Pith. sign in

REVIEW 5 major objections 5 minor 47 references

A three-stage LLM pipeline generates time-series forecast explanations that approach analyst-written reports in readability, consistency, and persuasiveness without domain-specific fine-tuning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 12:11 UTC pith:CVJKPVRC

load-bearing objection Plausible modular framework for grounded LLM explanation of time series, but the headline 'approached analyst-written' claim rests on an unvalidated LLM judge and a five-sample o3 comparison; deserving of peer review, needs major revision. the 5 major comments →

arxiv 2607.18271 v1 pith:CVJKPVRC submitted 2026-06-04 cs.AI

Using LLMs for Explainable, Data-Driven Insight Generation from Time Series

Ria Mundhra , Gustavo Sato dos Santos , Michael Benedikt This is my paper
classification cs.AI
keywords time series forecastingexplanation generationlarge language modelsretrieval-augmented generationfactor extractionevaluation of explanationsgroundednesspersuasiveness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that grounded, analyst-quality explanations for time series forecasts can be generated automatically at scale. It proposes a three-stage pipeline: extract structured explanatory factors from historical analyst reports; condition an LLM on encoded time-series evidence and summarized external news; and evaluate outputs along readability, logical consistency, and persuasiveness. In two case studies—a stock index and freight pricing—the generated reports approached, and in some pairwise comparisons matched, human analyst writing, with a stronger reasoning model closing most of the remaining gap. The authors argue that constraining generation to verifiable evidence reduces hallucination and makes this quality achievable without domain-specific fine-tuning. If correct, the framework offers a practical route to on-demand, evidence-linked explanations in finance and other forecast-heavy fields.

Core claim

On the paper's own terms, the central discovery is that constraining LLM generation to verifiable evidence—extracted factors, multi-scale STL-decomposed time-series summaries, and summarized external news—yields explanations that a battery of automated and human evaluations finds comparable to analyst-written ones in readability, logical consistency, and persuasiveness. The evidence shows that time-series encoding is the main lever for reducing internal contradictions, external news summaries drive persuasiveness, and few-shot examples align style; a stronger reasoning model substantially narrows the gap to analyst text in persuasive pairwise comparisons, though the sample is small and the e

What carries the argument

The key mechanism is the three-stage pipeline. First, historical factor extraction uses LLMs to parse analyst explanations into structured factors (name, evidence span, attributes) and clusters them into semantic themes. Second, evidence-conditioned generation maps each cluster to concrete evidence—quantitative time series encoded as natural-language dictionaries via multi-scale STL decomposition, and qualitative news summaries—and interleaves this evidence into a structured prompt that positions the LLM as a linguistic realizer. Third, multi-axis evaluation measures readability (Gunning-Fog plus LLM-based smart cloze), logical consistency (deductive-closure contradiction detection against h

Load-bearing premise

The load-bearing premise is that the LLM-based evaluation instruments—persuasiveness judge, contradiction verifier, and cloze readability—faithfully proxy expert human judgment; if they do not, the claimed parity with analyst-written explanations could be a measurement artifact.

