Integrating Large Language Models in Financial Investments and Market Analysis: A Survey

Jiating (Kristin) Chen; Lina Huertas Guativa; Pradeep Kumar Joshi; Sedigheh Mahdavi; Upmanyu Singh

REVIEW 4 major objections 5 minor 2 cited by

Integrating Large Language Models in Financial Investments and Market Analysis: A Survey

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This survey argues that large language models have become a practical force in financial decision-making, organizing recent research into four frameworks and concluding that LLM-based investment systems show promising results across stock…

desk verdict A useful but sloppy survey of LLM finance: the four-part taxonomy helps newcomers, but the citation errors and unverified performance numbers need cleanup before I'd trust it. read the letter →

arxiv 2507.01990 v1 pith:ZE3QFEGA submitted 2025-06-29 q-fin.GN cs.AIcs.LG

Sedigheh Mahdavi , Jiating (Kristin) Chen , Pradeep Kumar Joshi , Lina Huertas Guativa , Upmanyu Singh This is my paper

classification q-fin.GNcs.AIcs.LG

keywords largelanguagemodelsfinancialdecision-makingstockmarketpredictionsentimentanalysismulti-agentsystemsportfoliooptimizationfine-tuningretrieval-augmentedgeneration

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This survey tries to establish that large language models have become a practical ingredient in financial investment work, not just a laboratory curiosity. It organizes recent research into four frameworks—LLM-based pipelines, hybrid methods that pair LLMs with classical models, fine-tuned financial LLMs, and multi-agent systems—and argues that each has shown real capability in stock selection, sentiment analysis, risk assessment, trading, and forecasting. A sympathetic reader would care because the survey synthesizes a fast-moving literature into one map and extracts the field's own performance evidence, such as reported cumulative returns of 72% for MarketSenseAI, 125.9% for MarketSenseAI 2.0, and a 30.8% annualized return for Stock-Chain. The conclusion the paper is trying to establish is that this integration improves decision quality and predictive accuracy enough to matter for practice.

What carries the argument

The object that carries the survey is the four-category taxonomy itself, together with the enabling techniques that the reviewed systems reuse: retrieval-augmented generation (RAG), which pulls current or proprietary documents into the model's context; chain-of-thought (CoT) prompting, which makes the model reason step by step; in-context learning (ICL), which adapts the model to a new task from a few prompt examples; low-rank adaptation (LoRA) and parameter-efficient fine-tuning (PEFT), which update a small trainable subset of weights; reinforcement learning with human feedback (RLHF); and mixture-of-experts (MoE) routing. The taxonomy organizes the field, and the technique list explains how the reported gains are achieved—each category is a different way of getting textual and numerical data into an LLM and extracting a trading or rating decision from it.

What would settle it

Take the headline systems from Tables 1–4 and re-run their backtests out of sample with transaction costs, slippage, and significance testing; if the reported 72% MarketSenseAI return, the 125.9% MarketSenseAI 2.0 return, and the 30.8% Stock-Chain annualized return do not survive in most cases, the survey's conclusion that LLM integration is broadly promising loses its empirical footing.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that LLMs enhance financial decision-making across the board, and that the current research landscape can be productively read through a four-way taxonomy. The survey classifies roughly thirty recent systems into LLM-based frameworks and pipelines (MarketSenseAI, Ploutos, GPT-InvestAR, LLMoE), hybrid methods that combine LLMs with classical optimization or deep learning (ChatGPT-based portfolio selection, MuSA, SEP, LLM-augmented Transformer–CNN), fine-tuning and adaptation approaches (fine-tuned Llama 2 and GPT-3.5, FinLlama, Stock-Chain, StockTime, SAPPO), and agent-based architectures (FINCON, TradingAgents, StockAgent, Alpha-GPT 2.0, TwinMarket, MarketSenseAI 2.0). It argues that the common thread is the ability of LLMs to fuse unstructured text—news, filings, social media, earnings calls—with numerical market data in real time. The evidence it forwards is the reported performance of these systems: excess alpha of 10–30% for MarketSenseAI, 125.9% cumulative returns versus 73.5% for the S&P 100 index for MarketSenseAI 2.0, a 30.8% annualized return for Stock-Chain, and a 30.2% annualized return with a 1.90 Sharpe ratio for SAPPO. The paper's conclusion is that these results, taken together, show LLM integration to be a promising direction for investment strategy.

Load-bearing premise

The load-bearing premise is that the performance numbers the survey repeats from the papers it summarizes—like 72%, 125.9%, and 30.8% returns—are accurate and representative, because the survey itself does not verify them or adjust for backtest bias, transaction costs, or overfitting.

Editorial extensions

If this is right

If the survey's reading is correct, LLM-based systems have moved from text analysis to end-to-end investment signals: raw news and filings in, buy/sell/hold decisions out.
The reported numbers imply that LLM-driven pipelines can beat conventional benchmarks in backtests, with MarketSenseAI 2.0 reporting 125.9% cumulative returns against 73.5% for the S&P 100 index.
Fine-tuning and adaptation approaches are the cheapest route to domain competence, since LoRA and PEFT update a small fraction of weights instead of retraining.
Multi-agent systems are the emerging frontier: FINCON, TradingAgents, StockAgent, and TwinMarket all put several specialized LLM agents in one decision loop, suggesting that collaboration is where the field expects the next gains.
A practical corollary is that standardized benchmarking is the missing piece: the paper itself lists comprehensive benchmarking as a future direction, because cross-paper comparisons are currently hard.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The performance figures the survey repeats are the authors' own backtest claims; a reader should treat them as evidence of promise, not as audited results, since the survey does not adjust for transaction costs, slippage, or overfitting.
The taxonomy suggests a natural test: systems in the agent-based category should dominate in settings requiring many information sources, while fine-tuned single models should dominate in narrow, well-labeled tasks—an empirical comparison the survey does not run.
If the trend the survey maps continues, the bottleneck will shift from model capability to evaluation infrastructure: shared datasets, out-of-sample protocols, and risk-adjusted metrics that let one framework be compared fairly with another.
One implicit consequence is that retail investors, not only institutions, are the intended beneficiaries of several reviewed systems, which points toward a future of personalized LLM advisors aligned to individual risk preferences.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The main value here is organizational. The paper sorts recent LLM-finance work into four buckets—pipelines, hybrids, fine-tuning, and agent systems—and the accompanying tables give a quick map of who did what with which model and data. That is genuinely useful for someone entering the field. I also appreciate that it names the key techniques (RAG, CoT, ICL, LoRA, PEFT) and ties them to concrete papers. As a survey it does the basics well.

The soft spots are real but not fatal. The citation problems are verifiable: [39] is the CLIP paper, cited for RLHF; [33] is Jacobs 1991, the original MoE paper, which is not a good citation for modern MoE scaling; and [31] and [48] are the same paper listed twice. Those are sloppy mistakes that a careful revision should fix. More substantive is the lack of a stated paper-selection methodology. The survey never explains how the literature was gathered or filtered, so I can't tell whether the picture is representative or cherry-picked.

The stress-test concern about performance numbers holds up. Section 3 and Tables 1-4 transcribe claims like MarketSenseAI's 72% return and SAPPO's 30.2% annualized return without any caveat about backtest overfitting, transaction costs, or benchmark comparability. The conclusion's assertion of 'significant promise' rests entirely on those numbers. This is common in surveys, but the paper should say explicitly that it is reporting the authors' claims, not endorsing them, and should flag the usual pathologies. Right now it reads as if every backtest was sound.

That said, the paper is coherent on its own terms, and for a newcomer it provides a reasonable entry point. I would not cite it in my own work given the accuracy issues, but it could be a starting point for a reading group discussion about evaluation standards in this literature.

I'd send it to peer review. The field needs maps like this, and the problems are fixable. A serious referee should demand a methodology section, an accuracy pass on the citations, and a paragraph acknowledging the limits of the reported performance numbers.

Referee Report

4 major / 5 minor

Summary. This survey organizes recent research on applying LLMs to financial investment and market analysis into four categories: LLM-based frameworks and pipelines, hybrid integration methods, fine-tuning and adaptation approaches, and agent-based architectures. Sections 2 and 3 describe technical background and summarize roughly thirty papers, with tables listing models, data, and applications. The conclusion asserts that LLM integration in finance has shown 'significant promise' in improving decision-making, predictive accuracy, and actionable insights, and lists future research directions.

Significance. If the survey's organizing taxonomy and reported performance trends are reliable, the paper could serve as a useful entry point for researchers and practitioners mapping the LLM-finance landscape. The paper correctly names the main technical families used in this literature: RAG, CoT, ICL, PEFT/LoRA, MoE, and RLHF, and it compiles a broad bibliography across 2023-2025 work, including several preprint and workshop items. The four-way taxonomy is a sensible first-order partition, and the survey is presumably the first to assemble these particular papers into a single structured review. However, the paper's value as a survey is limited by its lack of methodological transparency (no search or inclusion criteria), its uncritical transcription of performance claims, and several verifiable reference errors. These issues do not invalidate the taxonomy, but they weaken the evidentiary base for the paper's central positive conclusion.

major comments (4)

[Section 4 and Tables 1-4] The conclusion's load-bearing claim that LLMs have shown 'significant promise' rests on performance figures transcribed from primary papers (e.g., MarketSenseAI 72% cumulative returns, MarketSenseAI 2.0 125.9%, Stock-Chain ARR 30.8%, SAPPO 30.2% and Sharpe 1.90, FinLlama 35% long-short returns). The survey reports these numbers without any stated inclusion criteria, without risk-adjustment or backtest-overfitting checks, without discussion of transaction costs or survivorship bias, and without acknowledging heterogeneous test periods and asset universes. As a result, the central positive conclusion is not supported by a critically vetted evidence base. Please add a limitations subsection that explicitly flags these caveats, and soften the conclusion to say that reported results are promising but not independently verified or directly comparable.
[Section 2.6, reference [39]] The RLHF discussion cites reference [39], which is Radford et al. (2021), 'Learning transferable visual models from natural language supervision' (CLIP). That paper does not introduce or describe RLHF. The appropriate citation would be Christiano et al. [40] (which is already listed) and possibly Stiennon et al. [41] for learning-to-summarize-from-human-feedback. This is a factual citation error in a core background section, and it undermines reader confidence in the survey's references; please correct it.
[Section 3, 'Materials and Methods'] The section is titled 'Materials and Methods,' but it describes no literature search strategy, no inclusion/exclusion criteria, no date-range specification, and no quality filter. For a survey whose organizing contribution is a taxonomy, the absence of a reproducible selection methodology makes it impossible for a reader to know whether the reviewed set is representative or biased. Please add a short methods paragraph describing how the papers were collected, screened, and assigned to the four categories.
[Section 3.1-3.4, taxonomy definitions] The four-way taxonomy is central to the paper, but the boundaries between categories are not operationalized. For example, Stock-Chain [27] is placed under Fine-Tuning despite being primarily a RAG-based framework, while MarketSenseAI 2.0 [23] and Ploutos [4] are placed under Frameworks/Pipelines and Agent-Based/other categories despite all three involving multi-component pipelines. Please provide explicit criteria for category assignment (e.g., what distinguishes a 'framework' from an 'agent-based architecture'), and either reclassify borderline entries or justify the current assignments.

minor comments (5)

[References] References [31] and [48] are the same paper (Lee, Choi, and Kwon, 'Quantifying Qualitative Insights: Leveraging LLMs to Market Predict,' arXiv:2411.08404); one duplicate entry should be removed and the in-text citations reconciled.
[Table 4, row 30] The table entry 'gpt-40-mini and gpt-4o' is a typo: it should be 'gpt-4o-mini and gpt-4o.' The same typo appears in the text description of TradingAgents.
[Section 3.4, paragraph on Henning et al.] The citation cluster '[46][48]' for Thomas Henning et al.'s experimental asset market paper includes an incorrect reference; the work is [46], and [48] is the duplicated Lee et al. paper.
[General, tables] The tables are not consistently filled: Table 4 lists '-' in the data column for [46], and several rows lack the year or venue information that would help readers locate the primary studies. A uniform citation style in tables would improve usability.
[Section 3.3, StockTime] The text says StockTime's LLM 'was not fine-tuned in this approach; instead, only the embedding and projection layers were trained,' which is a form of parameter-efficient fine-tuning. The presentation would be clearer if it explicitly acknowledged this as a PEFT-style approach rather than contrasting it with fine-tuning.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation chain: the survey's conclusion is an aggregation of externally reported performance figures, and the only self-citation is a summarized source that is not load-bearing.

full rationale

This paper is a literature survey. It contains no fitted parameters, no predictive model of its own, and no equations, so there is no derivation chain whose output could be equivalent to its input by construction. The concluding claim in Section 4 that 'the integration of LLMs into financial investment strategies has shown significant promise' is supported by transcribing performance numbers from cited works, such as MarketSenseAI's 72% cumulative returns, Stock-Chain's 30.8% ARR, and SAPPO's 30.2% annualized return. Those numbers are not derived from the survey's taxonomy or from any assumption in this paper; they are external literature claims. Building the four-category taxonomy from the same papers that are then summarized is normal review practice and does not make the organizational scheme circular. The one self-citation, reference [32] (Joshi, 2025), shares an author with this survey and is summarized in Section 3.2 as a prompt-engineering study for financial market integrity and risk management, but the survey's overall conclusion does not rest on that specific study, and no uniqueness theorem or fitted quantity is imported from it. Concerns about unverified, non-comparable, or possibly overfit benchmark figures are verification and correctness risks, not circularity, and would not change the circularity score.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No mathematical derivation, fitted parameters, or invented entities appear. The paper is a literature synthesis, so its ledger consists of assumptions about the reliability of cited results and the validity of its informal taxonomy.

assumptions (2)

domain assumption Reported performance numbers in the surveyed papers are accurate and representative.
Section 3 repeats figures such as MarketSenseAI's 72% cumulative return and Stock-Chain's 30.8% ARR without verification or caveat about backtest methodology.
ad hoc to paper The four-category taxonomy (frameworks/pipelines, hybrid, fine-tuning, agent-based) is a faithful partition of current work.
The Introduction asserts this categorization; no systematic clustering or search methodology justifies it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Integrating Large Language Models in Financial Investments and Market Analysis: A Survey." pith.science (2026). https://pith.science/paper/ZE3QFEGA

@misc{pith2026250701990,
  author       = {Pith},
  title        = {Pith review of: Integrating Large Language Models in Financial Investments and Market Analysis: A Survey},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZE3QFEGA}},
  note         = {Machine review of arXiv:2507.01990}
}

read the original abstract

Large Language Models (LLMs) have been employed in financial decision making, enhancing analytical capabilities for investment strategies. Traditional investment strategies often utilize quantitative models, fundamental analysis, and technical indicators. However, LLMs have introduced new capabilities to process and analyze large volumes of structured and unstructured data, extract meaningful insights, and enhance decision-making in real-time. This survey provides a structured overview of recent research on LLMs within the financial domain, categorizing research contributions into four main frameworks: LLM-based Frameworks and Pipelines, Hybrid Integration Methods, Fine-Tuning and Adaptation Approaches, and Agent-Based Architectures. This study provides a structured review of recent LLMs research on applications in stock selection, risk assessment, sentiment analysis, trading, and financial forecasting. By reviewing the existing literature, this study highlights the capabilities, challenges, and potential directions of LLMs in financial markets.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Talking to Digital Twins: Selective Disclosure and Belief Measurement in Financial Social Media
econ.GN 2026-08 conditional novelty 7.0 of 10

Daily real-time LLM digital-twin interviews of finfluencer accounts predict cross-sectional large-cap returns over the next ten trading days, mainly in the silent region with no concurrent public post.
AI Trading: Evaluating Large Language Models for Technical Market Analysis
cs.LG 2026-07 reject novelty 4.0 of 10

A comparative evaluation claims GPT-4 Turbo and FinGPT outperformed the S&P 500 in a 2023 simulated backtest, but flawed baselines and missing code/data undermine the result.

Reference graph

Works this paper leans on

47 extracted references · 26 canonical work pages · cited by 2 Pith papers

[39]

W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., et al

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., et al. (2021). Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (pp. 8748-8763). PMLR

work page 2021
[33]

A., Jordan, M

Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation , 3(1), 79-87

work page 1991
[48]

figure1_llm_models.png

Lee, H., Choi, Y., & Kwon, Y. (2024). Quantifying Qualita tive Insights: Leveraging LLMs to Market Predict. arXiv preprint arXiv:2411.08404 . This figure "figure1_llm_models.png" is available in "png" format from: http://arxiv.org/ps/2507.01990v1

arXiv 2024
[40]

F., Leike, J., Brown, T., Martic, M., Legg , S., & Amodei, D

Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg , S., & Amodei, D. (2017). Deep reinforcement learning from human preferences. Advances in Neural Information Processing Systems

work page 2017
[41]

Liu, F. (2020). Learning to summarize from human feedbac k. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 583-592)

work page 2020
[27]

Li, X., Li, Z., Shi, C., Xu, Y., Du, Q., Tan, M., Huang, J., & Lin, W. (2024). Alphaﬁn: Benchmarking ﬁnancial analysis with retrieval-augmented stock-chain f ramework. arXiv preprint arXiv:2403.12582

arXiv 2024
[23]

Fatouros, G., Metaxas, K., Soldatos, J., & Karathanassi s, M. (2025). MarketSenseAI 2.0: Enhancing Stock Analysis through LLM Agents. arXiv preprint arXiv:2502.00415

arXiv 2025
[4]

Tong, H., Li, J., Wu, N., Gong, M., Zhang, D., & