual context accompanying time\-series can outperform strong baselines on TimesX\.

## 1Introduction

Time\-series forecasting \(TSF\)Liu\_FOILis a ubiquitous task across numerous domains and is essential for informed decision\-making\. Motivated by success in NLP, there has been significant work in recent years on time\-series foundation models \(TFM\) for forecasting, ranging from re\-purposing LLMs directly for forecasting\(gruver2023large;tan2024\)to fine\-tuning pretrained LLMs on time\-series data\(zhou2023one;chang2023llm4ts\)to pretraining time\-series foundation models from scratch\(das2023decoder;goswami2024moment;woo2024unified;ansari2024chronos;kamarthi2024lptm\)\. These models have demonstrated promising zero\-shot TSF capabilities, often outperforming traditional statistical and supervised methods, using only the historical context of the time\-series at inference time\. Although historical numerical data provide a fundamental basis for prediction, they often lack the complete context necessary for reliable and accurate forecasts\. Human forecasters often integrate additional information such as background knowledge, events and constraints, which can usually be captured through natural language\(Liu\_et\_al\_2024a\_LSTPrompt\)\- we call these*textual contexts*\. Improving the ability of time\-series foundation models to effectively leverage and integrate diverse textual contexts remains an ongoing challenge\. This highlights a crucial need for high\-quality and comprehensive multimodal datasets and benchmarks to propel research in this area\.

At the same time, existing context\-enriched multimodal TSF evaluation benchmarks and their construction techniques present several limitations when benchmarking pretrained models on them\. The first limitation is the unknown generalization gap between the benchmarking metrics and the real world performance\. This is often caused by either the benchmark being restricted to asmall scalewith a narrow selection of domains, or it being fullysynthetic\.

The second significant challenge across existing benchmarks isdata leakage\. Pretrained TFMs and LLMs may have already ingested evaluation data, leading to unknown data contamination and unfair comparison between methods\. A benchmark itself is subject to short\-lived validity as pretrained models will continuously update their pretraining datasets, which results in their knowledge cut\-off becoming more recent than the benchmark’s data\.

The third limitation comes from our observation that the types and granularities of textual contexts in previous benchmarks are often limited and vary wildly\. Textual contexts derived from real data include metadata information about the time\-series, timestamp\-related information such as important dates and holidays, textual event data related to the time\-series, and textual data capturing statistics and trends of related exogenous variables correlated with the target time\-series\. The ability of multimodal models to process these different types of contexts, and the availability and quality of these different contexts in the datasets have a significant effect on the model’s benchmark performance\. When a multimodal model falls short on existing benchmarks, it is difficult to conclude whether the shortcoming is due to the fact that the method does not effectively utilize the provided information, or it is because the contexts provided in the benchmark are of a specific type, too vague, or lacking relevant details\.

To address these fundamental limitations and establish an unbiased benchmark for context\-enriched time series forecasting, we introduceTimesX, a novel multimodal benchmark designed to focus on \(1\) real, large\-scale, cross\-domain data, \(2\) a dataset generation pipeline providing continuous mitigation of data leakage, and \(3\) a collection of comprehensive, fine\-grained textual contexts\. We further introduce several promising baselines that combine TFMs and LLMs to achieve non\-trivial multimodal, context\-enriched forecasting performance on our benchmark\.

The primary contributions of this paper are threefold:

1. 1\.TimesX is, as far as we are aware, the first real\-world, large\-scale, and cross\-domain context\-enriched time series forecasting benchmark\. It encompasses 19 diverse domains with a total of 190 variables, covers diverse global geographical regions and includes daily and weekly frequencies\. It also links each target numeric series to multiple forms of detailed textual contexts\. Unlike existing multimodal time\-series benchmarks, all time series and textual data in TimesX are from real\-world observations\.
2. 2\.We propose two new mechanisms in the dataset generation pipeline of TimesX to help prevent data leakage and guarantee the validity of the textual context\. The first one is an automated data collection pipeline that adopts strict timestamp alignment and isolation during each of its steps\. The second one is a hypothesizer\-verifier\-enricher framework for fact\-checking and enriching the textual contexts\. These mechanisms combined ensure that TimesX is verifiable, leakage\-free, and can be updated in the future for benchmarking methods with new pretraining cut\-off dates\.
3. 3\.We empirically verify that TimesX addresses the shortcomings of existing benchmarks\. In addition, we conduct a thorough empirical comparison of current \(zero\-shot\) multimodal TSF approaches on TimesX, involving more than 312,000 independent LLM inferences and revealing new observations that were not discovered by existing benchmarks\. In particular, we discover that earlier benchmarks likeWilliams\_2024\(while being a great test for instruction following in forecasting\) tend to severely over\-estimate the performance of LLMs over TSF models\. At the same time, benchmarks likeLiu\_et\_al\_2024b\_TimeMMDunder\-report the importance of textual context information in forecasting accuracy\. Limitations and future works are detailed in Appendix[A](https://arxiv.org/html/2607.06973#A1)\.

## 2Related Work

In addition to the time\-series foundation models mentioned previously, recent worksjin2023time;Liu\_et\_al\_2024a\_LSTPrompt;wang2025chattimehave started to adapt pre\-trained LLMs to perform forecasting when textual contexts are available, often by aligning the modalities of time series and natural language\. The development of these models has been paralleled by new benchmarks designed to evaluate them\. ChatTS\(ChatTS\), for instance, generates synthetic time series paired with detailed attribute descriptions to train a multimodal LLM for understanding and reasoning tasks\. Similarly, the Time\-MQA framework\(Kong\_2025\)introduces the TSQA dataset, which contains real\-world time series with template\-based meta information and turns time\-series forecasting into a question\-answer format\. Benchmarks such as those referenced in CiK\(Williams\_2024\)contain real\-world time series and manually crafted textual contexts that are critical for forecasting\. Time\-MMD\(Liu\_et\_al\_2024b\_TimeMMD\)provides a context enriched benchmark constructed by keyword\-based web searches of nine domains, one variable for each\. MTBench\(Chen\_2025\)aligns stock prices with financial news and weather reports with temperature records, curating tasks that require reasoning about them\. The MoTime\(Zhou\_2025\)dataset suite provides a collection of multimodal datasets pairing time series with static modalities like text and images\. Noticeably, these early efforts were often constrained by a scarcity of high\-quality, large\-scale, real\-world datasets, leading some to rely on synthetic data and specific forms of textual contexts\. We summarize several representative benchmarks compared to our proposed TimesX in Table[1](https://arxiv.org/html/2607.06973#S2.T1)\.

Table 1:Comparison of multi\-modal TSF benchmarks\. Leakage\-Free indicates whether the benchmark can guarantee no data leakage for the latest pretrained models, detailed in Section[3\.1\.2](https://arxiv.org/html/2607.06973#S3.SS1.SSS2)\.
## 3The TimesX Benchmark

### 3\.1Design Principles of TimesX

We will first present the core principles that distinguish TimesX from other multimodal forecasting benchmarks: \(i\) real\-world data, \(ii\) leakage mitigation, and \(iii\) comprehensive, high\-quality context and \(iv\) large\-scale evaluation\. Note that we further provide empirical evidence that demonstrates the importance of these principles and where other benchmarks might fall short\.

#### 3\.1\.1Real\-World Data

Principle Description: TSF is highly complex due to real\-world dynamics, complex processes, and diverse domains\. On the other hand, many existing benchmarks use either synthetic time series, synthetic textual contexts, or both, without demonstrating how the conclusions on synthetic data can transfer to real\-world scenarios\(Williams\_2024;tan2024\)\. TimesX differentiates itself from those benchmarks byusing real\-world data for both the time series and the textual contexts\.

Why it matters: To test whether conclusions on synthetic data can transfer to real data, we evaluate three methods on a synthetic context dataset \(CiK\) and our real dataset \(TimesX\)\. We report the aggregated MASE \(Mean Absolute Scaled Error\) of the three methods which are: \(i\) TimesFM\-2\.5 \(a TFM\) using only the time series, \(ii\) Gemini\-2\.0\-Flash \(an LLM\) using both the time series and the textual contexts, and \(iii\)CodeRev, Gemini\-2\.0\-Flash writing and executing code to revise TimesFM\-2\.5 outputs based on the context\. More details are in Appendix[N](https://arxiv.org/html/2607.06973#A14)\.

As shown in Figure[1](https://arxiv.org/html/2607.06973#S3.F1),the rankings of the three methods on CiK \(synthetic\) and our real data are completely different\. On CiK, directly feeding the time\-series along with associated text context to Gemini has a huge lead over TimesFM\-2\.5\. Moreover, asking Gemini to edit the output of TimesFM\-2\.5 via code does even better, significantly outperforming the other two methods\. On the other hand on TimesX, we see that all the methods are much closer, and TimesFM\-2\.5 is actually better than Gemini\. FurtherCodeRevdegrades the performance of using both TimesFM\-2\.5 and Gemini directly\. This shows that CiK has a strong bias that clouds its conclusions: in CiK, textual contexts are both synthetic and extremely specific \- these specific contexts are then used to write code to modify an input time\-series and generate the final forecasting task\. Therefore, such tasks can be easily solved by instruction\-following language models and even better by language models that generate code to accomplish a simple modification task\. These conclusions, however, do not generalize to the real\-world tasks, whose contexts are less specific and do not relate to the forecasting task via the coding path\.

![Refer to caption](https://arxiv.org/html/2607.06973v1/x1.png)Figure 1:Synthetic \(CiK\) versus real\-world \(TimesX\) performance under the same setup \(lower MASE is better\)\. The ordering of TimesFM\-2\.5, Gemini\-2\.0\-Flash, andCodeRevflips between the two datasets, showing that synthetic generation can bias rankings toward instruction\-following LLM and coding\. We show results on the Future subset of CiK here\. Table[15](https://arxiv.org/html/2607.06973#A18.T15)further reports the average performance across all subsets with the same conclusion\.
#### 3\.1\.2Mitigating Data Leakage

Principle Description: A fair benchmark dataset should be able to mitigate any potential leakage of benchmark tasks into the pretraining datasets of the models being evaluated\. Given that many pretrained models do not release their pretraining datasets, existing benchmarks that claim to use data from unused sources \(data isolation\) for evaluation have three drawbacks: \(1\) Potentially unidentifiable data contamination\. \(2\) Limited number of clean tasks left for benchmarking multiple methods\. \(3\) Short\-lived validity i\.e\., future versions of the models considered may unintentionally leak the benchmark into their pretraining data\.

In contrast,TimesX adopts stricttime isolationto prevent data contamination: we timestamp TimesX thoroughly, and recommend strictly using benchmark tasks that occur after a target model’s knowledge cutoff date\. Moreover,the TimesX pipeline is designed as re\-updateable over time, i\.e\., data collection, validation, alignment, and model evaluation can all be automated\. To the best of our knowledge, TimesX is the first benchmark capable of being automatically refreshed for evaluating future pretrained methods\.

Table 2:Performance before and after the two LLMs’ knowledge cutoff \(June 2024\) on the Search Trend subset \(120 variables\)\. Lower MASE is better\. Setup is detailed in Appendix[B](https://arxiv.org/html/2607.06973#A2)\.Why it matters: To study how data contamination affects evaluation results, we evaluate on 120 variables from the Search Trend \(see Section[3\.2](https://arxiv.org/html/2607.06973#S3.SS2)\) subset of TimesX\. We keep the numeric series, windowing, and prompts fixed, and split the test period by the public knowledge cutoff of Gemini\-2\.0\-Flash and DeepSeek\-V3 \(June 2024\)\. We report the MASE aggregated by Geometric Mean in Table[2](https://arxiv.org/html/2607.06973#S3.T2)\.

The pretraining data descriptions of the two TFMs preclude any time\-series data leakage in both time periods, andboth TFMs report stable resultswith less than 2% delta\. In contrastboth LLMs see an increase of error by about 13% after their knowledge cutoff\.

#### 3\.1\.3Comprehensive and High\-Quality Context

Principle Description: Most multimodal forecasting benchmarks have a very limited set of text contexts \- restricted to either static metadata\(Zhou\_2025\), date and weather derived features\(wang2025chattime\)or synthetically generated contexts\(Williams\_2024\)\. In contrast, TimesX uses not onlya union of all common context types, but also a morehigh\-quality description of text eventsand corresponding alignment of these contexts\.

In particular, we categorize these various kinds of contexts into: \(1\)Metadata: descriptive high level summaries of the forecasting variable in question \(2\)Calendar\-derived features like holidays \(3\)Covariateswhich cover textual information around other related time series \(4\)Time\-stamped events: Textual description of related events aligned with time\-windows of the variable in question\. Section[3\.2\.2](https://arxiv.org/html/2607.06973#S3.SS2.SSS2)contains more details about how we generate these contexts\.

Why it matters: TimesX links each target time\-series to all these available context types and aligns them with the timestamps of the time series\. The end result is a benchmark with diverse, high\-quality textual contexts\. To quantify its effect on TSF accuracy, we run a controlled experiment where we take the time\-series in the Time\-MMD benchmark and use our methodology to replace the textual events using the construction rules in Appendix[E](https://arxiv.org/html/2607.06973#A5)\. We keep the numeric series, the LLM \(Gemini\-2\.0\-Flash\), and the prompt fixed\.

Table[3](https://arxiv.org/html/2607.06973#S3.T3)\(full results in Table[18](https://arxiv.org/html/2607.06973#A26.T18)\) compares the performance of Gemini\-2\.0\-Flash given these newly generated text events versus the original context in the Time\-MMD benchmark\. Across all nine Time\-MMD datasets, thiscontext swap reduces the geometric–mean aggregated MASE from0\.9060\.906to0\.8400\.840\(a7\.3%7\.3\\%relative drop\), suggesting that conclusions drawn under low–quality context might be misleading;high–quality context changes the picture\.

Table 3:Controlled replacement of textual context onTime\-MMD\. More detailed results are in Appendix[S](https://arxiv.org/html/2607.06973#A19)\.We further show thatwith the current dataset scale \(typically fewer than 20 variables\), model performance rankings are quite unstable, as illustrated in Figure[34](https://arxiv.org/html/2607.06973#A26.F34)and Appendix[T](https://arxiv.org/html/2607.06973#A20)\.

### 3\.2Construction of TimesX

We first provide a high level summary of our dataset and dataset construction methodology\. TimesX contains time\-series obtained from19 domains\. Each domain has1010variables, resulting in a total of190variables\(quantities evolving with time\)\. TimesX coversdifferent geographical regions: North America \(United States, Canada, Mexico\), Asia \(China, India among others\), Europe \(United Kingdom, Norway among others\), South America and Africa\.

Note that we have a separate portion of the dataset that is derived from the time\-series in Time\-MMDLiu\_et\_al\_2024b\_TimeMMD, where the text context is generated using our methodology, but only used for ablations in Section[3\.1\.3](https://arxiv.org/html/2607.06973#S3.SS1.SSS3)\.

#### 3\.2\.1Numerical Time Series Construction

The numerical component consists of target variables whose future values are forecasted given a historical window\.

Granularities and sources\.TimesX includes two temporal granularities:*Weekly*series from Google Search Trends across 12 domains111[https://trends\.google\.com/trends/](https://trends.google.com/trends/); and*Daily*series including \(i\) commodity prices from a market data API222[https://marketstack\.com/](https://marketstack.com/)and \(ii\) major USD exchange rates from a currency rates API333[https://frankfurter\.dev/](https://frankfurter.dev/)\. These API data can be*automatically refreshed*to extend the benchmark over time\. Monthly/quarterly coverage is discussed in Appendix[A\.1](https://arxiv.org/html/2607.06973#A1.SS1)\.

