Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Semantic, multi-level news pairing consistently beats keyword matching for stock forecasting.

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 19:15 UTC pith:VILEIJV5

load-bearing objection A genuinely useful financial text–TS dataset with a sensible pairing pipeline, but the headline claim that semantic relevance beats keyword matching is confounded by text volume and needs a matched baseline. the 4 major comments →

arxiv 2603.02702 v3 pith:VILEIJV5 submitted 2026-03-03 cs.AI cs.LG

FinTexTS: Financial Text-Paired Time-Series Dataset via Semantic-Based and Multi-Level Pairing

Jaehoon Lee , Suhwan Park , Taeyoon Lim , Seunghan Lee , Jun Seo , Dongwan Kang , Hwanil Choi , Minjae Kim
show 4 more authors
Sungdong Yoo Soonyoung Lee Yongjae Lee Wonbin Ahn
This is my paper
classification cs.AI cs.LG
keywords financial forecastingtext-time-series datasetsemantic pairingmulti-level pairingstock price predictionembedding-based retrievalLLM classificationmultimodal time series
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.

This paper argues that the way text is paired with financial time series matters as much as the text itself. It proposes a pairing framework that uses SEC filings as company-specific context to retrieve news articles by meaning rather than by explicit company mentions, and classifies each article into one of four levels—macro, sector, related-company, and target-company—so a stock is paired with a structured, multi-level text summary each day. On a new dataset covering 100 companies over five years, the authors report that semantic pairing outperforms keyword pairing across all 12 forecasting models under both MSE and MAE, and that adding levels progressively improves accuracy. If correct, this gives the time-series and finance communities a better-aligned resource for text-based forecasting and a reusable recipe for building such datasets.

Core claim

The paper's central claim is that semantic-based and multi-level pairing of news articles with stock price time series yields consistently better stock-forecasting performance than the standard keyword-based approach. Using SEC filing content as a query, a fine-tuned embedding model retrieves company-relevant news even when the company name is absent; an LLM classifier then assigns each article to macro, sector, related-company, or target-company level, and an LLM summarizer condenses each level into distinct event categories. In a pilot study across 12 time-series models, the semantic condition outperforms the keyword condition on every model under both evaluation metrics, and progressive a

What carries the argument

The central mechanism is the pairing framework itself, built from four components: (1) an LLM-based SEC filing parser that extracts five categories of company context (overview, strategy, governance/risks, financials, recent events) and forward-fills them to daily level; (2) an LLM-based news classifier that first assigns articles to macro, sector, or company level and then refines sector-level articles into specific industry sectors and company-level articles into target vs. related company; (3) an embedding-based retrieval module that uses parsed filing components as queries and a sector-fine-tuned embedding model to retrieve the top-N relevant company-level articles per day; and (4) an LL

Load-bearing premise

The experiments assume the improved forecasts stem from semantic relevance and level structure rather than from the semantic condition's larger and more consistent daily text volume, since the keyword baseline leaves many days with no paired news and no matching-volume control is run.

What would settle it

A matched-volume experiment where the keyword baseline receives the same number of articles per day as the semantic condition—for example by sampling all articles mentioning the company plus a random same-sector sample—would settle the causal claim; if keyword pairing then matches or beats semantic pairing, the paper's central attribution fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Across 12 time-series architectures, the semantic condition lowers MSE and MAE relative to keyword pairing for every model—a direct result reported in the paper.
  • Adding macro, then sector, then related-company, then target-company text progressively lowers normalized MSE, and the same holds when levels are added in reverse order.
  • Backtesting with a simple equal-weighted top-20 portfolio gives higher cumulative return and Sharpe ratio than keyword-paired or text-free alternatives.
  • Fine-tuning the embedding model on sector classification improves retrieval hit-rate by 1.4–5.9 percentage points across ten sampled companies.
  • Applying the framework to a proprietary, well-curated news source improves forecasting over public news on most of the 12 models.

Where Pith is reading between the lines

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

  • A controlled comparison that matches per-day text volume between the keyword and semantic conditions would separate the effect of semantic relevance from the effect of simply having more text; the paper does not provide such an ablation.
  • The multi-level text structure could be exploited by level-specific attention or by predicting each level's marginal contribution, which the current pilot does not attempt.
  • The framework's reliance on SEC filings and U.S.-listed companies means the recipe may not transfer directly to markets with less standardized filings; extending it to other geographies would require an analogous authoritative context source.
  • The LLM-as-judge quality scores suggest the pairing improves coverage and diversity, so the approach could also be used as a data-cleaning step for other financial-text tasks such as sentiment analysis or event extraction.

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

4 major / 5 minor

Summary. The paper introduces FinTexTS, a financial text-paired time-series dataset covering 100 U.S. companies over 2019–2023, built by a semantic-based and multi-level pairing framework. Company-specific context is extracted from SEC filings with an LLM, news articles are retrieved via a fine-tuned embedding model, and articles are classified into macro, sector, related-company, and target-company levels using LLMs, followed by LLM-based summarization. The dataset is evaluated as text augmentation for stock price forecasting with 12 time-series models against a keyword-based FNSPID baseline, a no-text baseline, and a proprietary news variant. The paper claims that semantic-based pairing consistently outperforms keyword-based pairing and that multi-level text progressively improves forecasting, with additional backtesting and quality analyses.

Significance. If the central claims hold, FinTexTS is a substantial resource for financial text–time-series research: it is large-scale, publicly released with code and a fine-tuned embedding model, and the evaluation spans 12 diverse forecasting architectures with consistent directional gains. The proposed framework addresses a real limitation of keyword-based pairing and the multi-level structure is a useful organizing principle. The paper also ships machine-checked reproducibility artifacts (dataset URLs, code, HuggingFace model) and includes a backtest, which are strengths. However, the load-bearing attribution of gains to semantic relevance and multi-level structure is currently confounded with text volume and coverage, and several evaluation components rely on LLM-generated labels without external validation.

major comments (4)
  1. [§4.2, Table 1] The headline claim that 'semantic-based pairing consistently outperforms the keyword-based approach across all 12 forecasting models' is confounded with text volume and coverage. The keyword condition (FNSPID) pairs only articles with explicit company mentions, has many company-days with no news (handled by forward filling), and supplies at most one level of text. The semantic condition retrieves from ~1 million articles across four levels, resulting in far more and more consistently available text per day. No baseline matches the token count, article count, or coverage of the semantic condition. The observed gains may therefore be due to quantity or availability of text rather than semantic relevance. A volume-matched baseline is needed—e.g., keyword-paired text augmented with random or coverage-matched articles per day—to isolate the effect of semantic selection.
  2. [§4.3, Figure 4] The multi-level ablation progressively adds text levels, so each step increases the amount of text fed to the model. The monotonic improvement shown in Figure 4 is compatible with a simple volume effect: more text (regardless of its semantic level or relevance) may improve forecasts through increased information supply. To support the claim that multi-level structure itself is beneficial, the paper should compare against adding an equivalent volume of text distributed at random, or hold total token count roughly constant across configurations (e.g., by subsampling articles per level). Without such a control, the 'complementary signals' interpretation is not established.
  3. [§5.2, Table 3] The retrieval quality evaluation uses hit-rate against 'target company-level news article' labels that are themselves generated by the LLM classification pipeline described in Section 3.3. This creates an internal feedback loop: the ground truth is produced by the same type of LLM used to filter retrieved articles. The comparison between Base and Fine-Tuned embeddings is still informative, but the absolute hit-rate numbers and the claim of 'improved retrieval performance' need external validation—e.g., human-annotated relevance judgments on a sample, or at least a baseline such as random retrieval. Reporting the chance-level hit-rate would also help interpret the improvements.
  4. [§4.1] The paper states that z-score normalization is applied to stock prices, but it does not state whether the normalization statistics are computed on the training split only. In time-series forecasting with a temporal split, using full-sample statistics for normalization is a form of leakage that can inflate apparent performance. The authors should explicitly state that normalization is fit on the training set and applied to validation/test, or correct the pipeline if this was not the case. This is a load-bearing detail for the forecasting evaluation.
minor comments (5)
  1. [Tables 1 and 4] The paper reports averages over three random seeds but omits standard deviations or confidence intervals. Given the small differences between some conditions (e.g., DLinear, FiLM in Table 4), error bars are needed to assess whether the claimed improvements are meaningful. Please add per-seed standard deviations.
  2. [Table 3] The text says 'consistent improvements' across 10 companies, but COST shows +0.0%p improvement. This should be acknowledged, and significance testing (or paired tests) across companies would strengthen the claim.
  3. [§5.5] The LLM-as-judge evaluation compares paired texts that were generated with LLM-based summarization. The judge may favor summaries that match its own stylistic preferences. A small human-annotated sample for each criterion would increase confidence in the quality comparison.
  4. [Figure 5] The x-axis label shows '0 10 15 20'; it would be clearer to state that N=0 means no company-level news is used, not zero retrieval size in the standard sense. Also, the caption references Figure 4 for normalization but the procedure is slightly different (across N settings); please clarify.
  5. [§4.2, Table 2] The backtesting experiment is a nice addition, but the portfolio construction (top 20 by predicted 3-day returns, rebalancing every 3 days) should specify transaction costs, liquidity considerations, and whether returns are gross or net. Also, the 2023 evaluation period is a single year; a multi-year backtest would be more robust.

Circularity Check

0 steps flagged

No significant circularity: the headline forecasting result rests on an external benchmark; the LLM-based component evaluations are self-referential but not derivationally circular.

full rationale

The paper's central claim, that semantic-based and multi-level pairing improves stock forecasting, is tested on an external forecasting benchmark (Section 4.2, Table 1) using 12 models and MSE/MAE on normalized OHLCV data. The comparison against FNSPID keyword pairing is an empirical comparison, not a derivation: the semantic condition differs from the keyword condition in text volume and coverage, but that is a confounding/validity threat, not a circular reduction. No equation defines the semantic condition in terms of the forecasting outcome, and no fitted parameter is renamed as a prediction in the headline result. The multi-level analysis (Figure 4) and retrieval-size sensitivity (Figure 5) are likewise empirical ablations. The internal validation components (Section 5.2 hit-rate against LLM-generated target labels; Section 5.5 LLM-as-judge on LLM-produced summaries) are self-referential in that ground truth and quality judgments come from the same LLM pipeline used to construct the data. This limits external validity but does not make any stated prediction equivalent to its inputs by construction: the fine-tuned embedding is trained on sector labels while hit-rate is measured on target-company labels, and the LLM judge is a separate evaluation instrument. Therefore no circular step meets the quoted-evidence bar. The volume confound and the absence of a stated train-only normalization should be treated as correctness/reproducibility risks, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The framework introduces a small set of hand-chosen configuration values and relies on several domain assumptions about financial news impact and LLM output quality. No invented entities are introduced.

free parameters (3)
  • retrieval size N = 10
    Selected from {10, 15, 20} based on test-set normalized MSE and cost; directly controls how much company-level news is paired.
  • summarization category counts = macro/sector: 5, company: 3
    Set manually to reflect news volume; affects the text representation and downstream forecasting performance.
  • embedding fine-tuning hyperparameters = 10 epochs, batch size 64, 50k samples
    Chosen by hand for the contrastive retrieval model; not tuned against the forecasting benchmark.
axioms (4)
  • domain assumption News about macro, sector, and related companies materially affects a target company's stock price beyond target-company-specific news
    This motivates multi-level pairing; if false, adding non-target text would not help forecasting. Invoked in Sections 1 and 3.
  • domain assumption LLM classifications (macro/sector/company and target vs related) are sufficiently accurate for retrieval and evaluation
    No human validation is reported; LLM-based labels are used both to train the retrieval model and to evaluate retrieval quality (Sections 5.2 and 5.5).
  • domain assumption SEC filing parsed content provides a reliable query for retrieving company-relevant news
    The parsed filing components are used as queries for embedding-based retrieval (Section 3.3); if they are stale or noisy, pairing quality drops.
  • domain assumption Linq-Embed-Mistral and SBERT embeddings capture financial semantic similarity after fine-tuning
    The retrieval and text-fusion steps assume these embeddings preserve financial relevance; no formal or human benchmark demonstrates this.

pith-pipeline@v1.3.0-alltime-deepseek · 15390 in / 9808 out tokens · 93340 ms · 2026-08-02T19:15:57.235031+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of FinTexTS: Financial Text-Paired Time-Series Dataset via Semantic-Based and Multi-Level Pairing." pith.science (2026). https://pith.science/paper/VILEIJV5

@misc{pith2026260302702,
  author       = {Pith},
  title        = {Pith review of: FinTexTS: Financial Text-Paired Time-Series Dataset via Semantic-Based and Multi-Level Pairing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VILEIJV5}},
  note         = {Machine review of arXiv:2603.02702}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The financial domain involves a variety of important time-series problems. Recently, time-series analysis methods that jointly leverage textual and numerical information have gained increasing attention. Accordingly, numerous efforts have been made to construct text-paired time-series datasets in the financial domain. However, financial markets are characterized by complex interdependencies, in which a company's stock price is influenced not only by company-specific events but also by events in other companies and broader macroeconomic factors. Existing approaches that pair text with financial time-series data based on simple keyword matching often fail to capture such complex relationships. To address this limitation, we propose a semantic-based and multi-level pairing framework. Specifically, we extract company-specific context for the target company from SEC filings and apply an embedding-based matching mechanism to retrieve semantically relevant news articles based on this context. Furthermore, we classify news articles into four levels (macro-level, sector-level, related company-level, and target company-level) using large language models (LLMs), enabling multi-level pairing of news articles with the target company. Applying this framework to publicly-available news datasets, we construct FinTexTS, a new large-scale text-paired stock price dataset. Experimental results on FinTexTS demonstrate the effectiveness of our semantic-based and multi-level pairing strategy in stock price forecasting. In addition to publicly-available news underlying FinTexTS, we show that applying our method to proprietary yet carefully curated news sources leads to higher-quality paired data and improved stock price forecasting performance.

Figures

Figures reproduced from arXiv: 2603.02702 by Dongwan Kang, Hwanil Choi, Jaehoon Lee, Jun Seo, Minjae Kim, Seunghan Lee, Soonyoung Lee, Suhwan Park, Sungdong Yoo, Taeyoon Lim, Wonbin Ahn, Yongjae Lee.

Figure 1
Figure 1. Figure 1: Overview of the proposed semantic-based and multi-level pairing framework. Given a target company and a date, the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Keyword-based vs. Semantic-based pairing method. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: LLM prompts used in our framework. Full version of LLM prompts are available in Appendix A. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of Multi-Level Text Pairing on Forecasting Performance. In the left figure, textual information is progressively [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis of the retrieval size [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of multi-level text on AMD (Advanced Micro [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Multi-level paired text examples for AMD, which is a U.S. semiconductor company that designs high-performance [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt for LLM-based SEC filing parser. • Materials: Companies producing raw materials such as metals, chemicals, and construction materials. • Industrials: Firms involved in manufacturing, transportation, aerospace, construction, and logistics. • Consumer Discretionary: Non-essential consumer goods and services such as automobiles, retail, and leisure. • Consumer Staples: Essential everyday products inclu… view at source ↗
Figure 9
Figure 9. Figure 9: Prompt for LLM-based news classification. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Prompt for LLM-based news summarization. [PITH_FULL_IMAGE:figures/full_fig_p012_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

63 extracted references · 5 linked inside Pith

  1. [1]

    Ethan Callanan, Amarachi Mbakwe, Antony Papadimitriou, Yulong Pei, Mathieu Sibue, Xiaodan Zhu, Zhiqiang Ma, Xiaomo Liu, and Sameena Shah. 2024. Can GPT models be Financial Analysts? An Evaluation of ChatGPT and GPT-4 on mock CFA Exams. InFinNLP-AgentScen Workshop @ IJCAI

    2024

  2. [2]

    Arik, and Tomas Pfister

    Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O. Arik, and Tomas Pfister. 2023. TSMixer: An All-MLP Architecture for Time Series Forecasting.Transactions on Machine Learning Research(2023)

    2023

  3. [3]

    Chanyeol Choi, Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, and Jy-yong Sohn. 2024. Linq-Embed-Mistral Technical Report. arXiv:2412.03223 [cs.CL] https://arxiv.org/abs/2412.03223

    Pith/arXiv arXiv 2024

  4. [4]

    Abhimanyu Das, Weihao Kong, Andrew Leber, Rajat Mathur, Rajat Sen, and Rose Yu. 2023. Long-term Forecasting with TiDE: Time-series Dense Encoder. Transactions on Machine Learning Research(2023)

    2023

  5. [5]

    Zihan Dong, Xinyu Fan, and Zhiyuan Peng. 2024. FNSPID: A Comprehensive Financial News Dataset in Time Series. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24). Association for Computing Machinery, New York, NY, USA, 4918–4927. doi:10.1145/3637528. 3671629

    doi:10.1145/3637528 2024

  6. [6]

    Sofia Giantsidi and Claudia Tarantola. 2025. Deep learning for financial forecast- ing: A review of recent trends.International Review of Economics & Finance104 (2025), 104719. doi:10.1016/j.iref.2025.104719

    arXiv 2025

  7. [7]

    Yoontae Hwang, Junhyeong Lee, Daham Kim, Seunghwan Noh, Joohwan Hong, and Yongjae Lee. 2023. SimStock: Representation Model for Stock Similarities. In Proceedings of the Fourth ACM International Conference on AI in Finance (ICAIF ’23)

    2023

  8. [8]

    Alex Kim, Maximilian Muhn, and Valeri Nikolaev. 2024. Financial Statement Analysis with Large Language Models. arXiv:2407.17866 [q-fin.ST] https://arxiv. org/abs/2407.17866

    Pith/arXiv arXiv 2024

  9. [9]

    Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The Efficient Transformer. InInternational Conference on Learning Representations

    2020

  10. [10]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. InInternational Conference on Machine Learning. PMLR, 12888– 12900

    2022

  11. [11]

    Tong Li, Zhaoyang Liu, Yanyan Shen, Xue Wang, Haokun Chen, and Sen Huang

  12. [12]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruc- tion Tuning. InAdvances in Neural Information Processing Systems, Vol. 36

    2023

  13. [13]

    Haoxin Liu, Shangqing Xu, Zhiyuan Zhao, Lingkai Kong, Harshavardhan Ka- marthi, Aditya B Sasanur, Megha Sharma, Jiaming Cui, Qingsong Wen, Chao Zhang, and B Aditya Prakash. 2024. Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysis. InAdvances in Neural Information Processing Systems, Vol. 37

    2024

  14. [14]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. InInternational Conference on Learning Representations

    2024

  15. [15]

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting. InAdvances in Neural Information Processing Systems, Vol. 35. 9881–9893

    2022

  16. [16]

    Alejandro Lopez-Lira and Yuehua Tang. 2023. Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models. arXiv:2304.07619 [q-fin.ST] https://arxiv.org/abs/2304.07619

    arXiv 2023

  17. [17]

    Rui Luo, Weinan Zhang, Xiaojun Xu, and Jun Wang. 2018. A Neural Stochastic Volatility Model. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32. 6401–6408

    2018

  18. [18]

    Kei Nakagawa, Tomoki Ito, Masaya Abe, and Kiyoshi Izumi. 2019. Deep Recurrent Factor Model: Interpretable Non-Linear and Time-Varying Multi-Factor Model. InAAAI-19 Workshop on Network Interpretability for Deep Learning

    2019

  19. [19]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In International Conference on Learning Representations

    2023

  20. [20]

    OpenAI. [n. d.]. OpenAI API Reference. https://platform.openai.com/docs/api- reference

  21. [21]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  22. [22]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. arXiv:1908.10084 [cs.CL] https://arxiv.org/abs/ 1908.10084

    Pith/arXiv arXiv 2019

  23. [23]

    Jun Seo et al. 2025. Adaptive Information Routing for Multimodal Time Series Forecasting.arXiv preprint arXiv:2512.10229(2025)

    arXiv 2025

  24. [24]

    Emi Soroka and Artem Arzyn. 2026. Data-Efficient Realized Volatility Forecasting with Vision Transformers. InNeurIPS 2025 Workshop: Generative AI in Finance. https://openreview.net/forum?id=0bxRD79zv1

    2026

  25. [25]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. InAdvances in Neural Information Processing Systems, Vol. 30

    2017

  26. [26]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. InAdvances in Neural Information Processing Systems, Vol. 34. 22419– 22430

    2021

  27. [27]

    Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, et al . 2024. FinBen: A Holistic Financial Benchmark for Large Language Models. InAdvances in Neural Information Processing Systems, Vol. 37

    2024

  28. [28]

    Yumo Xu and Shay B Cohen. 2018. Stock Movement Prediction from Tweets and Historical Prices. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1970–1979

    2018

  29. [29]

    Wenbo Yan and Ying Tan. 2024. Double-Path Adaptive-correlation Spatial-Temporal Inverted Transformer for Stock Time Series Forecasting. arXiv:2409.15662 [cs.LG] https://arxiv.org/abs/2409.15662

    Pith/arXiv arXiv 2024

  30. [30]

    Jaemin Yoo, Yejun Soun, Yong-chan Park, and U Kang. 2021. Accurate Multi- variate Stock Movement Prediction via Data-Axis Transformer with Multi-Level Contexts. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2037–2045

    2021

  31. [31]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are Transformers Effective for Time Series Forecasting?. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 11121–11128

    2023

  32. [32]

    Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer Utilizing Cross- Dimension Dependency for Multivariate Time Series Forecasting. InInternational Conference on Learning Representations

    2023

  33. [33]

    Lifan Zhao, Shuming Kong, and Yanyan Shen. 2024. DoubleAdapt: A Meta- learning Approach to Incremental Learning for Stock Trend Forecasting. arXiv:2306.09862 [q-fin.ST] https://arxiv.org/abs/2306.09862

    Pith/arXiv arXiv 2024

  34. [34]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 11106–11115

    2021

  35. [35]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, and Rong Jin. 2022. FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting. InAdvances in Neural Information Processing Systems, Vol. 35. 12677–12690

    2022

  36. [36]

    Zhuohang Zhu, Haodong Chen, Qiang Qu, and Vera Chung. 2025. FinCast: A Foundation Model for Financial Time-Series Forecasting. InProceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM). A LLM Prompts Used in Our Framework In this section, we present the LLM prompts used in our frame- work. Figure 8 shows the pro...