What would settle it

A concrete test: conduct a blinded, expert-rated head-to-head comparison between generated explanations and analyst-written explanations on a sample of dozens of reports, using the paper's own 5-point rubric, and check whether the parity claim survives human judgment. Separately, if the LLM persuasiveness judge fails its own paraphrase-tie sanity check on a larger set—labeling original-analyst and paraphrase pairs as non-ties—then the judge's validity, and with it the persuasiveness comparison, is undercut.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Organizations that rely on forecast explanations could move from slow, expert-written narratives to on-demand, evidence-linked reports produced automatically by this kind of pipeline.
  • The framework is intended to be domain-agnostic: applying it to a new domain mainly requires a corpus of historical expert explanations and aligned external signals, not domain-specific fine-tuning.
  • The ablation results give concrete engineering priorities: time-series encoding for logical consistency, external news for persuasiveness, few-shot examples for readability, with a combined configuration performing consistently well despite trade-offs.
  • Stronger reasoning models reduce the quality gap to human writing, suggesting the pipeline's ceiling rises with underlying LLM capability even without architectural changes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Since the evaluation's persuasiveness and consistency metrics themselves rely on LLM judgments, the parity result could be partly an artifact of LLM judges preferring LLM-style prose; a larger blinded human-expert study would test this directly.
  • The factor-to-evidence mapping is described as partially human-assisted, so 'domain-agnostic' is conditional on some per-domain configuration; automating that mapping is a natural next step.
  • The contradiction axioms are handcrafted; learning them from the corpus or from the extracted factors themselves could make the consistency check fully domain-adaptive.
  • The framework treats the forecast as given; an extension could couple explanations with forecast uncertainty, which would make it more directly useful for decision support.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes a modular, domain-agnostic framework for generating grounded natural-language explanations of time-series forecasts. The pipeline consists of (i) extracting structured explanatory factors from historical analyst-written reports (AnX) using LLM prompts, clustering, and a retrieval-based validation; (ii) generating explanations (GenX) by conditioning an LLM on encoded time-series summaries, external news summaries, and few-shot examples; and (iii) evaluating output along readability (Gunning–Fog and LLM-based smart cloze), logical consistency (LLM-based contradiction detection against author-written axioms), and persuasiveness (LLM-as-a-judge pairwise comparisons). The framework is applied to NASDAQ-100 forecasts (100 gpt-4o-mini reports, 5 o3 reports) and a secondary freight-pricing case study. The central claim is that generated explanations 'approached analyst-written explanations' on readability, consistency, and persuasiveness, and that this is achievable at scale without domain-specific fine-tuning. The paper includes an ablation study, a qualitative human evaluation of three reports, and detailed prompts in the appendix.

Significance. If the evaluation were fully valid, the framework would be a valuable step toward scalable, evidence-grounded explanation generation for time-series forecasting, a task currently dominated by expensive manual expert analysis. The modular design—factor extraction, evidence-conditioned generation, and multi-axis evaluation—is sensible and well described, with concrete prompts and a running example that make the pipeline reproducible. The explicit grounding of each claim in evidence spans and encoded time-series summaries is a real strength. However, the headline comparative claim rests on evaluation instruments whose validity is not established: the persuasiveness judge is an uncalibrated LLM, the consistency measure checks against author-written axioms rather than internal consistency, and the readability comparison lacks reported reference values and significance tests. The human evaluation is too small and not used to validate the automatic judges. These gaps are load-bearing and currently prevent the paper from supporting its central claim.

major comments (5)
  1. [§6, §7.1 F3, Table 2] The persuasiveness result is not validated. The LLM judge's only sanity check is a paraphrase-tie test, which checks that surface-form variation does not flip the decision; it does not show that the judge's preferences correlate with expert human preferences. The human evaluation in Appendix C rates three reports on a Likert rubric but never asks human evaluators to perform the same pairwise comparisons as the judge, so it cannot establish criterion validity. Given known LLM-as-a-judge biases (e.g., preferring LLM-style prose), the 'approached analyst-written' claim for persuasiveness is unsupported and could be an artifact of the judge.
  2. [§7, Table 2] The o3 result rests on n=5 generated explanations. With such a small sample, the contrast between o3 (50% preferred, 50% tie) and gpt-4o-mini (70% reference preferred) carries no statistical weight. The paper notes this in a caveat, but the abstract and conclusion state the 'approached analyst-written' result without this qualification. A sample of 5 cannot support the central comparative claim; at minimum, confidence intervals and a larger set of o3 outputs are needed.
  3. [§6, Appendix A.5] The logical-consistency measure is an LLM checking each extracted claim against a small set of author-written axioms. This measures consistency with those axioms, not internal consistency of the explanation, and the detector is not calibrated against human raters. The contradiction rates in Table 1 and Figure 9 could reflect axiom coverage and the judge's own biases rather than actual inconsistencies. Without a human-annotated set of true contradictions, the F2 finding (and the regeneration-loop refinement) is not established.
  4. [§6, Figure 8, Table 1] The readability claim is not backed by reported AnX comparison values. Table 1 reports GFI and smart cloze scores for the generator configurations, but no AnX row or statistical comparison is given; the text asserts 'no practically meaningful differences' without presenting the analyst-written reference scores or any significance test. Figure 8 is referenced but the actual numeric comparison is absent from the manuscript. The readability axis therefore cannot be independently assessed.
  5. [§4.4, §5] Factor extraction quality is validated only through an LLM-based retrieval task—whether a reasoning LLM can map attribute sets back to the source AnX. This is an internal loop: the same family of LLMs is used for extraction and validation, and the task measures source-attribution ability, not whether the factors are correct or useful for explanation. Moreover, the 'domain-agnostic without fine-tuning' claim is contradicted by the NASDAQ-specific extraction prompt (Appendix A.1) and by the statement that evidence retrieval 'is domain-specific and partially human-assisted' (§5). These weaken the central generality claim.
minor comments (5)
  1. [Throughout] Typos and wording: 'dat of report generation' (§5), 'a analyst-written explanation' (Definition 4.1), '4om' in Table 2 should be spelled out in the caption. Also 'etiological reasoning' (§4.4) is used without definition.
  2. [§6] The smart cloze procedure is described only briefly; details of how the distractor vocabulary is constructed and how content-bearing tokens are selected would improve reproducibility.
  3. [Appendix C] The human evaluation appendix describes the protocol but does not report the actual Likert scores, the 'Yes/No' trust responses, or any aggregate results—only qualitative recurring themes. Reporting the numerical outcomes would make the evaluation interpretable.
  4. [§7.1 F1] The Gunning–Fog formula is shown with ambiguous spacing; the standard formula is clear, but a clean equation would help. More importantly, the claim that 'all explanations fall within an advanced reading range' is not supported by any reported AnX GFI value.
  5. [§5] The number of external news articles (top 10) and other hyperparameters appear arbitrary; a brief sensitivity discussion or justification would strengthen the framework's credibility.

Circularity Check

1 steps flagged

Persuasiveness axis is self-referential: the LLM judge's rubric mirrors the generation prompt, so the 'approached analyst-written' claim partly measures prompt adherence.

specific steps
  1. self definitional [Section 7.1 (F3, Table 2); judge prompt in Appendix A.6; generation prompt in Appendix A.4]
    "A.6: 'we define convincing as coherent, relevant, specific, references important data points and external factors...' A.4: 'You should cite datapoints as much as possible... Use both the provided additional data, as well as external data.'"

    The persuasiveness judge's operational definition of 'convincing' is essentially the checklist given to the generator: be specific, cite data points, incorporate external factors. Thus the Table 2 comparison of GenX vs. AnX on persuasiveness largely measures how well each text follows the generator's own instructions, not an independently grounded quality. The only validation, the paraphrase-tie check, tests surface-form insensitivity rather than agreement with expert human preference, so it does not break the loop. Since F3 and the abstract's 'approached analyst-written... persuasiveness' claim rest on this judge, the result is partially self-definitional.

full rationale

The framework's derivation chain is mostly non-circular: factor extraction is validated by an AnX-retrieval task that checks information preservation; generation conditions on encoded time series and external evidence; readability uses the standard GFI plus a cloze procedure; and contradiction detection is checked against externally authored domain axioms. No fitted parameter is renamed as a prediction, and no load-bearing self-citation chain was found. The one significant circular element is the persuasiveness axis: the LLM judge's rubric (Appendix A.6) overlaps almost verbatim with the generation prompt's instructions (Appendix A.4), so the central 'approached analyst-written' claim on persuasiveness partly reduces to an LLM judging LLM output against the LLM's own conditioning criteria. The paper itself concedes in Section 8 that LLM judges 'remain imperfect proxies for expert human judgment,' and the human evaluation (Appendix C) covers only three reports and does not cross-validate the LLM judge's pairwise choices. Readability and consistency results retain independent content, and the freight qualitative feedback provides some external support, so the circularity is partial rather than total. Score 4 reflects one load-bearing self-definitional evaluation step without equation-level input-output equivalence.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The framework has no physical constants or fitted numerical laws; the free parameters are methodological choices (m=6, HDBSCAN/UMAP settings, top-k=10, forward-selection threshold, cloze settings). The empirical claims rest on assumptions that LLM judges and cloze tests proxy human judgment, that hand-authored domain axioms are valid, and that the AnX corpus is temporally appropriate for each forecast week.