Time alignment and missing values\.All series within the same frequency share the same timestamps\. Missing values are handled as described in Appendix[D\.1](https://arxiv.org/html/2607.06973#A4.SS1)\. Appendix[I](https://arxiv.org/html/2607.06973#A9)provides the full list of domains and variables\.

#### 3\.2\.2Text Context Construction

Our core contribution is aligning comprehensive and diverse text context information with the variables that evolve over time\. We will see that methods that capture both this context information along with the past of the time\-series can achieve superior performance compared to methods that only use the past time\-series component\. As mentioned before we provide 4 kinds of text contexts for all time\-series variables: Metadata, Calendar features, Covariates and Time\-Stamped Events\. Section[3\.2\.3](https://arxiv.org/html/2607.06973#S3.SS2.SSS3)shows an example of a time series with each of these four contexts\.

Metadata and calendar contexts\.The Metadata and Calendar text contexts are easily constructed from the sources of the numerical time\-series themselves and Python Holidays library respectively, as described in Appendix[C](https://arxiv.org/html/2607.06973#A3)\.

Covariate contexts\.For each time\-series in a given domain, we use the other time\-series from that domain to extract Covariate text contexts\. Specifically, we calculate features of these covariates over the historical window, including mean, median, max and min with the corresponding date, and overall trend direction\. These features are then transformed into natural language descriptions, detailed in Appendix[D](https://arxiv.org/html/2607.06973#A4)\.

Time\-stamped events via dataset agents\.Constructing and aligning textual events is the most challenging component\. Following our design principles in Section[3\.1](https://arxiv.org/html/2607.06973#S3.SS1), we must satisfy three goals at once:*real\-world*data,*leakage\-free \(refreshable\)*, and*high\-quality*\. A single LLM or a naive web search cannot reliably meet these goals\. Therefore, we adopt amulti\-agent automated workflowwith four agents:*Hypothesizer*,*Verifier*,*Enricher*, and*Synthesizer*, as illustrated in Figure[2](https://arxiv.org/html/2607.06973#S3.F2)\. Each agent interacts with an LLM under constraints and has dedicated tools \(time\-bounded web search and lightweight crawlers\)\. Overall, the Hypothesizer and the Verifier act adversarially to ensure event truthfulness and timestamp accuracy, which prevents information leakage; the Verifier then guides the Enricher to supply missing details thus ensuring quality\. The whole workflow is automated, enabling regular updates of the benchmark data\. We provide the details of each LLM role:

\(i\) the*Hypothesizer*identifies points of interest in the time series \(e\.g\., local maxima, unexpected movements or peaks in the corresponding search trend\) and iteratively calls an LLM agent with integrated web search to build an initial event set that is able to match these points\.

\(ii\) then, the*Verifier*uses crawlers to fetch relevant URLs corresponding to each event window and conducts fact checks, filters out any hallucination or leakage, and prepares a checklist of missing details\.

\(iii\) The*Enricher*resolves the checklist using strictly time\-bounded web searches, merging multiple sources to fill in the details\.

\(iv\) finally, the*Synthesizer*, a reasoning LLM, aggregates all pieces of evidence to finalize the event description, adjust timestamps, and discard events with unresolved doubts\.

Using this agent, TimesX automatically produces an event corpus that matches the time window, contains verifiable facts \(each claim has a supporting URL\), accurate time stamps, and rich detail\. More details and hyperparameter analysis experiments are provided in Appendix[E](https://arxiv.org/html/2607.06973#A5)\. An execution log demo is in Appendix[F](https://arxiv.org/html/2607.06973#A6)\.

![Refer to caption](https://arxiv.org/html/2607.06973v1/iclr2026/Pictures/agent_workflow_highres_V3.png)Figure 2:Overview of the dataset agents for automatic construction of event contexts\. The multi\-agent pipeline iteratively retrieves, verifies, and enriches event information\. In Phase 3, strict programmatic time range constraints are enforced to ensure timestamp accuracy and prevent future info leakage\. Detailed execution logs are provided in Appendix[F](https://arxiv.org/html/2607.06973#A6)\.
#### 3\.2\.3Data Example

We provide below an example from TimesX corresponding to theGAS\_PRICEtime\-series\.

![Refer to caption](https://arxiv.org/html/2607.06973v1/iclr2026/Pictures/gas_price.png)Figure 3:The numerical example from theGAS\_PRICE\.Time series: See Figure[3](https://arxiv.org/html/2607.06973#S3.F3)\.

Metadata: This time series records gasoline price \(USD/GAL\) in the Commodity Price domain, with daily frequency\. Prediction target period: from 2024\-09\-01 to 2024\-09\-15\.

Calendar: In the format of \(timestamp, value\), historical data is: \(2024\-06\-02, 2\.4116\), \(2024\-06\-03, 2\.3279\),… Upcoming holidays in the prediction window: Labor Day \(2024\-09\-02\)\.

Covariates: from 2024\-06\-01 to 2024\-08\-31: \(1\) Brent Crude Oil \(USD/BBL\): The maximum value was87\.4387\.43, occurring on July 4, the minimum…, showing an overall downward trend\. \(2\)…

Events: \(1\) On June 2, 2024, OPEC\+ agreed to extend deep oil output cuts \[1,3\], the cut of 2\.2 million bpd would be extended until September 2024, after which it would be gradually phased out \[2,3\]\. Source: \[1\] \[2\] \[3\]; \(2\) …

#### 3\.2\.4Quality Checks\.

Diversity\.We profile numerical and textual features of TimesX \(Appendix[J](https://arxiv.org/html/2607.06973#A10)\) and visualize them with PCA/t\-SNE \(Appendix[M](https://arxiv.org/html/2607.06973#A13)\)\. Samples are broadly dispersed, and variables from different sources form distinct clusters, supporting the benefit of multi\-source integration\.

Factual correctness\.Here, we focus on the most sensitive component: the accuracy of automatic timestamp annotation\. We ensure the accuracy via a three\-stage pipeline \(Appendix[G](https://arxiv.org/html/2607.06973#A7)\): verification, programmatic time\-bounded search \(±kkdays\), and cross\-source adjudication in the agent workflow\. A manual audit of 50 samples shows*94% exact matches; 4% conservative \(later\) offsets; and 2% earlier due to date ambiguity*\(Appendix[G](https://arxiv.org/html/2607.06973#A7)\)\.

### 3\.3TimesX Release and Refresh Consideration

The core TimesX dataset now contains 190 variables spanning Jan 2018 to Oct 2025 to support both training \(2018–2022\) and evaluation \(2023–2025\)\. We plan to update the TimesX dataset every three months and version each release\.

The extended TimesX further supports 11 languages beyond English \(Afrikaans, French, German, Hindi, Japanese, Korean, Portuguese, Simplified Chinese, Spanish, Swahili, and Turkish\), as well as 5 rare diseases\. See Appendix[A\.5](https://arxiv.org/html/2607.06973#A1.SS5)

## 4Empirical Study

### 4\.1Evaluation Settings

We focus on zero\-shot solutions and consider the following:

Pretrained TFMs: We select three SOTA TFMs including TimesFM\-2\.5, Moirai\-2\.0, and Sundial\. All are towards the top on GIFT\-Evalaksu2024giftbenchmark, with TimesFM\-2\.5 being the top open model with no leakage\.

Pretrained LLMs: We consider three well\-adopted LLMs \(implementation details in Appendix[Q](https://arxiv.org/html/2607.06973#A17)\): closed\-source Gemini\-2\.0\-Flash and GPT\-4o, and open\-source DeepSeek\-V3\.

Composed \(Ensemble And Agentic\) Solutions: We construct the following four solutions: \(1\)AvgEns: An ensemble that averages the forecasts from a pair of TFM and LLM\. We simply set equal weights\. \(2\)TextRev: An LLM taking the forecast of a TFM as text and revising it according to the context\. \(3\)CodeRev: An LLM writing and executing code to revise the forecast of a TFM according to the context\. \(4\)FuncRev: Similar toCodeRev, but the LLM is limited to a selection of functions\. See Appendix[Q](https://arxiv.org/html/2607.06973#A17)\.

For each variable in TimesX and its context corpus, the look\-back window and the forecast horizon are set to 96 and 12, respectively\. The rolling window is set to 4 for weekly data and 12 for daily data to balance sample count and sample diversity\. We then include all relevant metadata, calendar info and covariates info\. To avoid future info leakage, we select theKKmost recent textual events whose announcement dates strictly precede the first timestamp of the prediction horizon to add to the context\. We fixk=10k=10to balance prompt length and information effectiveness, following common practicesLiu\_et\_al\_2024b\_TimeMMD;li2025language\.

TimesX spans 2018–2025 and is refreshable\. However, to avoid pretraining data contamination, we construct evaluation examples whose forecast horizon begins after the pretraining cutoff of all involved models, detailed in Appendix[P](https://arxiv.org/html/2607.06973#A16)\. Unless stated otherwise, we use2024\-07\-01as the cutoff, i\.e\., evaluating only the subset whose horizon start time is 2024\-07\-01 or later\. See models’ knowledge cutoffs in Appendix[P](https://arxiv.org/html/2607.06973#A16)\. Web access is limited to the build\-time agent; the benchmark runs offline\. Appendix[P\.1](https://arxiv.org/html/2607.06973#A16.SS1)further discusses how to prevent test\-time leakage\.

Following the conventions in other popular forecasting benchmarks like GIFT\-Eval\(aksu2024gift\), we use normalized MASE \(mean absolute scaled error\) as our main metric\. Since the different variables have very different scales, we calculate the average MASE over all rolling windows of a variable and normalize that by the average MASE of a seasonal naive baseline\. Then we take the Geometric Mean \(GM\), for robustness to normalization choice, of these normalized MASE ratios across all variables\. See Appendix[N](https://arxiv.org/html/2607.06973#A14)\. We also compute the average MASE rank of each method over the variables\. For both metrics, smaller is better\.

Table 4:Overall benchmark results \(mean over 10 runs\) of the 13 selected methods\. The top 3 methods per metric are numbered in the parentheses\. The continuous ranked probability score is further used to reveal how context helps model uncertainty, in Appendix[U\.1](https://arxiv.org/html/2607.06973#A21.SS1)\.
### 4\.2Benchmarking Results

Table[4](https://arxiv.org/html/2607.06973#S4.T4)shows the benchmark results on TimesX \(details in Appendix[Y\.1](https://arxiv.org/html/2607.06973#A25.SS1)\)\. As multimodal solutions,*though the zero\-shot LLMs have the edge over the unimodal zero\-shot TFMs, this edge is not as significant on TimesX as observed on other synthetic benchmarks*\. This points out the gap between synthetic contexts and real world contexts, that some synthetic benchmarks tend to severely over\-estimate the performance of LLMs over TSF models because the synthetic contexts were providing the exact information required for bettering the forecast\. Real world contexts are in contrast too nuanced to always deliver a large performance boost even when guaranteed to be related\. This conclusion is further confirmed by the observation that*the agenticCodeRevmethod is performing worse*than its two components working alone, contradicting the claim of\(Williams\_2024\)\.

In terms of the composed solutions, to our surprise*the best performers on TimesX are the simple average ensembles*of different pretrained models \(i\.e\.,AvgEns\)\. We by no means want to suggest it as the optimal composition, but instead want to reiterate the stochastic and flexible nature of LLMs and the fact that it would take a great deal of effort to design the interaction with them in a composed solution to just outperform simple averaging ensembles\.

We also use the continuous ranked probability score \(CRPS\) metrics to measure uncertainty, detailed in Appendix[U\.1](https://arxiv.org/html/2607.06973#A21.SS1)\. We initially observe that*all three LLMs achieve lower CRPS than the TFMs*\. This suggests that our constructed context helps LLMs model future uncertainty better than TFMs, which only use numerical values\.

As an early experiment, we also evaluate reasoning LLMs, including GPT\-5, Gemini\-2\.5\-Flash, and DeepSeek\-R1\. To reduce data contamination, we only use evaluation samples after Jan 2025\. Table[17](https://arxiv.org/html/2607.06973#A22.T17)\(detailed in Appendix[Y\.4](https://arxiv.org/html/2607.06973#A25.SS4)\) shows that these*reasoning models do not provide a clear advantage*\.

Table 5:Gemini\-2\.0\-Flash MASE and MASE rank when provided with different context types\. Including all context types provides significant improvement over any other combinations\. See detailed results in Appendix[Y\.3](https://arxiv.org/html/2607.06973#A25.SS3)
### 4\.3Ablation: Context Types

Table[5](https://arxiv.org/html/2607.06973#S4.T5)presents the MASE and MASE rank when different combinations of contexts are provided to a chosen multimodal method \(Gemini\-2\.0\-Flash\)\. Comparing to only using the high level metadata, the inclusion of either calendar, calendar and covariates, or calendar and events brings significant improvement\. Using all context types improves the accuracy further eventually being around 16% better than using only static metadata\. It is worth noticing that we cannot observe the incremental gain from adding covariates or events alone\. We speculate there is crucial interaction between those two context types, for instance the effect of an event on the target variable can be quantified by a similar leading effect on a covariate\. Figure[4](https://arxiv.org/html/2607.06973#S4.F4)suggests this speculation may generalize to other LLMs\. This observation is aligned with our expectation that*a multimodal method can compound the gains by composing the information in different context types*\. Appendix[V](https://arxiv.org/html/2607.06973#A22.SS0.SSS0.Px1)reports domain\-level holiday and covariate effects\.

DomainMethodA&EC&EEconE\. TechFinP&APub\. H\.PPGSciShopSSSGTrafCropsEnergyLvstk\.RMCSAMSHVMCurrSeasonalNaive13131313131313131313131312131312111312Sundial10111211111210121011121091091091213Moirai\-2\.0121210101010111111121091445557TimesFM\-2\.5244641323714882881010AvgEns: TimesFM \+ Moirai9988876978662217688DeepSeek\-V3111011121211121012511117361213Gemini\-2\.0\-Flash6657998881886683434GPT\-4o7369755694574754142FuncRev: TimesFM \+ Gemini8893289759711112111313119CodeRev: TimesFM \+ Gemini577536756109121311121112911TextRev: TimesFM \+ Gemini3534542446331091091071AvgEns: TimesFM \+ GPT4122634322453136766AvgEns: TimesFM \+ Gemini1211121113226683436

Table 6:Breakdown of the MASE ranks by domain\. See Appendix[Z](https://arxiv.org/html/2607.06973#A26)for MASE\. Acronyms are used to shorten domain names \(see Appendix[I](https://arxiv.org/html/2607.06973#A9)\)\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x2.png)Figure 4:The compounded gains from all available context types are present for all LLMs\. See detailed results in Appendix[Y\.2](https://arxiv.org/html/2607.06973#A25.SS2)
### 4\.4Ablation: Domains

Table[6](https://arxiv.org/html/2607.06973#S4.T6)breaks down the ranking of the benchmark methods on each of the 19 domains \(see Appendix[Z](https://arxiv.org/html/2607.06973#A26)for MASE\)\. There is no dominance of one zero\-shot method over others \- in fact the orderings of TimesFM\-2\.5 and Gemini\-2\.0\-Flash are complementary\. In terms of composed solutions, the observation is consistent with Table[4](https://arxiv.org/html/2607.06973#S4.T4)thatAvgEnsperforms well while other compositions can be worse than its components acting alone\. We also observe two domain groups where multimodal solutions show a clear advantage over TFMs\. First, in Shopping, calendar information likely provides strong signals about when sales occur, which is hard to infer from the time series alone\. Second, in RMC \(Raw Materials & Constructions\), SAM \(Specialty & Advanced Materials\), SHVM \(Strategic & High\-Value Materials\), and Curr \(Currency\), the series are strongly affected by external policies, which are captured in the event context\.

### 4\.5Why Agentic \(Revision\) Methods Underperform?

None of the three revision methods \(FuncRev,CodeRev, andTextRev\) outperforms the simpleAvgEnson TimesX\. Figure[5](https://arxiv.org/html/2607.06973#S4.F5)shows that their forecast errors have a wider spread and more severe outliers than TimesFM\-2\.5, which leads to worse geometric means\. Gemini\-2\.0\-Flash alone shows a similar pattern\. We speculate that this is due to LLM stochasticity and the fact that LLM\-based revision does not reliably preserve the temporal structure in the initial forecast\. Among the three,TextRevperforms best, which suggests that effective code\-based revision may require instruction fine\-tuning\.

![Refer to caption](https://arxiv.org/html/2607.06973v1/x3.png)Figure 5:The boxplot of TimesFM\-2\.5, Gemini\-2\.0\-Flash and their composed method’s performances\.Table 7:Geometric mean MASE on the last three \(to avoid demo\-caused information leakage\) evaluation instances per variable\.
### 4\.6Future Solutions: Could Training Help?

This work focuses on zero\-shot methods\.To support future training\-based research, TimesX provides both training \(2018–2023\) and testing \(2023–2025\) splits\.

As a first step, weuse in\-context learning \(ICL\)brown2020languageas a probeto verify whether training\-based solutions, i\.e\., learning from historical instances, are promising\.

Specifically, we testTextRev\-ICL, which retrieves one temporally nearby historical example as a demonstration for eachTextRevinstance, andTextRev\-ICL\-C, which further adds an explicit conservative no\-change option to the prompt inspired by Figure[5](https://arxiv.org/html/2607.06973#S4.F5)which highlights the instability issue \(see Section[X](https://arxiv.org/html/2607.06973#A24)\)\. Table[7](https://arxiv.org/html/2607.06973#S4.T7)shows that both variants makeTextRevmore effective: they outperform Gemini\-2\.0\-Flash and TimesFM\-2\.5, and even surpass the best\-performingAvgEns\. Thus, we suggest that future training\-based methods consider effectiveness and stability\.

## 5Conclusions

We introduce a new multimodal time\-series forecasting benchmark, TimesX, which contains real\-world, cross\-domain time series with high\-quality, detailed textual context\. TimesX includes a dataset generation pipeline that keeps the benchmark leakage\-free via time isolation and enables automatic refresh via a dataset agent\. We conduct a detailed empirical study of zero\-shot multimodal TSF approaches on TimesX and find that earlier benchmarks either overestimate the performance of LLMs compared with TSF models or understate the importance of textual context for forecasting accuracy\. Our evaluation also shows that simple ensemble methods outperform seemingly stronger agentic solutions on TimesX\.

## Impact Statement

This paper presents work whose goal is to advance the field of machine learning\. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here\.

## Acknowledgment

This work was partly supported by the NSF \(Expeditions CCF\-1918770, CAREER IIS\-2028586, Medium IIS\-1955883, Medium IIS\-2403240, Medium IIS\-2106961\), NIH \(1R01HL184139\), Meta, Modal\.com and Dolby faculty gifts\.

## References

## Appendix

### Appendix Roadmap

This appendix is organized as follows\. We first discuss limitations and future directions, including low\-frequency coverage, bias, out\-of\-distribution variables, and future training\-oriented extensions\. We then provide dataset overview details, textual\-event structure, domain and variable breakdowns, construction details, leakage controls, evaluation protocols, and additional quantitative results\.

Table 8:Statistics of sparse\-event rare\-disease variables in TimesX\. “Events” is the number of constructed events per variable between 2023 and mid\-2025; “Avg\. summary len” is the average length of event descriptions; the remaining columns are time\-series characteristics computed as in Tables 11 and 15\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x4.png)Figure 6:Search\-trend signals for the five rare\-disease variables in TimesX\. Each panel shows a normalized trend series for one disease, where spikes often align with news or outbreak\-related events\.

## Appendix ALimitations and future directions

### A\.1Coverage of low\-frequency time series

The current release of TimesX focuses on daily and weekly frequencies\. There are two main reasons for this choice\. First, our goal is to perform leakage\-free evaluation using data strictly after the knowledge cutoffs of mainstream LLMs \(around June 2024\)\. Under this constraint, even if we collect monthly or quarterly data from July 2024 to August 2025 \(submission deadline\), there would be only about 13 monthly points or 4 quarterly points per variable, which is too short for a meaningful forecasting evaluation\.

Second, many real\-world monthly or quarterly indicators can be constructed by aggregating higher\-frequency series\. Users of TimesX can aggregate our daily or weekly variables into monthly or quarterly indicators when they wish to study lower\-frequency behavior\. This provides a practical way to analyze coarser temporal patterns on top of TimesX\.

Native monthly and quarterly series would further improve coverage\. We are actively searching for real\-world, regularly updated low\-frequency series that satisfy the same leakage\-free constraint, and we plan to add such variables in future versions of TimesX\.

### A\.2Coverage of multivariate time series

The current release focuses on multimodal univariate forecasting\. Homogeneous multivariate multimodal TSF, such as forecasting over multiple related sensors, is an important direction for future releases\.

### A\.3Coverage of more solutions

This work focuses on the evaluation of zero\-shot multimodal time series \(TS\) forecasting\. Accordingly, we benchmark four advanced zero\-shot TSFMs, six LLMs, and their architectural compositions, including ensembles and agents\. In contrast, works dedicated to TS understanding and reasoning operate under a distinct paradigm from our forecasting setting\. We recommend them as promising backbones for future fine\-tuning research and provide the performance of ChatTSChatTSas a representative reference\.

Future work could also consider exploring additional methodologies, such as visual tokenizationLiu\_picture, test\-time computation strategiesLiu\_eval, and training\-based fusion schemeswang2025chronosteer, and general event forecastingLiu\_Future\.

### A\.4Geographic and media bias

Event texts in the current version of TimesX are written in English\. We adopt English as a starting point, in line with common practice in NLP where a robust English setup is built first and then extended to multilingual settings\. At the same time, the numeric variables themselves already cover multiple regions, including North America \(United States, Canada\), Asia \(China, India\), Europe \(United Kingdom, Norway, Germany\), South America, and Africa\.

Geographic and media biases can still arise\. As shown in the workflow diagram in Figure[2](https://arxiv.org/html/2607.06973#S3.F2), our dataset construction agent already takes several steps to mitigate these biases\. In Steps 1b and 3b, the web search is configured with a global scope rather than a single country or region\. In Step 3b, the search engine retrieves information from multiple sources\. In Step 4a, the Synthesizer LLM is instructed to cross\-check evidence across sources and to give higher weight to authoritative outlets when producing the final event description\. These design choices help, but they cannot completely remove bias\.

In future versions of TimesX, we plan to make biases more transparent and more controllable by explicitly disentangling each stored event description into two parts: an “Objective Facts” segment that records dates, numbers, and events that have already occurred, and a “Subjective Analysis” segment that records market sentiment and speculative commentary\. This structured separation will allow users to focus on objective information when needed and to design more targeted robustness analyses\.

### A\.5Out\-of\-distribution evaluation: multilingual and sparse\-event variables

To explore out\-of\-distribution \(OOD\) generalization, we extend TimesX with two types of additional variables that are not part of the main in\-distribution set: multilingual variables and sparse\-event rare\-disease variables\.

##### Multilingual variables\.

The main release of TimesX uses English event texts; multilingual and region\-specific events are also important\. To this end, we construct 11 new non\-English variables that span five continents and cover the following languages: Afrikaans, French, German, Hindi, Japanese, Korean, Portuguese, Simplified Chinese, Spanish, Swahili, and Turkish\. Each variable focuses on region\-specific topics in the corresponding language\. These multilingual variables are available in an extended split of TimesX\. We reserve them for OOD evaluation rather than including them in the main in\-distribution benchmark\.

##### Sparse\-event rare\-disease variables\.

We also construct five rare\-disease variables to study sparse\-event settings: Chagas disease, Huntington’s disease, Guinea worm disease, Marburg virus disease, and Nipah virus\. For these variables, the numeric component currently uses search\-trend signals; we are actively looking for stable, regularly updated sources of case counts or incidence rates\.

Figure[6](https://arxiv.org/html/2607.06973#Ax1.F6)visualizes the search\-trend signals for these five rare\-disease variables\. Even though the events are sparse, the series still show noticeable spikes around news or outbreak\-related periods\.

We further compute statistics over the constructed events for these rare\-disease variables\. Table[8](https://arxiv.org/html/2607.06973#Ax1.T8)reports the number of events, the average event summary length, and several time\-series characteristics\. Compared with typical variables in TimesX \(see Tables 11 and 15\), the rare\-disease variables indeed have fewer events, but still more than 30 events per disease between 2023 and mid\-2025 on average\. This suggests that our dataset agent remains effective even in sparse\-event domains and can recover a reasonable amount of external context\.

### A\.6Support for finetuning and future exploration

The main focus of this paper is to address the lack of an appropriate multimodal TSF evaluation benchmark\. We see this as a key bottleneck before fully exploring finetuning strategies\. At the same time, we aim to make TimesX a useful testbed for future work on finetuning TFMs and LLMs\.

To support finetuning, we construct an additional dataset that covers the years 2018–2022 for the same 190 in\-distribution variables, including both numeric series and textual contexts\. This split is intended as a training resource, while the 2023–2025 split serves as the main evaluation period\.

We have observed that adding a single ICL example already improves the performance in Table[7](https://arxiv.org/html/2607.06973#S4.T7)\. This suggests that finetuning or more advanced adaptation schemes on TimesX could be promising\. A thorough study of finetuning TFMs and LLMs on TimesX would, however, require substantial additional effort in terms of computation and method design \(including tokenization, alignment strategies, and loss functions\), which is beyond the scope of this dataset\-and\-benchmark paper\. We hope that the finetuning split, the refreshable leakage\-free evaluation set, and our empirical findings can serve as a solid foundation for future work on finetuning models for multimodal time\-series forecasting\.

## Appendix BDetailed Experiment Setup of Table[2](https://arxiv.org/html/2607.06973#S3.T2)

For evaluation cost consideration, we conduct the data leakage validation only on the 121 variables in theSearchTrendsubset of TimesX\. All other experimental settings follow Section[4\.1](https://arxiv.org/html/2607.06973#S4.SS1)\.

## Appendix CDetails on Metadata and Calendar Context Construction

### C\.1Metadata Construction

For each variable in TimesX, we create a static metadata description that summarizes the essential attributes of the time series\. The metadata is generated using a fixed template with three components: \(1\) the variable name and its measurement unit, \(2\) the domain the variable belongs to, and \(3\) the collection frequency and the target prediction window\.

The general template is as follows:

> “Meta Info": "This time series records \[*variable name and unit*\] in the \[*domain*\] domain, with a collection frequency of \[*frequency*\]\. Prediction target period: from \[*start date*\] to \[*end date*\]\."

For example, for the gasoline price series in the Commodity Price domain, the metadata is:

> “Meta Info": "This time series records gasoline price \(USD/Gal\) in the Commodity Price domain, with a collection frequency of daily\. Prediction target period: from 2024\-09\-01 to 2024\-09\-15\."

### C\.2Calendar Context Construction

We generate calendar\-based context features by automatically identifying holidays and special dates that fall within the forecasting horizon\. Specifically, we use thePython Holidayslibrary444[https://pypi\.org/project/holidays/](https://pypi.org/project/holidays/)to retrieve country\- and region\-specific holidays\.

The construction process follows three steps:

1. 1\.Convert each time series into a sequence of\(timestamp,value\)\(\\text\{timestamp\},\\text\{value\}\)pairs\.
2. 2\.For each forecasting horizon, query the library to extract all holidays that overlap with the horizon window\.
3. 3\.Format the results into textual annotations describing the holiday names and dates, which are then aligned with the corresponding timestamps\.

For example, if the prediction horizon is from 2024\-09\-01 to 2024\-09\-15, the generated calendar context includes:

> “Upcoming holidays in the prediction window: Labor Day \(2024\-09\-02\)\."

## Appendix DCovariate Context Construction

To generate covariate\-based textual contexts, we compute descriptive statistics for each covariate in the same domain as the target series\. Specifically, for each covariate we calculate:

- •The average and median value over the observation window\.
- •The maximum value and the corresponding date\.
- •The minimum value and the corresponding date\.
- •The overall trend \(upward or downward\)\.

These statistics are automatically extracted using simple Python scripts and converted into structured textual descriptions\. For example, one covariate may be described as:

> “from 2024\-06\-01 to 2024\-08\-31: The average value ofBrent Crude Oil \(USD/BBL\)is 81\.5\. The maximum value of 87\.4 occurred on 2024\-07\-04, and the minimum value of 76\.0 occurred on 2024\-08\-21, showing an overall downward trend\.”

### D\.1Missing\-value handling

We handle missing values separately for the input and target windows\. For target\-window missing values, evaluation metrics are computed only on observed timestamps\. For input\-window missing values, short gaps are linearly interpolated for time\-series foundation models\. In the current data, the average maximum contiguous missing gap per sample is 0\.7962\. We also remove 3 anomalous samples whose maximum contiguous gap exceeds 6\.

## Appendix EDetails of Textual Event Construction

Execution log demo is provided in Section[F](https://arxiv.org/html/2607.06973#A6)

##### Stage 1: Time\-Series\-Aware Event Hypothesis Generation

Input\.A target variable with its numeric seriesy1:Ty\_\{1:T\}and a target period\[ts,te\]\[t\_\{s\},t\_\{e\}\]\. The period is partitioned into fixed\-length time blocksℬ=\{B1,…,BM\}\\mathcal\{B\}=\\\{B\_\{1\},\\dots,B\_\{M\}\\\}\(e\.g\., week or month\)\.

Process\.For each blockBB, an LLM proposes an initial set of event hypothesesHB=\{h1,…\}H\_\{B\}=\\\{h\_\{1\},\\dots\\\}\. Eachhhcontains a tentative title, a draft timestamp, involved entities, and at least one candidate source URL\. We aim to ensure through multiple iterations thatHBH\_\{B\}sufficiently explains the prominent movements inyty\_\{t\}withinBB\. Specifically, we define a peak set𝒫B\\mathcal\{P\}\_\{B\}fromyty\_\{t\}\(by a standard peak detector with fixed hyperparameters\)\. A peakp∈𝒫Bp\\in\\mathcal\{P\}\_\{B\}is*covered*if at least oneh∈HBh\\in H\_\{B\}is temporally aligned withpp\(within a small window\) and topically relevant to the target variable\. The coverage is

cov​\(HB\)=\|\{p∈𝒫B:p​is covered by​HB\}\|\|𝒫B\|\.\\mathrm\{cov\}\(H\_\{B\}\)\\;=\\;\\frac\{\|\\\{p\\in\\mathcal\{P\}\_\{B\}:p\\text\{ is covered by \}H\_\{B\}\\\}\|\}\{\|\\mathcal\{P\}\_\{B\}\|\}\.We keep querying the LLM to add hypotheses iteratively and stop whencov​\(HB\)≥θ\\mathrm\{cov\}\(H\_\{B\}\)\\geq\\thetaor when a step limitKmaxK\_\{\\max\}is reached\. This rule balances completeness and cost without relying on unrestricted search\.

In our ablation, this peak\-based heuristic reduces construction cost by about 21%\.

Output\.For each blockBB, a hypothesis setHBH\_\{B\}with draft timestamps, entities, and seed URLs\. All retrieval during Stage 1 is time\-bounded by\[ts,te\]\[t\_\{s\},t\_\{e\}\]to keep the search window consistent with the evaluation period and to avoid leakage\.

Leakage control\.All queries use an explicit upper boundtet\_\{e\}\. Sources with edited or republished pages aftertet\_\{e\}are kept only if the original publication date is within\[ts,te\]\[t\_\{s\},t\_\{e\}\]and the content is accessible in that state\.

Prompt\. The prompt for this role is detailed in Fig[7](https://arxiv.org/html/2607.06973#A5.F7)

##### Stage 2: Rigorous Verification and Temporal Characterization

Input\.Hypotheses\{HB\}\\\{H\_\{B\}\\\}from Stage 1\.

Process\.A verification role calls LLMs to re\-fetch evidence under the same time bound\[ts,te\]\[t\_\{s\},t\_\{e\}\]and constructs a structured temporal view for each hypothesis\. The verifier normalizes titles, resolves canonical entities, and extracts two dates:announcement\_dateandoccurrence\_date\. It then assigns a temporal type from a closed set:*Scheduled*\(announced in advance\),*Contemporaneous*\(announcement and occurrence are near in time\),*Retrospective*\(backward\-looking report\),*Predictive*\(forward\-looking signal\), or*Mixed*\. For each atomic claim, the verifier requires an accessible source URL \(HTTP 200 at crawl time\), stores the access date, and records a short quote that supports the extracted field\. Multi\-source cross\-checking removes items with unresolved contradictions and de\-duplicates near\-duplicates by normalized title, entity set, and date tuple\.

Output\.A verified event set with\(announcement\_date,occurrence\_date,type\)\(\\text\{announcement\\\_date\},\\text\{occurrence\\\_date\},\\text\{type\}\), consolidated sources, and a confidence score that reflects agreement across sources and the precision of dates \(day\-level preferred over month\-level\)\. The final timestamp and type come with a concise rationale that explains corrections to the Stage 1 draft\.

Leakage control\.Verification refuses any evidence whose first\-publication date is aftertet\_\{e\}\. If a page is updated aftertet\_\{e\}but preserves the original content and date within\[ts,te\]\[t\_\{s\},t\_\{e\}\], the verifier keeps the archived or cited original\. Otherwise the evidence is discarded\.

Prompt\. The prompt for this role is detailed in Fig[8](https://arxiv.org/html/2607.06973#A5.F8)

##### Stage 3: Conditional Enrichment for Narrative Depth

Input\.Verified events from Stage 2\.

Process\.An evaluation role checks information sufficiency for forecasting\. If key fields are missing \(e\.g\., actors, locations, magnitudes, explicit dates, or links to related variables\), the pipeline runs an iterative but bounded deep search under the same time bound\[ts,te\]\[t\_\{s\},t\_\{e\}\]\. Each step adds at most one new high\-value source\. The process stops when all required fields are present or when a small step capLmaxL\_\{\\max\}is reached\. The reporting role then writes a concise narrative \(a few sentences\) that states what happened, when, who is involved, and why it likely relates to the target variable\. Each factual sentence is grounded by one or more quotes with URLs\.

Prompt\. The prompt for this role is detailed in Fig[9](https://arxiv.org/html/2607.06973#A5.F9)\.

The prompt for the final synthesizer role is detailed in[10](https://arxiv.org/html/2607.06973#A5.F10)

Across the construction pipeline, the Synthesizer discards about 32\.20% of unverified candidate events\.

Note that, in our configuration, we empirically setKmax=3K\_\{\\max\}=3,θ=90%\\theta=90\\%, andLmax=3L\_\{\\max\}=3\. Under this setting, the system can efficiently construct high\-quality corpora with reasonable runtime and cost\. We also conduct a small\-scale sensitivity test: when we increaseKmaxK\_\{\\max\}from 3 to 5 on a subset of variables, the total number of accepted events increases by only about2%2\\%\. Overall, we encourage users to adjust these hyperparameters according to their budget and domain, while using our configuration as a default recommendation\.

Under our current configuration \(Gemini 2\.5 Pro plus Gemini 2\.5 Flash\), the construction cost is about $0\.7 per variable per three\-month time block, including reruns due to network errors\. This cost can be further reduced by using open\-source LLMs or by batching verification steps so that multiple candidate events share the same LLM calls\.

[⬇](data:text/plain;base64,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)prompt=f"""Youareaprofessionalresearchanalystspecializinginthe\{domain\}field\.Yourtaskistouseyourwebsearchcapabilitiestoidentifyandstructuresignificanteventsrelatedto’\{keyword\}’in\{geography\_str\}thatoccurredbetween\{start\_date\}and\{end\_date\}\.\*\*KeyGuidelines:\*\*1\.\*\*SourceofInformation:\*\*Pleasebaseyourresponsesontheinformationretrievedfromyourwebsearchandgeneralcommonsense\.Itisimportanttoavoidrelyingoninternalknowledgeorgeneratingspeculativedetails\(hallucinations\)\.2\.\*\*DateExtractionPrinciples:\*\*Itishelpfultodistinguishbetweentwokeytypesofdates\.Ifadatecannotbefoundfromthesources,pleaseuse‘null‘\.\*‘announcement\_date‘:Thedatewhenthenewsabouttheeventwas\*\*publishedorfirstannounced\*\*\.Forexample,ifanewsarticlefrom\*\*2024\-01\-10\*\*announcesanupcomingproductlaunch\.\*‘occurrence\_date‘:Thedatewhentheevent\*\*actuallytookplaceorisscheduledtotakeplace\*\*\.Forexample,iftheproductlaunchmentionedabovehappenson\*\*2024\-02\-02\*\*\.3\.\*\*EventTypeClassification:\*\*Pleaseclassifytheeventintothefollowingtypesbasedonitscertaintyandtiming\.\*‘ScheduledEvent‘:Ahigh\-certaintyeventthathasbeenofficiallyannouncedtooccuratafuturedate\.\*‘PredictiveInformation‘:Alower\-certaintypieceofinformationaboutthefuture,suchasananalystforecast,atargetpricechange,oracrediblerumor\.\*‘ContemporaneousEvent‘:Aneventthatoccursatthesametimeitisannounced,oftenunexpected\.\*‘RetrospectiveReport‘:Ananalysisorreportaboutaneventorperiodthathasalreadypassed\.4\.\*\*GeographicFocus:\*\*Pleasefocusoneventswithinthe\{geography\_str\}region\.5\.\*\*SourceVerifiability:\*\*Eacheventshouldbesupportedbyatleastoneverifiable,high\-qualityURL\.\*\*OutputFormat:\*\*PleaseprovideyourresponseasaJSONarrayofeventobjects\.Eachobjectinthearrayshouldconformtothefollowingstructure\.Ifanyfield’svaluecannotbedeterminedfromthesources,use‘null‘\.‘‘‘json\[\{\{"event\_summary":"Aconcise,factualdescriptionoftheevent\.","announcement\_date":"YYYY\-MM\-DDornull","occurrence\_date":"YYYY\-MM\-DDornull","event\_type":"ScheduledEvent\|PredictiveInformation\|ContemporaneousEvent\|RetrospectiveReportorMixedornull","source\_urls":\["url1","url2"\],"confidence\_score":0\.8\}\}\]‘‘‘PleaseensuretheentireresponseisonlythevalidJSONarray,withoutanysurroundingtextorexplanations\."""

Figure 7:Prompt used for Hypothesizer Role: Initial Event Discovery via LLM with Web Search Integration\.[⬇](data:text/plain;base64,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)prompt=f"""Youareameticulousfact\-checker\.Yourtaskistoverifyclaimsagainstsourcecontentandidentifyinvalidpages\.\*\*CRITICAL:Firstdetermineifthesourcecontentisvalid\.\*\*\*\*ImportantDefinitions:\*\*\-\*\*content\_date\*\*:Thedatewhentheeventitselfoccurred\(e\.g\.,productlaunch,announcement\)\-\*\*publish\_date\*\*:Theoriginalpublicationdateofthesource\(NOT"updated"or"lastmodified"dates\)\*\*Few\-ShotLearningExamples:\*\*\*\*Example1\-ValidContent:\*\*EventClaim:"AppleVisionProwillbeavailableonFebruary2,2024"SourceText:‘‘‘<h2\>AppleVisionProAvailableintheU\.S\.onFebruary2</h2\><spanclass="publish\-date"\>POSTEDONJANUARY8,2024</span\><p\>AppletodayannouncedAppleVisionProwillbeavailablebeginningFriday,February2\.\.\.</p\><footer\>Lastupdated:January10,2024</footer\>‘‘‘ExpectedOutput:‘‘‘json\{\{"page\_status":"valid\_content","verified\_statements":\[\{\{"statement":"AppleVisionProwillbeavailableonFebruary2,2024","status":"Confirmed","supporting\_quote":"AppleVisionProwillbeavailablebeginningFriday,February2"\}\}\],"overall\_timing":\{\{"content\_date":"2024\-02\-02","publish\_date":"2024\-01\-08"\}\},"reasoning":"Validpressreleasecontent\.Usedoriginalpublishdate\(Jan8\),ignored’lastupdated’footer\."\}\}‘‘‘\*\*Example2\-404ErrorPage:\*\*EventClaim:"iPhone16rumorssurfaceinMarch2024"SourceText:‘‘‘<title\>PageNotFound\-TechNews</title\><h1\>404\-PageNotFound</h1\><p\>Thepageyou’relookingfordoesn’texist\.</p\><divclass="sidebar"\>Today’sHotTopics:July28,2025</div\>‘‘‘ExpectedOutput:‘‘‘json\{\{"page\_status":"error\_page\_404","verified\_statements":\[\],"overall\_timing":\{\{"content\_date":null,"publish\_date":null\}\},"reasoning":"Thisisa404errorpagewithnovalidcontent\.Sidebardatesareirrelevanttemplatecontent\."\}\}‘‘‘\*\*Nowanalyzetheactualcontent:\*\*\*\*1\.TheEventClaimtoAnalyze:\*\*\{event\_claim\}\*\*2\.TheEvidence\(SourceText\):\*\*\-\-\-\{source\_evidence\}\-\-\-\*\*Instructions:\*\*1\.First,determinepage\_status:valid\_content,error\_page\_404,access\_denied,orlogin\_wall2\.Ifpage\_statusisNOT"valid\_content",returnemptyverified\_statementsandnulldates3\.Ifvalid\_content,decomposeclaimintoatomicfactsandverifyeachone4\.Fordates:UseORIGINALpublishdates,ignore"updated","modified",orsidebardates5\.Classifyfactstatus:Confirmed,Anticipated,Speculation,orNot\_Found\*\*JSONOutputformat:\*\*\{\{"page\_status":"<valid\_content\|error\_page\_404\|access\_denied\|login\_wall\>","verified\_statements":\[\{\{"statement":"<atomicfactualstatement\>","status":"<Confirmed\|Anticipated\|Speculation\|Not\_Found\>","supporting\_quote":"<exactquotefromtextornull\>"\}\}\],"overall\_timing":\{\{"content\_date":"<YYYY\-MM\-DDwhentheeventoccurredornull\>","publish\_date":"<YYYY\-MM\-DDwhensourcewasoriginallypublishedornull\>"\}\},"reasoning":"<Briefexplanationofyouranalysisprocess\>"\}\}RespondwithONLYtheJSONobject,noadditionaltext\."""

Figure 8:Prompt used for Verifier Role: Atomic Fact Verification with Evidence Matching\.[⬇](data:text/plain;base64,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)prompt=f"""Youarearesearchstrategistanalyzinginformationgapsandplanningnextsteps\.\*\*IMPORTANTACTIONLIMIT\*\*:Pleaselimityournext\_actionstoamaximumof\{max\_actions\}items\.Focusonthemostcriticalinformationgapsthatneedtobeaddressed\.Eachactionshouldbehigh\-qualityandtargeted\.\*\*OriginalEventClaim:\*\*\{original\_claim\}\*\*CurrentlyVerifiedInformation:\*\*\{statements\_summary\}\*\*YourTask:\*\*1\.\*\*Analyzecompleteness\*\*:Compareverifiedinformationagainsttheoriginalclaim2\.\*\*Identifyinformationgaps\*\*:Listmissingorunconfirmedfacts3\.\*\*Plannextactions\*\*:Foreachgap,determineifit’scommonknowledgeorrequiressearchForeachinformationgap,classifyas:\-\*\*Commonknowledge\*\*:Factsthatcanberesolvedinternally\(e\.g\.,"Apple’sfiscalQ1isOct\-Dec"\)\-\*\*Requiressearch\*\*:Factsneedingexternalverification\*\*JSONOutputformat:\*\*\{\{"is\_sufficient":<trueifallkeyfactsareconfirmed,falseotherwise\>,"next\_actions":\[\{\{"info\_gap":"<descriptionofmissinginformation\>","is\_common\_knowledge":<true\|false\>,"action\_type":"<resolve\_internally\|search\>","resolved\_answer":"<answerifcommonknowledge,nullotherwise\>","query":"<searchqueryifaction\_typeissearch,nullotherwise\>"\}\}\]\}\}PleaserespondwithONLYtheJSONobject\."""

Figure 9:Prompt used for Enricher Role \(Phase 1\): Information Sufficiency Evaluation and Action Planning\.[⬇](data:text/plain;base64,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)prompt=f"""YouareanInformationIntegrationSpecialist\.Yourtaskistoproducethefinal,authoritativeversionofaneventbyreviewingallprovidedevidence\.Yoursummarymust:\-Correctandenrichtheoriginalclaimwithadditionalverifieddetails\.\-Forfactualorscheduledevents,prioritizethemostauthoritativesources\.\-Forsubjectiveanalysesorpredictions,explicitlyincludemultiplecredibleviewpoints,ifavailable\.Clearlyacknowledginganyconflictingoruncertainclaimsalongwiththeirsources,ifavailable\-Accuratelyadjudicateorrevisetheevent’sannouncementdate\(thedatethenewswaspublished\)andoccurrencedate\(theactualdateoftheevent\),usingwebsearchifnecessarytoensureaccuracy\.\-Ifyouareunsure,useNAandavoidmakingupinformation\.\*\*1\.OriginalEventClaim:\*\*\{event\_summary\}\*\*2\.DetailedFactualEvidence:\*\*\*\*ConfirmedFacts:\*\*\{confirmed\_facts\}\*\*Anticipated/PlannedFacts:\*\*\{anticipated\_facts\}\*\*InternalKnowledgeResolutions:\*\*\{internal\_facts\}\*\*3\.DetailedTimingEvidencefromSources:\*\*\{timing\_summary\_detailed\}\*\*4\.InitialDate:\*\*Thefollowingdatesarepreliminaryfindingsanddonotrepresent100%accuracy\.Pleaseselectthemostreasonabledatebasedonthecontentanduseonlinesearchtoolsifnecessary\.\-\*\*AllPublishDatesFound:\*\*\{unique\_publish\_dates\}\-\*\*AllContentDatesFound:\*\*\{unique\_content\_dates\}\*\*YourFinalTask:\*\*RespondwithONLYasingleJSONobject\.Donotaddanytext,explanations,ormarkdownformattingbeforeoraftertheJSONblock\.\*\*JSONOutputFormat:\*\*\{\{"final\_summary\_text":"<Yourcomprehensivesummaryhere\.Thistextshouldbewell\-written,accurate,detailedandreflectyourfinaldecisiononthedates\.\>","authoritative\_dates":\{\{"announcement\_date":"<Thesingle,mostcredibleYYYY\-MM\-DDpublishdateofcontent\.Ifnone,useNA\.\>","occurrence\_date":"<Thesingle,mostcredibleYYYY\-MM\-DDdatewhentheeventactuallytookplace\.Ifnone,useNA\.\>"\}\},"reasoning\_for\_date\_choice":"<Abrief,one\-sentenceexplanationforyourdateselection\.e\.g\.,’Chosetheearliestpublishdatefromaprimarynewssource\.’\>"\}\}"""

Figure 10:Prompt used for Synthesizer Role: Final Information Synthesis with Authoritative Date Determination\.

## Appendix FDetailed Execution Log: Cotton Price Case Study

To further elucidate the leakage prevention and verification mechanisms of the Hypothesizer\-Verifier\-Enricher \(H\-V\-E\) framework, we present a detailed step\-by\-step execution log for the variable “cotton price” over the historical window from 2024\-01\-01 to 2024\-03\-31\.

#### Phase 1: Hypothesizer

- •Input:Keyword: “cotton price”; Time Range: 2024\-01\-01 to 2024\-03\-31\.
- •Action 1 \(Peak Discovery\):The system analyzed the time\-series data and identified 10 significant peaks requiring explanation \(e\.g\., 2024\-01\-08, 2024\-02\-19, etc\.\)\.
- •Action 2 \(Event Discovery\):The LLM \(Gemini\-2\.5\-Pro\) performed an initial search and generated 14 candidate events\. Examples include: - –Event 66e0: “The USDA’s January 2024 WASDE report…” \(Source: cottongrower\.com/…\) - –Event 57e9: “The USDA’s weekly export sales report…” \(Source: ccfgroup\.com/…\) - –Event a7a7: “An early 2024 report highlighted… Panama Canal…” \(Source: terrain\.sc\.eg/…\)

#### Phase 2 & 3: Verifier & Enricher \(Parallel Streams\)

The system spawned 14 independent verification tasks for the candidate events\. We illustrate the robustness of the pipeline using three representative tasks:

##### Task 1: Successful Verification \(Event 66e0 \- USDA WASDE Report\)

- •Verify:Crawler accessedcottongrower\.com/\.\.\.\(Status:Success\)\.
- •Verify:Theclaim\_verifieragent \(Gemini\-2\.5\-Flash\) cross\-referenced the crawled content with the claim and confirmed consistency\.
- •Enrich \(Evaluate\):Theinfo\_sufficiency\_evaluatordeemed the information sufficient\.
- •Finalize:Thefinal\_synthesisagent generated the final summary\.
- •Output:StatusVERIFIED\.

##### Task 2: Verification Failure and Discard \(Event 57e9 \- USDA Weekly Sales\)

- •Verify:Crawler attempted to accessccfgroup\.com/\.\.\.\(Status: Failed, error\_page\_404\)\.
- •Enrich \(Evaluate\):Evaluator deemed information insufficient, triggering an Action Plan\.
- •Enrich \(Act\):Agent generated a new query: “cotton price official announcement”\.
- •Enrich \(Act \- Leakage Prevention\):The system executed a time\-bounded search\. > INFO \- Executing Date\-Restricted Search: 2024\-01\-05 to 2024\-01\-19 INFO \- \[SEARCH DEBUG\] tbs: cdr:1,cd\_min:01/05/2024,cd\_max:01/19/2024
- •Verify \(Loop 2\):System crawled new URLs \(e\.g\., usda\.gov\), but theclaim\_verifiercould not verify the specific statistics \(“262,500 running bales”\) from the new sources\.
- •Finalize:Verification failed after max attempts\.
- •Output:StatusUNVERIFIED\. The event was discarded\.

##### Task 3: Recovery via Enrichment \(Event a7a7 \- Panama Canal\)

- •Verify:Crawler attempted to accessterrain\.sc\.eg/\.\.\.\(Status: Failed, net::ERR\_NAME\_NOT\_RESOLVED\)\.
- •Enrich \(Evaluate\):Evaluator deemed information insufficient\.
- •Enrich \(Act\):Agent generated a new query: “cotton price… Panama Canal… details”\.
- •Enrich \(Act \- Leakage Prevention\):The system executed a time\-bounded search to find alternative sources\. > INFO \- Executing Date\-Restricted Search: 2024\-01\-03 to 2024\-01\-17 INFO \- \[SEARCH DEBUG\] tbs: cdr:1,cd\_min:01/03/2024,cd\_max:01/17/2024
- •Enrich \(Act\):Search successfully retrieved valid new URLs \(e\.g\.,windward\.ai/\.\.\.andporttechnology\.org/\.\.\.\)\.
- •Verify \(Loop 2\):Crawler accessedwindward\.ai/\.\.\.\(Status:Success\)\.
- •Verify \(Loop 2\):Theclaim\_verifiersuccessfully verified the facts regarding the Panama Canal drought impact\.
- •Finalize:Thefinal\_synthesisagent generated the final summary using the verified facts from the new source\.
- •Output:StatusVERIFIED\.

## Appendix GDetails on Time Isolation and Event Timestamp Verification

Addressing the risk of sample\-level information leakage—specifically, the inclusion of future events within the historical window—is a critical challenge in constructing time\-series datasets\. As stated in Section 4\.1, our evaluation protocol strictly incorporates textual events with timestamps that precede the prediction window\. Consequently, the integrity of our benchmark hinges on the Dataset Agent’s capability to accurately annotate event timestamps\.

Our Dataset Agent does not operate as a simple, unrestricted web search tool, but rather as a rigorous, multi\-stage verification pipeline governed by the Hypothesizer\-Verifier\-Enricher framework \(see Figure[2](https://arxiv.org/html/2607.06973#S3.F2)\)\. To mitigate hallucinations and prevent timestamp errors, we design a three\-tier correction mechanism:

- •Verifier \(Phase 2a\-2b\):Following the initial event list generation by LLM A in the Hypothesizer phase, the Verifier employs a web crawler to fetch content from specific URLs\. It then utilizes LLM B to strictly validate whether the dates within the webpage content align with the proposed event, thereby filtering out discrepancies\.
- •Enricher \(Phase 3\):For details that the Verifier cannot conclusively confirm, the Enricher formulates search queries\. Crucially, in Phase 3b, we enforce hard constraints on the search engine API \(e\.g\., thetbsparameter in the Google Search API\), restricting results to a time window of±k\\pm kdays around the candidate event date\. This API\-level constraint provides a strong guarantee for timestamp validity\.
- •Synthesizer \(Phase 4\):LLM E aggregates all verified evidence to make a final adjudication\. Any event that fails to achieve cross\-source verification regarding its date is rejected\.

### G\.1Manual Audit of Timestamp Accuracy

To empirically validate this mechanism, we conducted a manual audit on 50 randomly sampled events generated by the pipeline\.

- •In 47 cases \(94%94\\%\), the automatically annotated timestamps matched the human annotation exactly\.
- •In 2 cases \(4%4\\%\), the agent’s timestamp was 1–2 days later than the human annotation\. This represents a conservative error that does not constitute leakage\.
- •Only 1 case \(2%2\\%\) was dated 1 day earlier than the human annotation\. Upon inspection, this discrepancy arose from ambiguity in defining the date for a complex event involving multiple sequential developments\.

### G\.2Case Study: Correcting Reporting Lag

We observe that the agent is capable of reducing reporting lag while maintaining strict leakage prevention\. We present a representative case study regarding Medicare drug costs to demonstrate this capability:

- •
- •Agent Action:The agent traces the primary source cited within the news report\.
- •Correction:The agent successfully retrieves the original JAMA Health Forum article published on February 14 and corrects the event timestamp from February 18 to February 14\.

This precision ensures that the event is correctly aligned with the historical window, capturing the earliest valid signal without violating temporal causality\.

## Appendix HOverview of TimesX

TimesX contains 20 domains and 200 variables in total \(balanced design:20×1020\\times 10\)\. The collection window spans from2022\-01\-01to2025\-06\-30for the daily subset, and from2023\-01\-01to2025\-06\-30for the weekly subset\.666Weekly Google Trends series are included from 2022\-01\-01 due to stable availability and consistent retrieval settings\.Geographical coverage includes North America \(United States, Canada, Mexico\), Asia \(for example, China, India\), Europe \(for example, United Kingdom, Norway\), South America \(for example, Brazil\), and Africa\.

Table 9:Dataset summary\.##### Variables and Frequencies\.

Weeklyseries consist of Google Search Trend signals888[https://trends\.google\.com/trends/](https://trends.google.com/trends/)across 12 domains\.Dailyseries include \(i\) commodity prices999[https://marketstack\.com/](https://marketstack.com/)covering raw materials, energy, metals, and agriculture, and \(ii\) major USD exchange rates101010[https://frankfurter\.dev/](https://frankfurter.dev/)\. All series are aligned to a unified calendar per frequency, with clear missing\-value handling policies documented in the dataset card\.

### H\.1Structure of Textual Events

To maximize scientific utility and trust, we choose two complementary representations of the events: astructured event corpusfor modeling and thecomplete verification logsfor audit\. The former provides clean, time\-aligned annotations with compact narratives\. The latter records the search queries, source URLs, access timestamps, and short evidence quotes produced by the verifier\.

Each event is organized into two layers that match modeling needs and evidence needs:

##### Core semantics for modeling\.

A short, fact\-checked narrative, distinct*announcement*and*occurrence*dates, and a categorical*event type*\(*Scheduled*,*Contemporaneous*,*Retrospective*,*Predictive*, or*Mixed*\)\.

##### Evidence and provenance\.

A small set of independent sources that support each claim, with verbatim text snippets and the corresponding access timestamps\. For pages updated after the evaluation cut\-off, archived versions or original publication records are linked\.

## Appendix IBreakdown of TimesX by Domains and Variables

Here are the acronyms we used for each domain when applicable:

- •A&E: Arts & Entertainment;
- •C&E: Climate & Environment;
- •Econ: Economy;
- •E\. Tech: Electronic Technology;
- •Fin: Finance;
- •P&A: Pets & Animals;
- •Pub\. H\.: Public Health;
- •PPG: Public Policy & Governance;
- •Sci: Science;
- •Shop: Shopping;
- •SSSG: Society Security & Social Good;
- •Traf: Traffic;
- •Crops: Crops & Staples;
- •Energy: Energy & Fuels;
- •Lvstk\.: Livestock & Food Products;
- •RMC: Raw Materials & Construction;
- •SAM: Specialty & Advanced Materials;
- •SHVM: Strategic & High\-Value Materials;
- •Curr: Currency\.

Tables[10](https://arxiv.org/html/2607.06973#A9.T10),[11](https://arxiv.org/html/2607.06973#A9.T11)and[12](https://arxiv.org/html/2607.06973#A9.T12)list the variables under each of the 19 domains in TimesX\.

Table 10:Search Trend \(weekly\) coverage by domain and variables \(2023\-01\-01–2025\-06\-30\)\. Each domain lists ten representative keywords\.Table 11:Daily dataset coverage by domain and variables \(2022\-01\-01–2025\-06\-30\)\. Each domain lists ten representative instruments\.DomainVariables \(Exchange Rate\)CurrencyUSDtoAUD\-ExchangeRate; USDtoBRL\-ExchangeRate; USDtoCAD\-ExchangeRate; USDtoCHF\-ExchangeRate; USDtoGBP\-ExchangeRate; USDtoHKD\-ExchangeRate; USDtoINR\-ExchangeRate; USDtoKRW\-ExchangeRate; USDtoMXN\-ExchangeRate; USDtoSGD\-ExchangeRate

Table 12:ExchangeRate: domains and variables
## Appendix JFeature Definition of Time Series

Let a univariate series be\{xt\}t=1T\\\{x\_\{t\}\\\}\_\{t=1\}^\{T\}\. We decompose it with STL into trendTtT\_\{t\}, seasonal componentStS\_\{t\}, and remainderRtR\_\{t\}:

xt=Tt\+St\+Rt\.x\_\{t\}\\;=\\;T\_\{t\}\\;\+\\;S\_\{t\}\\;\+\\;R\_\{t\}\.\(1\)Define the de\-trended seriesxtdetr=xt−Ttx^\{\\text\{detr\}\}\_\{t\}=x\_\{t\}\-T\_\{t\}and the de\-seasonalized seriesxtdeseas=xt−Stx^\{\\text\{deseas\}\}\_\{t\}=x\_\{t\}\-S\_\{t\}\.

##### Seasonality

Seasonality=max⁡\(0,1−Var​\(Rt\)Var​\(xtdetr\)\)\.\\mathrm\{Seasonality\}\\;=\\;\\max\\\!\\left\(0,\\;1\\;\-\\;\\frac\{\\mathrm\{Var\}\(R\_\{t\}\)\}\{\\mathrm\{Var\}\(x^\{\\text\{detr\}\}\_\{t\}\)\}\\right\)\.\(2\)Higher values mean a clearer periodic pattern explains more variance inxtx\_\{t\}\.

##### Trend

Trend=max⁡\(0,1−Var​\(Rt\)Var​\(xtdeseas\)\)\.\\mathrm\{Trend\}\\;=\\;\\max\\\!\\left\(0,\\;1\\;\-\\;\\frac\{\\mathrm\{Var\}\(R\_\{t\}\)\}\{\\mathrm\{Var\}\(x^\{\\text\{deseas\}\}\_\{t\}\)\}\\right\)\.\(3\)Higher values mean a smoother long\-term trend explains more variance inxtx\_\{t\}\.

##### Nonstationarity

We report the Augmented Dickey–Fullerpp\-valueliu2022non\.

##### Short\-term distributional change \(Short\_term\_jsd\)

Using a short window of lengthws=30w\_\{s\}=30, for each windowWWform a histogram estimatep^W\\hat\{p\}\_\{W\}on fixed bins and a Gaussian referenceq^W=𝒩​\(μW,σW2\)\\hat\{q\}\_\{W\}=\\mathcal\{N\}\(\\mu\_\{W\},\\sigma\_\{W\}^\{2\}\)discretized on the same bins, whereμW\\mu\_\{W\}andσW\\sigma\_\{W\}are the window mean and standard deviation\. The Jensen–Shannon divergence in windowWWis

JSD\(p^W,q^W\)=12KL\(p^W∥p^W\+q^W2\)\+12KL\(q^W∥p^W\+q^W2\)\.\\mathrm\{JSD\}\(\\hat\{p\}\_\{W\},\\hat\{q\}\_\{W\}\)=\\tfrac\{1\}\{2\}\\,\\mathrm\{KL\}\\\!\\left\(\\hat\{p\}\_\{W\}\\,\\middle\\\|\\,\\tfrac\{\\hat\{p\}\_\{W\}\+\\hat\{q\}\_\{W\}\}\{2\}\\right\)\+\\tfrac\{1\}\{2\}\\,\\mathrm\{KL\}\\\!\\left\(\\hat\{q\}\_\{W\}\\,\\middle\\\|\\,\\tfrac\{\\hat\{p\}\_\{W\}\+\\hat\{q\}\_\{W\}\}\{2\}\\right\)\.\(4\)The metric is the average over all windows:

Short​\_​term​\_​jsd=1Nw​∑WJSD​\(p^W,q^W\)\.\\mathrm\{Short\\\_term\\\_jsd\}\\;=\\;\\frac\{1\}\{N\_\{w\}\}\\sum\_\{W\}\\mathrm\{JSD\}\(\\hat\{p\}\_\{W\},\\hat\{q\}\_\{W\}\)\.\(5\)Larger values indicate that short\-term empirical distributions deviate more from a Gaussian shape\.

##### Shifting

It summarizes typical level changes while retaining the influence of rare but large deviations\.

##### Transition

Discretize\{xt\}\\\{x\_\{t\}\\\}into three equiprobable statesst∈\{1,2,3\}s\_\{t\}\\in\\\{1,2,3\\\}\(tertiles\)\. Letπi=Pr⁡\(st=i\)\\pi\_\{i\}=\\Pr\(s\_\{t\}=i\)andTi​j=Pr⁡\(st\+1=j\|st=i\)T\_\{ij\}=\\Pr\(s\_\{t\+1\}=j\\,\|\\,s\_\{t\}=i\)\. The score is the sum of diagonal covariances between successive states:

Transition\\displaystyle\\mathrm\{Transition\}=∑i=13\[Pr\(st=i,st\+1=i\)\\displaystyle=\\sum\_\{i=1\}^\{3\}\\\!\\Bigl\[\\Pr\(s\_\{t\}=i,s\_\{t\+1\}=i\)\(6\)−Pr\(st=i\)Pr\(st\+1=i\)\]\\displaystyle\\qquad\\qquad\-\\Pr\(s\_\{t\}=i\)\\Pr\(s\_\{t\+1\}=i\)\\Bigr\]=∑i=13\(πi​Ti​i−πi2\)\.\\displaystyle=\\sum\_\{i=1\}^\{3\}\\\!\\left\(\\pi\_\{i\}T\_\{ii\}\-\\pi\_\{i\}^\{2\}\\right\)\.
Higher values indicate that the process tends to stay in the same state more often than expected by chance\.

## Appendix KFeature Analysis of TimesX

The features of both time series and text aggregated by domain are summarized in Table[13](https://arxiv.org/html/2607.06973#A13.T13)\.

The feature aggregated by data source is summarized in Table[14](https://arxiv.org/html/2607.06973#A13.T14)\.

## Appendix LVisualization of Numeric Dataset in TimesX by Domain

We provide visualizations of the numeric time series in TimesX grouped by domain, to give an overview of the series patterns and scales across domains \(Figs\.[13](https://arxiv.org/html/2607.06973#A13.F13)–[31](https://arxiv.org/html/2607.06973#A13.F31)\)\.

## Appendix MPCA and t\-SNE visualizations

In this section we visualize the diversity of TimesX in both the textual and numeric spaces\. For textual features, we compute embeddings for each variable’s context and then apply PCA and t\-SNE to obtain two\-dimensional representations\. For numeric features, we extract summary statistics or learned representations of each time series and apply the same dimensionality\-reduction procedures\. We color the points either by domain or by data source\.

Across all plots we make two consistent observations\. First, there is clear diversity in both the textual and numeric spaces, with points spread over the two\-dimensional maps rather than concentrating in a few tight clusters\. Second, the three data sources form distinct clusters in feature space, which supports the value of integrating multiple sources to achieve broad coverage and diversity\.

![Refer to caption](https://arxiv.org/html/2607.06973v1/x5.png)

Text PCA by domain

![Refer to caption](https://arxiv.org/html/2607.06973v1/x6.png)

Text PCA by source

![Refer to caption](https://arxiv.org/html/2607.06973v1/x7.png)

Text t\-SNE by domain

![Refer to caption](https://arxiv.org/html/2607.06973v1/x8.png)

Text t\-SNE by source

Figure 11:PCA and t\-SNE visualizations of textual features in TimesX\. Each point corresponds to a variable, colored either by domain or by data source\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x9.png)

Time\-series PCA by domain

![Refer to caption](https://arxiv.org/html/2607.06973v1/x10.png)

Time\-series PCA by source

![Refer to caption](https://arxiv.org/html/2607.06973v1/x11.png)

Time\-series t\-SNE by domain

![Refer to caption](https://arxiv.org/html/2607.06973v1/x12.png)

Time\-series t\-SNE by source

Figure 12:PCA and t\-SNE visualizations of numeric time\-series features in TimesX\. Each point corresponds to a variable, colored either by domain or by data source\.Table 13:Domain\-level textual and numeric features\. "AvgEventSummaryLen" is the average number of characters in the event summary text within each domain\. The numeric feature set follows TFB\[qiu2024tfb\]\. The exact computation is detailed in Section[J](https://arxiv.org/html/2607.06973#A10)\.Table 14:Data\-source\-level mean of numeric characteristics \(Transition, Shifting, Seasonality, Trend, NonStationarity \(ADF p\-value\), Short\_term\_jsd\)\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x13.png)Figure 13:Numeric series visualization for the Arts and Entertainment domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x14.png)Figure 14:Numeric series visualization for the Climate and Environment domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x15.png)Figure 15:Numeric series visualization for the Economy domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x16.png)Figure 16:Numeric series visualization for the Electronic Technology domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x17.png)Figure 17:Numeric series visualization for the Finance domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x18.png)Figure 18:Numeric series visualization for the Pets and Animals domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x19.png)Figure 19:Numeric series visualization for the Public Health domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x20.png)Figure 20:Numeric series visualization for the Public Policy and Governance domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x21.png)Figure 21:Numeric series visualization for the Science domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x22.png)Figure 22:Numeric series visualization for the Shopping domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x23.png)Figure 23:Numeric series visualization for the Society, Security, and Social Good domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x24.png)Figure 24:Numeric series visualization for the Traffic domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x25.png)Figure 25:Numeric series visualization for the Crops and Staples domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x26.png)Figure 26:Numeric series visualization for the Currency domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x27.png)Figure 27:Numeric series visualization for the Energy and Fuels domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x28.png)Figure 28:Numeric series visualization for the Livestock and Food Products domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x29.png)Figure 29:Numeric series visualization for the Raw Materials and Construction domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x30.png)Figure 30:Numeric series visualization for the Specialty and Advanced Materials domain\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x31.png)Figure 31:Numeric series visualization for the Strategic and High\-Value Materials domain\.
## Appendix NEvaluation Metrics

##### Principle

We use the Mean Absolute Scaled Error \(MASE\) as the core metric and normalize it by a seasonal naive baseline\. This choice of MASE is followingGIFT\-Eval\[aksu2024gift\]and Chronos Benchmark 2\[ansari2024chronos\]\. We aggregate performance across datasets using the*geometric mean*of normalized scores rather than the arithmetic mean, since prior work proved that the geometric mean is more robust to the choice of normalization baseline\[fleming1986not\]\.

##### Per\-window MASE\.

Let dataseti∈\{1,…,D\}i\\in\\\{1,\\dots,D\\\}be a single variable\. Its full series is\{yt\(i\)\}t=1Ti\\\{y^\{\(i\)\}\_\{t\}\\\}\_\{t=1\}^\{T\_\{i\}\}\. Windoww∈\{1,…,Wi\}w\\in\\\{1,\\dots,W\_\{i\}\\\}has forecast originτi,w\\tau\_\{i,w\}and horizonHi,wH\_\{i,w\}, so the history is\{yt\(i\)\}t=1τi,w\\\{y^\{\(i\)\}\_\{t\}\\\}\_\{t=1\}^\{\\tau\_\{i,w\}\}\. Given a seasonalitymm, the in\-history seasonal scale is

Qi,w=1τi,w−m​∑t=m\+1τi,w\|yt\(i\)−yt−m\(i\)\|\.Q\_\{i,w\}\\;=\\;\\frac\{1\}\{\\tau\_\{i,w\}\-m\}\\sum\_\{t=m\+1\}^\{\\tau\_\{i,w\}\}\\big\|\\,y^\{\(i\)\}\_\{t\}\-y^\{\(i\)\}\_\{t\-m\}\\,\\big\|\.\(7\)For a model with forecasts\{y^τi,w\+h\(i\)\}h=1Hi,w\\\{\\widehat\{y\}^\{\(i\)\}\_\{\\tau\_\{i,w\}\+h\}\\\}\_\{h=1\}^\{H\_\{i,w\}\}, the per\-window MASE is

MASEi,w​\(model\)=1Hi,w​∑h=1Hi,w\|yτi,w\+h\(i\)−y^τi,w\+h\(i\)\|Qi,w\.\\mathrm\{MASE\}\_\{i,w\}\(\\text\{model\}\)\\;=\\;\\frac\{1\}\{H\_\{i,w\}\}\\sum\_\{h=1\}^\{H\_\{i,w\}\}\\frac\{\\big\|\\,y^\{\(i\)\}\_\{\\tau\_\{i,w\}\+h\}\-\\widehat\{y\}^\{\(i\)\}\_\{\\tau\_\{i,w\}\+h\}\\,\\big\|\}\{Q\_\{i,w\}\}\.\(8\)This follows theGIFT\-Evalscaling but replaces a separate training split with the history up to the forecast origin\.

##### Seasonal naive baseline\.

The seasonal naive forecast repeats the last observed seasonal cycle from the history\. Let𝐬\(i,w\)=\(yτi,w−m\+1\(i\),…,yτi,w\(i\)\)\\mathbf\{s\}^\{\(i,w\)\}=\\big\(y^\{\(i\)\}\_\{\\tau\_\{i,w\}\-m\+1\},\\dots,y^\{\(i\)\}\_\{\\tau\_\{i,w\}\}\\big\)\. Then

y^τi,w\+h\(i\),SNAIVE=𝐬1\+\(\(h−1\)modm\)\(i,w\),h=1,…,Hi,w\.\\widehat\{y\}^\{\(i\),\\mathrm\{SNAIVE\}\}\_\{\\tau\_\{i,w\}\+h\}\\;=\\;\\mathbf\{s\}^\{\(i,w\)\}\_\{\\,1\+\\big\(\(h\-1\)\\bmod m\\big\)\},\\quad h=1,\\dots,H\_\{i,w\}\.\(9\)We computeMASEi,w​\(SNAIVE\)\\mathrm\{MASE\}\_\{i,w\}\(\\mathrm\{SNAIVE\}\)by substituting \([9](https://arxiv.org/html/2607.06973#A14.E9)\) into \([8](https://arxiv.org/html/2607.06973#A14.E8)\)\.

##### Per\-dataset normalization\.

For each datasetii, we aggregate across its windows and form a normalized*MASE ratio*:

Ri​\(model\)=∑w=1WiMASEi,w​\(model\)∑w=1WiMASEi,w​\(SNAIVE\)\.R\_\{i\}\(\\text\{model\}\)\\;=\\;\\frac\{\\sum\_\{w=1\}^\{W\_\{i\}\}\\mathrm\{MASE\}\_\{i,w\}\(\\text\{model\}\)\}\{\\sum\_\{w=1\}^\{W\_\{i\}\}\\mathrm\{MASE\}\_\{i,w\}\(\\mathrm\{SNAIVE\}\)\}\.\(10\)ValuesRi<1R\_\{i\}<1indicate improvement over the seasonal naive baseline on datasetii\.

##### Primary aggregate: geometric mean of ratios\.

We report the geometric mean across all datasets as the primary summary:

GM​\(model\)=\(∏i=1DRi​\(model\)\)1D\.\\mathrm\{GM\}\(\\text\{model\}\)\\;=\\;\\left\(\\prod\_\{i=1\}^\{D\}R\_\{i\}\(\\text\{model\}\)\\right\)^\{\\frac\{1\}\{D\}\}\.\(11\)In our release,D=200D=200\.

##### Secondary aggregate: average rank\.

As a complementary, scale\-free indicator, we rank models on each dataset byRiR\_\{i\}\(lower is better\)\. Letranki​\(model\)∈\{1,2,…\}\\mathrm\{rank\}\_\{i\}\(\\text\{model\}\)\\in\\\{1,2,\\dots\\\}be the rank of a model on datasetii\. We report the average rank

AvgRank​\(model\)=1D​∑i=1Dranki​\(model\)\.\\mathrm\{AvgRank\}\(\\text\{model\}\)\\;=\\;\\frac\{1\}\{D\}\\sum\_\{i=1\}^\{D\}\\mathrm\{rank\}\_\{i\}\(\\text\{model\}\)\.\(12\)

##### More Details\.

Unless otherwise specified, we set the seasonality tom=12m=12for monthly data,m=4m=4for weekly data, andm=7m=7for daily data, which matches the construction of our series and the seasonal naive baseline used for normalization\.

## Appendix OEvaluation Samples Statistics

Under our rolling\-window evaluation setup, we obtain a total of 2,434 forecasting samples on TimesX\. These samples cover 19 domains and 190 variables\. On average, each domain contributes about 128\.1 samples, and each variable contributes about 12\.8 samples\.

## Appendix PKnowledge Cutoff of Evaluated Models

Since pretrained models may have ingested training data only up to specific points in time, we explicitly document the knowledge cutoff date of each model considered in our experiments\. This ensures a fair evaluation by avoiding potential data leakage from future information\. The knowledge cutoffs are as follows:

- •GPT\-5: September 30, 2024
- •Gemini\-2\.5\-Flash: January 2025
- •Gemini\-2\.0\-Flash: June 2024
- •GPT\-4o: October 2023
- •DeepSeek\-R1: prior to June 2024
- •DeepSeek\-V3: prior to June 2024

For all evaluations, we align the forecasting horizons such that the prediction targets fall strictly after each model’s knowledge cutoff date, thereby minimizing the risk of contamination\.

### P\.1Test\-time access control

We additionally control test\-time access to external information\. Plain LLM baselines are evaluated with no web search, no tools, and no function calling\. For code/function/agent\-style methods, generated code is executed in a no\-network sandbox\. The allowed function list excludes API\-access functions, and we manually checked execution logs without observing test\-time leakage\.

## Appendix QImplementation Details of Methods

Implementation Details of Pretrained LLMsWe use the LLM decoding settings recommended by OpenRouter\.111111https://openrouter\.ai/\. We follow the prompt format used in CiKWilliams\_2024\.

Implementation Details of Agentic Forecasting Methods

These three methods demonstrate different strategies for integrating TFM with LLMs, ranging from simple textual corrections to complex code generation, providing diverse approaches for context\-aware time series forecasting\.

### Q\.1Text Revision Method \(TextRev\)

#### Q\.1\.1Method Overview

The Text Revision method employs a two\-stage approach: first generating initial numerical forecasts using TimesFM, then leveraging large language models to perform context\-aware textual corrections on these predictions\. This method transforms time series forecasting into a text manipulation task, enabling LLMs to understand and modify numerical predictions through natural language processing\.

#### Q\.1\.2Implementation Steps

1. 1\.Foundation Forecast Generation: TimesFM generates point forecasts based on historical time series data
2. 2\.Text\-based Revision: The TimesFM predictions are converted to timestamp\-value pairs and fed to the LLM along with contextual information
3. 3\.Result Parsing: The corrected forecast values are extracted from the LLM response

#### Q\.1\.3Correction Prompt Template

See Figure[35](https://arxiv.org/html/2607.06973#A26.F35)\.

### Q\.2Function Call Revision Method \(FuncRev\)

#### Q\.2\.1Method Overview

The Function Call Revision method extends the text revision approach by providing LLMs with a structured set of predefined functions for forecast adjustments\. This method incorporates multi\-round conversation mechanisms with text/visual/hybrid critic feedback modes, offering systematic and reproducible forecast modifications\.

#### Q\.2\.2Implementation Steps

1. 1\.Initial Prediction: TimesFM generates the baseline forecast
2. 2\.Multi\-round Revision Loop: - •Critic analyzes current forecast and provides feedback - •Forecaster calls predefined functions to adjust predictions based on feedback - •Process repeats until maximum rounds reached
3. 3\.Final Output: Returns the forecast from the last revision round

#### Q\.2\.3Predefined Function Set

The system provides 13 forecast adjustment functions organized into five categories:

- •Basic Transformations:shift\(offset\),scale\(factor\),linear\_transform\(slope, intercept\)
- •Trend Adjustments:add\_linear\_trend\(slope\),add\_exponential\_trend\(base, growth\_rate\),adjust\_trend\_strength\(factor\)
- •Smoothing Operations:moving\_average\_smooth\(window\),exponential\_smooth\(alpha\)
- •Seasonality Modifications:add\_seasonal\_pattern\(period, amplitude, phase\),remove\_seasonal\_pattern\(period\)
- •Data Normalization:standardize\(\),normalize\_range\(min\_val, max\_val\),clip\_outliers\(lower\_percentile, upper\_percentile\)

#### Q\.2\.4Function Call Prompt Template

See Figure[36](https://arxiv.org/html/2607.06973#A26.F36)\.

### Q\.3Code Revision Method \(CodeRev\)

#### Q\.3\.1Method Overview

The Code Revision method provides maximum adjustment flexibility by allowing LLMs to generate and execute free\-form Python code for forecast modifications\. This approach operates within a secure execution environment, supports multiple scientific computing libraries, and includes timeout and retry mechanisms for robust operation\.

#### Q\.3\.2Implementation Steps

1. 1\.Initial Prediction: TimesFM generates the baseline forecast
2. 2\.Multi\-round Revision Loop: - •Critic provides feedback on current forecast - •LLM generates Python code for forecast adjustments - •Code executes safely in restricted environment - •Retry mechanism activates if execution fails \(maximum 3 attempts\)
3. 3\.Code Execution Environment: Pre\-imported scientific libraries and forecast data variables are provided

#### Q\.3\.3Supported Libraries

The execution environment includes the following pre\-imported libraries:numpy,pandas,math,datetime,requests,sqlite3,csv,json,sympy,statsmodels,networkx\.

#### Q\.3\.4Code Generation Prompt Template

See Figure[37](https://arxiv.org/html/2607.06973#A26.F37)\.

### Q\.4Additional ChatTS result

As an additional related baseline, we evaluated ChatTS Qwen3\-8B on TimesX\. Its aggregated MASE on TimesX is 0\.8933\.

## Appendix RAdditional results on CiK dataset

See results on all the five CiK subsets in Table[15](https://arxiv.org/html/2607.06973#A18.T15)\.

Table 15:Results on all the five CiK subsets \(MASE↓\\downarrow\)\. The geometric\-mean column shows thatCodeRevperforms best, followed by Gemini\-2\.0\-Flash, and then TimesFM\-2\.5\. This ordering is the opposite of what we observe on real\-world data \(TimesX\), illustrating that synthetic generation can flip model rankings\. However, carefully designed synthetic benchmarks like CiK are very useful for testing specific capabilities such as instruction following and different types of reasoning over controlled contexts\. We highly recommend using both synthetic and real\-world benchmarks for a more complete and robust evaluation
## Appendix SAdditional results for context–quality study on Time\-MMD dataset

##### Setup\.

We evaluate on all nineTime\-MMDdatasets\. For LLM methods we fix the prompt template, the decoding settings, and the modelGemini 2\.0 Flash; we*only*replace the textual context \(ours vs\. the originalTime\-MMDcontext\)\. The numeric series remain unchanged\. We use the period from2021\-06\-30to2024\-04\-01, which is the cutoff date ofTime\-MMDdataset\. Fordailydatasets we sethist\_window=365\\text\{hist\\\_window\}=365,pred\_window=120\\text\{pred\\\_window\}=120,slide\_window=40\\text\{slide\\\_window\}=40\. Forweeklydatasets we sethist\_window=96\\text\{hist\\\_window\}=96,pred\_window=12\\text\{pred\\\_window\}=12,slide\_window=4\\text\{slide\\\_window\}=4\. Formonthlydatasets we sethist\_window=16\\text\{hist\\\_window\}=16,pred\_window=4\\text\{pred\\\_window\}=4,slide\_window=2\\text\{slide\\\_window\}=2\. These choices balance sample count and sample diversity while keeping the same evaluation protocol across methods\.

##### Complete per\-domain results\.

Table[18](https://arxiv.org/html/2607.06973#A26.T18)reports normalized performance \(MASE↓\\downarrow\) for each domain and the geometric mean across domains, together with the average rank\. We use two decimals for all numbers and do not report the arithmetic mean\.

## Appendix TModel Rankings Robustness Analysis to Benchmark Size

We study how the size of the evaluation benchmark affects the stability of model performance\. Starting from the 190 in\-distribution variables in TimesX, we construct smaller benchmark variants by randomly sampling subsets of variables without replacement\.

For a target subset sizeK∈\{10,20,…,190\}K\\in\\\{10,20,\\ldots,190\\\}, we draw 20 subsets ofKKdistinct variables\. For each subset and each forecasting model, we compute the geometric\-mean normalized MASE over the variables in that subset\. This gives, for every\(K,model\)\(K,\\text\{model\}\)pair, a distribution of MASE values across random subsets\.

Figure[34](https://arxiv.org/html/2607.06973#A26.F34)summarizes these results\. The horizontal axis is the number of variablesKK, and for each model we plot the mean MASE over the 20 subsets \(solid line\) together with the 5th–95th percentile band \(shaded area\)\. WhenKKis between 10 and 40—which is similar to the sizes of most existing multimodal TSF benchmarks listed in Table[1](https://arxiv.org/html/2607.06973#S2.T1)—the percentile bands are very wide and strongly overlap across models\. This means that, in this small\-scale regime, the apparent ranking of models can change substantially depending on which variables are included in the benchmark\. AsKKincreases, the bands shrink and the relative ordering becomes more stable\.

These observations support our motivation of designing a large\-scale benchmark for more stable and reliable comparisons\.

## Appendix UMore Evaluation Results on TimesX

### U\.1CRPS\-based probabilistic evaluation

In the main text we focus on the \(normalized\) MASE because most LLM\-based methods in our study output point forecasts rather than full predictive distributions\. Here we complement this view with a probabilistic evaluation based on the continuous ranked probability score \(CRPS\)\.

For a predictive distributionFFover a scalar outcomeyy, the CRPS is defined as

CRPS⁡\(F,y\)=∫−∞\+∞\(F​\(z\)−𝟏​\{z≥y\}\)2​dz\.\\operatorname\{CRPS\}\(F,y\)=\\int\_\{\-\\infty\}^\{\+\\infty\}\\bigl\(F\(z\)\-\\mathbf\{1\}\\\{z\\geq y\\\}\\bigr\)^\{2\}\\,\\mathrm\{d\}z\.\(13\)When we only have samplesx1,…,xMx\_\{1\},\\ldots,x\_\{M\}from the predictive distribution, we estimate the CRPS following CiK\[Williams\_2024\]\. Letx1≤⋯≤xMx\_\{1\}\\leq\\cdots\\leq x\_\{M\}denote the samples sorted in ascending order\. An efficient estimator is

CRPS^​\(X~,y\)≈\\displaystyle\\widehat\{\\operatorname\{CRPS\}\}\(\\tilde\{X\},y\)\\approx\{\}1M​∑n=1M\|xn−y\|\+1M​∑n=1Mxn\\displaystyle\\frac\{1\}\{M\}\\sum\_\{n=1\}^\{M\}\\lvert x\_\{n\}\-y\\rvert\+\\frac\{1\}\{M\}\\sum\_\{n=1\}^\{M\}x\_\{n\}\(14\)−2M​\(M−1\)​∑n=1M\(n−1\)​xn,\\displaystyle\-\\frac\{2\}\{M\(M\-1\)\}\\sum\_\{n=1\}^\{M\}\(n\-1\)\\,x\_\{n\},whereX~=\{x1,…,xM\}\\tilde\{X\}=\\\{x\_\{1\},\\ldots,x\_\{M\}\\\}and thexnx\_\{n\}are sorted\. This estimator has𝒪​\(M​log⁡M\)\\mathcal\{O\}\(M\\log M\)time complexity due to the sorting step and is numerically equivalent to the standard unbiased estimator based on pairwise distances\.

In our experiments we treat each stochastic run of a method as one sample from its predictive distribution\. For every method and every evaluation instance on TimesX, we produceM=10M=10stochastic forecasts, compute the CRPS using Equation \([14](https://arxiv.org/html/2607.06973#A21.E14)\), and then aggregate CRPS across variables using geometric mean\. Table[16](https://arxiv.org/html/2607.06973#A21.T16)reports the resulting aggregated CRPS values \(lower is better\)\.

Table 16:Geometric\-mean CRPS on TimesX when estimating predictive uncertainty with 10 stochastic samples per instance\. Lower is better\.All three LLMs achieve lower geometric\-mean CRPS than the three TFMs\. This is consistent with our MASE results and suggests that LLMs can use the rich textual context in TimesX to assign probability mass to multiple plausible futures, while TFMs rely only on numeric series\. We view this CRPS study as an initial step toward more systematic probabilistic evaluation on TimesX; future work may explore improved uncertainty estimation and training objectives that directly optimize probabilistic scores such as CRPS\.

## Appendix VEffects of various variable characteristics

In this section we investigate what variable level features can impact the ordering of LLM based multimodal solutions vs TFMs\.

In Fig\.[32](https://arxiv.org/html/2607.06973#A22.F32), we plot the win rate of Gemini\-2\.0\-Flash against TimesFM\-2\.5 as we climb the quartiles of event counts, length of event details and seasonality\. We can see that with increasing event information the multimodal solutions that leverage these become better than TFMs which are time\-series only\. In the case of seasonality, extremely seasonal series are easy to predict and therefore the edge that TFMs have over LLMs in pure forecasting tasks reduces\.

In Fig\.[33](https://arxiv.org/html/2607.06973#A22.F33), we plot the same while varying various time\-series characteristics like trend, non\-stationarity and transition/ change\-points\. Increasing quartiles of these indicate the hardness of the pure time\-series forecasting task irrespective of the text context, and therefore TFMs can perform better on these time\-series tasks\. Consequently, very strong trends and high non\-stationarity reduce Gemini’s edge over TFMs\.

##### Domain\-level effects of calendar and covariate contexts\.

Holidays are effective in 12/19 domains, with an average intra\-domain win rate of 65\.2%\. The largest gains appear in Shopping, Climate & Environment, Public Health, and Public Policy & Governance\. Covariates are effective in 12/19 domains, with an average intra\-domain win rate of 57\.9%\. The largest gains appear in Specialty & Advanced Materials, Raw Materials & Construction, Livestock & Food Products, and Currency\.

![Refer to caption](https://arxiv.org/html/2607.06973v1/x32.png)Figure 32:We plot the aggregated MASE \(lower is better\) as a function of variable level features like event count, length of event details and seasonality\.![Refer to caption](https://arxiv.org/html/2607.06973v1/x33.png)Figure 33:We plot the aggregated MASE \(lower is better\) as a function of variable level time\-series features like trend, non\-stationarity and transitions\.Table 17:Relationships between LLMs’ multimodal TSF performance and four core capabilities\. We use data with sample start dates after January 2025 on TimesX\. On TimesX, lower MASE indicates better performance, while on the other four benchmarks, higher scores indicate better performance\.

## Appendix WDetailed Experiment Results for Advanced Reasoning Models

We further extend our experiments to advanced reasoning LLMs\. Specifically, for GPT\-4o, Gemini\-2\.0\-Flash, and DeepSeek\-V3, we introduce their corresponding reasoning versions, GPT\-5, Gemini\-2\.5\-Flash, and DeepSeek\-R1\. To avoid data contamination, we select evaluation examples whose forecast horizons begin after January 2025\. As shown in Table[17](https://arxiv.org/html/2607.06973#A22.T17), we observe that while GPT\-5 outperforms GPT\-4o, the reasoning models in the other two pairs perform significantly worse than their non\-reasoning counterparts\. To further understand the drivers of TSF performance, we cross reference the TSF performance with four other core LLM capabilities: factuality \(SimpleQA\[wei2024measuring\]\), multimodal long\-context understanding \([MRCR](https://huggingface.co/datasets/openai/mrcr)\), mathematics \(AIME\[guha2025openthoughts\]\), and science \(GPQA\[rein2024gpqa\]\)\. The results suggest that multimodal TSF is unrelated to the mathematics and science skills emphasized by current reasoning models\. Instead, it depends more on factuality and multimodal long\-context understanding, which are better captured by non\-reasoning LLMs\.

## Appendix XDetails of in\-context learning experiments

Experimental setup\.We denote in\-context learning asICL, andICL\-Cas its conservative variant that allows a no\-change option\. Due to the computational cost and to avoid leakage from in\-context demonstrations, we evaluate all methods on only the last three evaluation instances per variable, and use the remaining instances as the candidate pool for demonstration retrieval\. All other experimental settings remain unchanged\.

Sample selection \(one\-shot, leakage\-free nearest neighbor\)\.For each target evaluation instance, we select a single historical instance as the in\-context demonstration\. To guarantee no leakage, we require the demonstration instance to satisfy*horizon\-end*<<*target\-horizon\-start*\. Among all candidates that satisfy this constraint, we choose the temporally closest one \(i\.e\., with the smallest time gap\), as it is expected to be more similar to the target instance\.

ICL\-C: conservative no\-change option\.In ICL\-C, the prompt allowsTextRevto take a conservative strategy when it is uncertain about the revision, i\.e\., keep the unimodal forecast unchanged\.

Prompt templates\.Figure[38](https://arxiv.org/html/2607.06973#A26.F38)shows the prompt template forTextRev\-ICL\. In the demonstration, the<demo\_forecast\>block uses the ground\-truth future values of the historical \(demonstration\) instance\. For TextRev\-ICL\-C, we further add the following sentence to explicitly allow a conservative*no\-change*option:*If you are not confident about the revision based on the context, you are allowed to keep the initial forecast unchanged\.*

## Appendix YDetailed Per\-Series MASE Results

### Y\.1Detailed Main Benchmarking Results

The detailed per\-series MASE results corresponding to Table[4](https://arxiv.org/html/2607.06973#S4.T4)are reported in Tables[19](https://arxiv.org/html/2607.06973#A26.T19)–[56](https://arxiv.org/html/2607.06973#A26.T56)\.

### Y\.2Detailed MASE Results of Event Type Attribution Analysis with all LLMs

The detailed per\-series MASE results of the event type attribution analysis corresponding to Figure[4](https://arxiv.org/html/2607.06973#S4.F4)are reported in Tables[57](https://arxiv.org/html/2607.06973#A26.T57)–[94](https://arxiv.org/html/2607.06973#A26.T94)\.

### Y\.3Detailed MASE Results of Event Type Deep Analysis Using Gemini

The detailed per\-series MASE results of the event type deep analysis with Gemini corresponding to Table[5](https://arxiv.org/html/2607.06973#S4.T5)are reported in Tables[95](https://arxiv.org/html/2607.06973#A26.T95)–[132](https://arxiv.org/html/2607.06973#A26.T132)\.

### Y\.4Detailed MASE Results of Reasoning Models with Data after January 2025

The detailed per\-series MASE results of reasoning models evaluated on data after January 2025 corresponding to Table[17](https://arxiv.org/html/2607.06973#A22.T17)are reported in Tables[133](https://arxiv.org/html/2607.06973#A26.T133)–[170](https://arxiv.org/html/2607.06973#A26.T170)\.

## Appendix ZDetailed Benchmarking Results by Domain

The detailed per\-domain MASE results corresponding to Table[4](https://arxiv.org/html/2607.06973#S4.T4)are reported in Table[171](https://arxiv.org/html/2607.06973#A26.T171)\.

Table 18:Per\-dataset results onTime\-MMD\(MASE↓\\downarrow\)\. Rows are datasets and summary metrics; columns are methods\. LLM settings \(model/prompt/temperature\) are fixed; only textual context differs\.![Refer to caption](https://arxiv.org/html/2607.06973v1/iclr2026/Rebuttal-Pic/gm_scale_band.png)Figure 34:Effect of benchmark size on the geometric\-mean normalized MASE\. For each subset sizeKK, we sample 20 subsets ofKKvariables from TimesX, compute the metric for each model on each subset, and plot the mean \(solid line\) and the 5th–95th percentile band \(shaded area\)\. WhenKKis at existing benchmark scale \(10–40\), the bands are wide and strongly overlapping, indicating unstable rankings; at largerKKthe bands narrow and the ordering becomes more stable\.[⬇](data:text/plain;base64,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)Ihaveatimeseriesforecastingcorrectiontaskforyou\.Hereissomecontextaboutthetask\.Pleaseconsiderthisinformationwhenreviewingtheforecast:<context\>\{backgroundinformation,constraints,scenariodescriptions,holidayinformation,etc\.\}</context\>Hereisthehistoricaltimeseriesin\(timestamp,value\)format:<history\>\{historicaldatapoints\}</history\>Aninitialforecasthasbeengeneratedusingadeepmodel\.Hereitis:<initial\_forecast\>\{TimesFMpredictionresults\}</initial\_forecast\>Pleasereviewtheinitialforecastandadjustthevaluesconsideringtheprovidedcontext\.Makereasonablemodificationswherethecontextprovidesrelevantinformationthatcouldimprovetheforecast\.Returnyourcorrectedforecastin\(timestamp,value\)formatbetween<forecast\>and</forecast\>tags\.Donotincludeanyotherinformation\(e\.g\.,comments\)intheforecast\.Exampleformat:<forecast\>\(2024\-01\-0112:00:00,123\.45\)\(2024\-01\-0113:00:00,124\.67\)</forecast\>

Figure 35:Prompt template for Text Revision method \(TextRev\)\.[⬇](data:text/plain;base64,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)Youareanexperttimeseriesforecasterwithaccesstoforecastadjustmenttools\.Taskcontext:<context\>\{contextualinformation\}</context\>Criticfeedback:<feedback\>\{criticanalysisandsuggestions\}</feedback\>Currentforecast:<current\_forecast\>\{currentpredictiondata\}</current\_forecast\>Availableadjustmentfunctions:\{functionlist\}Pleaseanalyzethecriticfeedbackandselectappropriatefunctionstoadjusttheforecast\.Youcan:1\.Callasinglefunctionforspecificadjustments2\.Callmultiplefunctionsforcombinedadjustments3\.ChoosenottocallanyfunctionsifthecurrentforecastisalreadyreasonablePleaseexplainyouradjustmentstrategyandcallthecorrespondingfunctions\.

Figure 36:Prompt template for Function Call Revision method \(FuncRev\)\.[⬇](data:text/plain;base64,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)YouareanexperttimeseriesforecasterwithPythonprogrammingcapabilities\.Insteadofusingpredefinedfunctions,youshouldwritePythoncodetoadjusttheforecastbasedonthecritic’sfeedback\.Hereisthecontextaboutthetask:<context\>\{contextualinformation\}</context\>Thecritic’sfeedback:<critic\_feedback\>\{criticfeedback\}</critic\_feedback\>PROGRAMMINGENVIRONMENT:YouhaveaccesstoarestrictedPythonenvironmentwithpre\-importedlibrariesandvariables\.Pre\-importedlibraries:numpy,pandas,math,datetime,requests,sqlite3,csv,json,sympy,statsmodels,networkxAvailablevariablesinyourcode:\-current\_forecast:Dictionarymappingtimestampstoforecastvalues\-timestamps:Listofpredictiontimestamps\(stringsin"YYYY\-MM\-DDHH:MM:SS"format\)\-forecast\_values:ListofforecastvaluescorrespondingtotimestampsINSTRUCTIONS:1\.WritePythoncodetoadjusttheforecastbasedonthecritic’sfeedback2\.DONOTincludeanyimportstatements\-alllibrariesarealreadyimported3\.Yourcodecanperformanymathematicaloperations,transformations,oradjustments4\.Youmustassignthefinaladjustedforecasttoavariablecalled’adjusted\_forecast’5\.The’adjusted\_forecast’shouldbeadictionarymappingtimestampstoadjustedvalues6\.Youcanmodifyforecast\_valueslistandthenreconstructthedictionary,orworkdirectlywithcurrent\_forecast7\.Becreativewithyouradjustments\-you’renotlimitedtopredefinedfunctionsEXAMPLECODESTRUCTURE:‘‘‘python\#Youranalysisandadjustmentlogichere\#DONOTincludeimportstatements\-librariesarepre\-imported\#Example:Applysomeadjustmentbasedoncriticfeedbackfori,timestampinenumerate\(timestamps\):\#Yourlogichereforecast\_values\[i\]=forecast\_values\[i\]\*some\_factor\#exampleadjustment\#Finalresultadjusted\_forecast=\{timestamp:valuefortimestamp,valueinzip\(timestamps,forecast\_values\)\}‘‘‘CRITICALREQUIREMENTS:\-DONOTincludeanyimportstatements\(librariesarepre\-imported\)\-IssyntacticallycorrectPython\-Usesonlythepre\-importedlibraries\-Assignsthefinalresultto’adjusted\_forecast’variable\-HandlestheforecastdataappropriatelyYourPythoncode\(withoutanyimportstatements\):

Figure 37:Prompt template for Code Revision method \(CodeRev\)\.[⬇](data:text/plain;base64,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)Ihaveatimeseriesforecastingcorrectiontaskforyou\.Hereissomecontextaboutthetask\.Pleaseconsiderthisinformationwhenreviewingtheforecast:<context\>\{backgroundinformation,constraints,scenariodescriptions,holidayinformation,etc\.\}</context\>Hereisthehistoricaltimeseriesin\(timestamp,value\)format:<history\>\{historicaldatapoints\}</history\>Aninitialforecasthasbeengeneratedusingadeepmodel\.Hereitis:<initial\_forecast\>\{TimesFMpredictionresults\}</initial\_forecast\>\#One\-shotdemonstration\(historicalinstance\)\#NOTE:demo\_forecastusestheground\-truthfuturevaluesforthedemonstrationinstance\.<demo\><demo\_context\>\{contextforthedemonstrationinstance\}</demo\_context\><demo\_history\>\{historicaldatapointsforthedemonstrationinstance\}</demo\_history\><demo\_initial\_forecast\>\{initialforecastforthedemonstrationinstance\}</demo\_initial\_forecast\><demo\_forecast\>\{ground\-truthfuturevaluesforthedemonstrationinstance\}</demo\_forecast\></demo\>Pleasereviewtheinitialforecastandadjustthevaluesconsideringtheprovidedcontext\.Makereasonablemodificationswherethecontextprovidesrelevantinformationthatcouldimprovetheforecast\.Returnyourcorrectedforecastin\(timestamp,value\)formatbetween<forecast\>and</forecast\>tags\.Donotincludeanyotherinformation\(e\.g\.,comments\)intheforecast\.Exampleformat:<forecast\>\(2024\-01\-0112:00:00,123\.45\)\(2024\-01\-0113:00:00,124\.67\)</forecast\>

Figure 38:Prompt template forTextRev\-ICL\(one\-shot demonstration\)\.Table 19:Detailed MASE results of Table 1\. \(part 1/38\)Table 20:Detailed MASE results of Table 1\. \(cont’d, part 2/38\)Table 21:Detailed MASE results of Table 1\. \(cont’d, part 3/38\)Table 22:Detailed MASE results of Table 1\. \(cont’d, part 4/38\)Table 23:Detailed MASE results of Table 1\. \(cont’d, part 5/38\)Table 24:Detailed MASE results of Table 1\. \(cont’d, part 6/38\)Table 25:Detailed MASE results of Table 1\. \(cont’d, part 7/38\)Table 26:Detailed MASE results of Table 1\. \(cont’d, part 8/38\)Table 27:Detailed MASE results of Table 1\. \(cont’d, part 9/38\)Table 28:Detailed MASE results of Table 1\. \(cont’d, part 10/38\)Table 29:Detailed MASE results of Table 1\. \(cont’d, part 11/38\)Table 30:Detailed MASE results of Table 1\. \(cont’d, part 12/38\)Table 31:Detailed MASE results of Table 1\. \(cont’d, part 13/38\)Table 32:Detailed MASE results of Table 1\. \(cont’d, part 14/38\)Table 33:Detailed MASE results of Table 1\. \(cont’d, part 15/38\)Table 34:Detailed MASE results of Table 1\. \(cont’d, part 16/38\)Table 35:Detailed MASE results of Table 1\. \(cont’d, part 17/38\)Table 36:Detailed MASE results of Table 1\. \(cont’d, part 18/38\)Table 37:Detailed MASE results of Table 1\. \(cont’d, part 19/38\)Table 38:Detailed MASE results of Table 1\. \(cont’d, part 20/38\)Table 39:Detailed MASE results of Table 1\. \(cont’d, part 21/38\)Table 40:Detailed MASE results of Table 1\. \(cont’d, part 22/38\)Table 41:Detailed MASE results of Table 1\. \(cont’d, part 23/38\)Table 42:Detailed MASE results of Table 1\. \(cont’d, part 24/38\)Table 43:Detailed MASE results of Table 1\. \(cont’d, part 25/38\)Table 44:Detailed MASE results of Table 1\. \(cont’d, part 26/38\)Table 45:Detailed MASE results of Table 1\. \(cont’d, part 27/38\)Table 46:Detailed MASE results of Table 1\. \(cont’d, part 28/38\)Table 47:Detailed MASE results of Table 1\. \(cont’d, part 29/38\)Table 48:Detailed MASE results of Table 1\. \(cont’d, part 30/38\)Table 49:Detailed MASE results of Table 1\. \(cont’d, part 31/38\)Table 50:Detailed MASE results of Table 1\. \(cont’d, part 32/38\)Table 51:Detailed MASE results of Table 1\. \(cont’d, part 33/38\)Table 52:Detailed MASE results of Table 1\. \(cont’d, part 34/38\)Table 53:Detailed MASE results of Table 1\. \(cont’d, part 35/38\)Table 54:Detailed MASE results of Table 1\. \(cont’d, part 36/38\)Table 55:Detailed MASE results of Table 1\. \(cont’d, part 37/38\)Table 56:Detailed MASE results of Table 1\. \(cont’d, part 38/38\)Table 57:Detailed MASE Results of Event Type Attribution Analysis\. \(part 1/38\)Table 58:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 2/38\)Table 59:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 3/38\)Table 60:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 4/38\)Table 61:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 5/38\)Table 62:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 6/38\)Table 63:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 7/38\)Table 64:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 8/38\)Table 65:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 9/38\)Table 66:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 10/38\)Table 67:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 11/38\)Table 68:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 12/38\)Table 69:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 13/38\)Table 70:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 14/38\)Table 71:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 15/38\)Table 72:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 16/38\)Table 73:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 17/38\)Table 74:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 18/38\)Table 75:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 19/38\)Table 76:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 20/38\)Table 77:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 21/38\)Table 78:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 22/38\)Table 79:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 23/38\)Table 80:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 24/38\)Table 81:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 25/38\)Table 82:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 26/38\)Table 83:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 27/38\)Table 84:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 28/38\)Table 85:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 29/38\)Table 86:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 30/38\)Table 87:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 31/38\)Table 88:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 32/38\)Table 89:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 33/38\)Table 90:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 34/38\)Table 91:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 35/38\)Table 92:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 36/38\)Table 93:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 37/38\)Table 94:Detailed MASE Results of Event Type Attribution Analysis\. \(cont’d, part 38/38\)Table 95:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(part 1/38\)Table 96:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 2/38\)Table 97:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 3/38\)Table 98:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 4/38\)Table 99:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 5/38\)Table 100:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 6/38\)Table 101:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 7/38\)Table 102:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 8/38\)Table 103:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 9/38\)Table 104:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 10/38\)Table 105:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 11/38\)Table 106:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 12/38\)Table 107:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 13/38\)Table 108:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 14/38\)Table 109:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 15/38\)Table 110:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 16/38\)Table 111:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 17/38\)Table 112:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 18/38\)Table 113:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 19/38\)Table 114:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 20/38\)Table 115:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 21/38\)Table 116:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 22/38\)Table 117:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 23/38\)Table 118:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 24/38\)Table 119:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 25/38\)Table 120:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 26/38\)Table 121:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 27/38\)Table 122:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 28/38\)Table 123:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 29/38\)Table 124:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 30/38\)Table 125:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 31/38\)Table 126:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 32/38\)Table 127:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 33/38\)Table 128:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 34/38\)Table 129:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 35/38\)Table 130:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 36/38\)Table 131:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 37/38\)Table 132:Detailed MASE Results of Event Type Deep Analysis Using Gemini\-2\.0\-Flash \(cont’d, part 38/38\)Table 133:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(part 1/38\)Table 134:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 2/38\)Table 135:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 3/38\)Table 136:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 4/38\)Table 137:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 5/38\)Table 138:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 6/38\)Table 139:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 7/38\)Table 140:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 8/38\)Table 141:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 9/38\)Table 142:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 10/38\)Table 143:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 11/38\)Table 144:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 12/38\)Table 145:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 13/38\)Table 146:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 14/38\)Table 147:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 15/38\)Table 148:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 16/38\)Table 149:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 17/38\)Table 150:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 18/38\)Table 151:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 19/38\)Table 152:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 20/38\)Table 153:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 21/38\)Table 154:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 22/38\)Table 155:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 23/38\)Table 156:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 24/38\)Table 157:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 25/38\)Table 158:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 26/38\)Table 159:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 27/38\)Table 160:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 28/38\)Table 161:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 29/38\)Table 162:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 30/38\)Table 163:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 31/38\)Table 164:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 32/38\)Table 165:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 33/38\)Table 166:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 34/38\)Table 167:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 35/38\)Table 168:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 36/38\)Table 169:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 37/38\)Table 170:Detailed MASE Results of Reasoning Models with Data after 2025 Jan \(cont’d, part 38/38\)Table 171:Breakdown of MASE of all methods on each of the 19 domains\. Acronyms are used to shorten domain names for clarity: see Appendix[I](https://arxiv.org/html/2607.06973#A9)\.