    2025

  37. [39]

    overviewProduct: Provides a high-level summary of the company's business, including its mission, core products or services, key customer groups, business segments, and primary geographic markets

  38. [40]

    strategyMarketOps: Describes the company's strategic direction and competitive strengths (e.g., proprietary technology, intellectual property, and regulatory expertise), along with its target markets, regulatory context, and operating model such as manufacturing footprint, supply chain structure, and major partnerships

  39. [41]

    Risk factors are organized by category (e.g., regulatory, market, operational, and cybersecurity) and may include both long-term structural risks and near-term concerns

    governanceRisks: Covers the company's governance framework, including notable changes in leadership or the board, and summarizes key risks disclosed in filings. Risk factors are organized by category (e.g., regulatory, market, operational, and cybersecurity) and may include both long-term structural risks and near-term concerns

  40. [42]

    financialStatement: Summarizes the company's financial statements with key figures, and provides an interpretation of its financial health based on filings-discussing major performance drivers, liquidity, funding and capital resources, capital allocation decisions, accounting updates, and material obligations

  41. [43]

    recentEventCatalyst: Highlights significant developments from roughly the past 12 months, including changes to earnings outlook, major product releases, regulatory decisions, M\&A progress, leadership updates, and other events that could meaningfully affect market perception or performance. # Rules

  42. [44]

    Focus on the content rather than the format

    The SEC filing can be long and may contain noisy formatting. Focus on the content rather than the format

  43. [45]

    Only use information explicitly stated in the SEC filing

    Do not hallucinate. Only use information explicitly stated in the SEC filing

  44. [46]

    # Output Format

    If there is no relevant information for a category, return an empty string for that category. # Output Format

  45. [47]

    overviewProduct

    Example: { "overviewProduct": ..., "strategyMarketOps": ..., "financialStatement": ..., "governanceRisks": ..., "recentEventCatalyst": ... } # SEC Filing [sec_filing_text] Figure 8: Prompt for LLM-based SEC filing parser. • Materials:Companies producing raw materials such as metals, chemicals, and construction materials. • Industrials:Firms involved in ma...

    2026

  46. [48]

    # COMPANY_PROFILE (provided only for company-level classification into Target vs

    N/A: The article does not fit any of the above categories. # COMPANY_PROFILE (provided only for company-level classification into Target vs. Related Companies)

  47. [49]

    Financial Statement: [FINANCIAL_STATEMENT]

  48. [50]

    Governance Risks: [GOVERNANCE_RISKS]

  49. [51]

    Overview Product: [OVERVIEW_PRODUCT]

  50. [52]

    Recent Event Catalyst: [RECENT_EVENT_CATALYST]

  51. [53]

    Strategy Market Ops: [STRATEGY_MARKET_OPS] # Rules

  52. [54]

    # Output

    Please ignore any unusual or inconsistent formatting in the article. # Output

  53. [55]

    category

    Example: {"category": ...}

  54. [56]

    category

    "category" must be one of [CATEGORY 1], [CATEGORY 2], [CATEGORY 3], or N/A. # Article Headline: [HEADLINE] Body: [BODY] Figure 9: Prompt for LLM-based news classification. # Main Instruction You will be given a list of multiple ARTICLES that may impact [TAG]. Your task is to review all ARTICLES and identify up to [N] key categories of significant events t...

  55. [57]

    Please ignore any unusual or inconsistent formatting in the ARTICLES

  56. [58]

    Select a category only if it has a meaningful impact on [TAG]

  57. [59]

    If more than [N] categories are identified, select only the [N] most important ones

  58. [60]

    Each category must address one single, distinct topic only

  59. [61]

    Write only based on the given ARTICLES

    Do not hallucinate. Write only based on the given ARTICLES. # Output Format

  60. [62]

    category1

    Example: { "category1": ..., "category2": ..., ... } # Articles

  61. [63]

    Headline: [HEADLINE_1] 1-2

    Article 1: 1-1. Headline: [HEADLINE_1] 1-2. Body: [BODY_1] ... Figure 10: Prompt for LLM-based news summarization. Received 8 February 2026; accepted 16 May 2026

    2026

  62. [2021]

    InInternational Conference on Machine Learning

    Learning Transferable Visual Models from Natural Language Supervision. InInternational Conference on Machine Learning. PMLR, 8748–8763

  63. [2024]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    MASTER: Market-Guided Stock Transformer for Stock Price Forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 162–170