free parameters (5)
  • m=6 candidate AnX in retrieval task = 6
    Number of candidate analyst reports in the factor-extraction retrieval diagnostic; chosen by the authors and affects the reported extraction-quality accuracy (§4.4).
  • HDBSCAN/UMAP hyperparameters = not reported
    Factor clustering depends on these settings, but no values are given (§4.3).
  • Forward variable selection stopping threshold = not specified
    Retained attribute set for generation depends on a 'statistically significant gain' rule that is not quantified (§4.4).
  • Top-k news articles for context = 10 (running example)
    Retrieval of external context is illustrated with top-10 articles; no ablation or general rule is given (§5).
  • Smart cloze blank/distractor settings = not fully specified
    Domain-specific readability depends on POS tagging, blank selection, and distractor sampling; exact parameters are adapted from [33] but not fully reported (§6).
axioms (4)
  • ad hoc to paper LLM judges and LLM-filled cloze tests are valid proxies for expert human judgment of persuasiveness, consistency, and readability.
    The headline evaluation depends on this; the only external check is a described-but-unquantified paraphrase-tie test and a 3-report human evaluation (§6, Appendix A.5–A.6, Appendix C).
  • domain assumption Hand-authored domain axioms such as 'increase in index prices cannot be caused by sustained increase in interest rates' are correct and sufficient for contradiction detection.
    Used in the consistency evaluation (§6); if an axiom is oversimplified, flagged 'contradictions' become false positives and the reported consistency rates lose meaning.
  • domain assumption The 100 AnX in the NASDAQ corpus are representative of the explanatory vocabulary needed for future forecasts, and can be used without temporal lookahead.
    The corpus was retrieved July 23, 2025 and used to generate explanations for 2024 weeks; no temporal split is described (§7).
  • domain assumption STL-derived text summaries preserve the temporal structure needed for causal explanation.
    The generation pipeline relies on these compact encodings rather than raw numbers (§5); no direct validation shows the summaries are sufficient for faithful reasoning.
invented entities (1)
  • Factor tuple (n, e, A) no independent evidence
    purpose: Structured representation of explanatory influences extracted from analyst text, used to condition downstream LLM generation.
    A representational construct with no external causal handle; it is validated only through LLM retrieval tasks and downstream generation quality, not independently.

pith-pipeline@v1.3.0-alltime-deepseek · 17647 in / 14377 out tokens · 143475 ms · 2026-08-02T12:11:14.930879+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Using LLMs for Explainable, Data-Driven Insight Generation from Time Series." pith.science (2026). https://pith.science/paper/CVJKPVRC

@misc{pith2026260718271,
  author       = {Pith},
  title        = {Pith review of: Using LLMs for Explainable, Data-Driven Insight Generation from Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CVJKPVRC}},
  note         = {Machine review of arXiv:2607.18271}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Time series forecasts are widely used in decision-critical domains, where they are rarely consumed without accompanying explanations. Producing such explanations is usually a manual and costly process, and attempts to automate it using large language models often suffer from hallucination when applied to temporal data. We propose a domain-agnostic framework for grounded natural language explanation generation for time series forecasts, illustrated in Figure 1. The framework consists of three components: (i) extraction of structured explanatory factors from historical analyst-written explanations, (ii) evidence-conditioned explanation generation, and (iii) scalable evaluation for readability, logical consistency, and persuasiveness. The design explicitly constrains generation to verifiable evidence, reducing unsupported claims. We evaluate the framework on a financial forecasting case study involving the NASDAQ-100 index and a freight pricing case study using data from Vortexa. Results show that generated explanations approached analyst-written explanations in terms of readability, consistency and persuasiveness. These findings demonstrate that grounded explanation generation for time series forecasting can be achieved at scale without domain-specific fine-tuning.

Figures

Figures reproduced from arXiv: 2607.18271 by Gustavo Sato dos Santos, Michael Benedikt, Ria Mundhra.

Figure 1
Figure 1. Figure 1: A modular framework for generating grounded natural language explanations for time series forecasts. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Target time series for the running NASDAQ-100 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The historical factor extraction pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The AnX retrieval pipeline. A reasoning model is [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The grounded explanation generation stage. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example encoding for weekly Unemployment Rate [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Domain-specific readability comparison using the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Contradiction rates across explanation types. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: C.4 Qualitative Feedback In addition to numeric ratings, evaluators were encouraged to pro￾vide free-form qualitative comments. These comments proved par￾ticularly valuable for identifying strengths and weaknesses not fully captured by scalar scores. Recurring themes in the feedback included: • The importance of explicitly distinguishing between short￾term and long-term effects of the same market driver. … view at source ↗
Figure 10
Figure 10. Figure 10: Full evaluation framework provided to human [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 13 linked inside Pith

  1. [1]

    Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. 2001. On the surprising behavior of distance metrics in high dimensional space. InInternational Conference on Database Theory. Springer, 420–434

    2001

  2. [2]

    Afra Feyza Akyürek, Ekin Akyürek, Leshem Choshen, Derry Wijaya, and Jacob Andreas. 2024. Deductive closure training of language models for coherence, accuracy, and updatability.arXiv preprint arXiv:2401.08574(2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Shruthi Bannur, Kenza Bouzid, Daniel C Castro, Anton Schwaighofer, Anja Thieme, Sam Bond-Taylor, Maximilian Ilse, Fernando Pérez-García, Valentina Salvatelli, Harshita Sharma, et al . 2024. Maira-2: Grounded radiology report generation.arXiv preprint arXiv:2406.04449(2024)

    Pith/arXiv arXiv 2024

  4. [4]

    Michael M Bechtel and Dirk Leuffen. 2010. Forecasting European Union politics: Real-time forecasts in political time series analysis.European Union Politics11, 2 (2010), 309–327

    2010

  5. [5]

    Angana Borah, Rada Mihalcea, and Verónica Pérez-Rosas. 2025. Persuasion at Play: Understanding Misinformation Dynamics in Demographic-Aware Human- LLM Interactions.arXiv preprint arXiv:2503.02038(2025)

    arXiv 2025

  6. [6]

    Patrick T Brandt and John R Freeman. 2006. Advances in Bayesian time se- ries modeling and the study of politics: Theory testing, forecasting, and policy analysis.Political Analysis14, 1 (2006), 1–36

    2006

  7. [7]

    Ching Chang, Wen-Chih Peng, and Tien-Fu Chen. 2023. Llm4ts: Two-stage fine-tuning for time-series forecasting with pre-trained llms.CoRR(2023)

    2023

  8. [8]

    Yew Ken Chia, Guizhen Chen, Luu Anh Tuan, Soujanya Poria, and Lidong Bing

  9. [9]

    Robert B Cleveland, William S Cleveland, Jean E McRae, Irma Terpenning, et al

  10. [10]

    Bhuwan Dhingra, Manaal Faruqui, Ankur Parikh, Ming-Wei Chang, Dipanjan Das, and William Cohen. 2019. Handling Divergent Reference Texts when Evaluating Table-to-Text Generation. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Anna Korhonen, David Traum, and Lluís Màrquez (Eds.). Association for Computational Ling...

    doi:10.18653/v1/p19-1483 2019

  11. [11]

    Pedro Domingos. 2012. A few useful things to know about machine learning. Commun. ACM55, 10 (2012), 78–87

    2012

  12. [12]

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234(2022)

    Pith/arXiv arXiv 2022

  13. [13]

    Hao Fei, Yuan Yao, Zhuosheng Zhang, Fuxiao Liu, Ao Zhang, and Tat-Seng Chua. 2024. From Multimodal LLM to Human-level AI: Modality, Instruction, Reasoning, Efficiency and beyond. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024): Tutorial Summaries, Roman Klinger, Na...

    2024

  14. [14]

    Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal, Pawan Sasanka Am- manamanchi, Aremu Anuoluwapo, Antoine Bosselut, Khyathi Raghavi Chandu, Miruna Clinciu, Dipanjan Das, Kaustubh D Dhole, et al. 2021. The gem bench- mark: Natural language generation, its evaluation and metrics.arXiv preprint arXiv:2102.01672(2021)

    Pith/arXiv arXiv 2021

  15. [15]

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al . 2024. A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594(2024)

    Pith/arXiv arXiv 2024

  16. [16]

    Robert Gunning. 1952. The technique of clear writing.McGraw-Hill(1952)

    1952

  17. [17]

    Isabelle Guyon and André Elisseeff. 2003. An introduction to variable and feature selection.Journal of machine learning research3, Mar (2003), 1157–1182

    2003

  18. [18]

    Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al

  19. [19]

    2018.Forecasting: principles and practice

    Rob J Hyndman and George Athanasopoulos. 2018.Forecasting: principles and practice. OTexts

    2018

  20. [20]

    Yushan Jiang, Wenchao Yu, Geon Lee, Dongjin Song, Kijung Shin, Wei Cheng, Yanchi Liu, and Haifeng Chen. 2025. Explainable multi-modal time series predic- tion with llm-in-the-loop.arXiv preprint arXiv:2503.01013(2025)

    arXiv 2025

  21. [21]

    Large language models: a comprehensive survey of its applications, chal- lenges, limitations, and future prospects.Authorea Preprints1 (2023), 1–26

    2023

  22. [22]

    Mitesh M Khapra and Ananya B Sai. 2021. A tutorial on evaluation metrics used in natural language generation. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Tutorials. 15–19

    2021

  23. [23]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33 (2020), 9459–9474

    2020

  24. [24]

    Shruti Kaushik, Abhinav Choudhury, Pankaj Kumar Sheron, Nataraj Dasgupta, Sayee Natarajan, Larry A Pickett, and Varun Dutt. 2020. AI in healthcare: time- series forecasting using statistical, neural, and ensemble architectures.Frontiers in big data3 (2020), 4

    2020

  25. [25]

    Spyros Makridakis, Evangelos Spiliotis, and Vassilios Assimakopoulos. 2018. Sta- tistical and Machine Learning forecasting methods: Concerns and ways forward. PloS one13, 3 (2018), e0194889

    2018

  26. [26]

    Leland McInnes, John Healy, Steve Astels, et al . 2017. hdbscan: Hierarchical density based clustering.J. Open Source Softw.2, 11 (2017), 205

    2017

  27. [27]

    Chris P Long and Sim B Sitkin. 2023. Contradictions that erode institutional trust & opportunities for addressing them.Behavioral Science & Policy9, 2 (2023), 1–6

    2023

  28. [28]

    Mike A Merrill, Mingtian Tan, Vinayak Gupta, Tom Hartvigsen, and Tim Althoff

  29. [29]

    Ning Miao, Yee Whye Teh, and Tom Rainforth. 2023. SelfCheck: Using LLMs to Zero-Shot Check Their Own Step-by-Step Reasoning. arXiv:2308.00436 [cs.AI] https://arxiv.org/abs/2308.00436

    Pith/arXiv arXiv 2023

  30. [30]

    Leland McInnes, John Healy, and James Melville. 2018. Umap: Uniform man- ifold approximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426(2018)

    Pith/arXiv arXiv 2018

  31. [31]

    Andrew Pang, Hyeju Jang, and Shiaofen Fang. 2024. Generating Descriptive Explanations of Machine Learning Models Using LLM. In2024 IEEE International Conference on Big Data (BigData). IEEE, 5369–5374

    2024

  32. [32]

    Min Park and Tzachi Zach. 2025. Analysts’ forecasting models and uncertainty about the past.Review of Accounting Studies(2025), 1–43

    2025

  33. [33]

    Elissa Redmiles, Lisa Maszkiewicz, Emily Hwang, Dhruv Kuchhal, Everest Liu, Miraida Morales, Denis Peskov, Sudha Rao, Rock Stevens, Kristina Gligorić, et al

  34. [34]

    Kanghui Ning, Zijie Pan, Yu Liu, Yushan Jiang, James Y Zhang, Kashif Rasul, Anderson Schneider, Lintao Ma, Yuriy Nevmyvaka, and Dongjin Song. 2025. Ts-rag: Retrieval-augmented generation based time series foundation models are stronger zero-shot forecaster.arXiv preprint arXiv:2503.07649(2025)

    arXiv 2025

  35. [35]

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. 2020. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019.Applied soft computing90 (2020), 106181

    2020

  36. [36]

    Jialiang Tang, Shuo Chen, Chen Gong, Jing Zhang, and Dacheng Tao. 2025. LLM-PS: Empowering Large Language Models for Time Series Forecasting with Temporal Patterns and Semantics. arXiv:2503.09656 [cs.LG] https://arxiv.org/ abs/2503.09656

    Pith/arXiv arXiv 2025

  37. [37]

    Cloze procedure

    Wilson L Taylor. 1953. “Cloze procedure”: A new tool for measuring readability. Journalism quarterly30, 4 (1953), 415–433

    1953

  38. [38]

    2022.Unlocking the Power of Sentence Embeddings with all-MiniLM- L6-v2

    Rahul Tiwari. 2022.Unlocking the Power of Sentence Embeddings with all-MiniLM- L6-v2. Accessed: 2025-07-29

    2022

  39. [39]

    Alexander Rogiers, Sander Noels, Maarten Buyl, and Tijl De Bie. 2024. Persuasion with large language models: a survey.arXiv preprint arXiv:2411.06837(2024)

    Pith/arXiv arXiv 2024

  40. [40]

    Silin Yang, Dong Wang, Haoqi Zheng, and Ruochun Jin. 2025. Timerag: Boosting llm time series forecasting via retrieval-augmented generation. InICASSP 2025- 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    2025

  41. [41]

    Yunkai Zhang, Yawen Zhang, Ming Zheng, Kezhen Chen, Chongyang Gao, Ruian Ge, Siyuan Teng, Amine Jelloul, Jinmeng Rao, Xiaoyuan Guo, et al. 2023. Insight miner: A large-scale multimodal model for insight mining from time series. In NeurIPS 2023 AI for Science Workshop

    2023

  42. [42]

    AAPL", "ABNB

    Tian Zhou, Peisong Niu, Xue Wang, Liang Sun, and Rong Jin. 2023. One fits all: Universal time series analysis by pretrained lm and specially designed adaptors. arXiv preprint arXiv:2311.14782(2023). A Prompts This appendix lists the full prompts used in this paper. Conference’17, July 2017, Washington, DC, USA Ria Mundhra, Gustavo Sato dos Santos, and Mic...

    Pith/arXiv arXiv 2023

  43. [44]

    Oskar Triebe, Hansika Hewamalage, Polina Pilyugina, Nikolay Laptev, Christoph Bergmeir, and Ram Rajagopal. 2021. NeuralProphet: Explainable Forecasting at Scale. arXiv:2111.15397 [cs.LG] https://arxiv.org/abs/2111.15397

    Pith/arXiv arXiv 2021

  44. [1990]

    STL: A seasonal-trend decomposition.J. off. Stat6, 1 (1990), 3–73

    1990

  45. [2019]

    InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)

    Comparing and developing tools to measure the readability of domain- specific texts. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 4831–4842

    2019

  46. [2023]

    Contrastive chain-of-thought prompting.arXiv preprint arXiv:2311.09277 (2023)

    Pith/arXiv arXiv 2023

  47. [2024]

    Language models still struggle to zero-shot reason about time series.arXiv preprint arXiv:2404.11757(2024)

    Pith/arXiv arXiv 2024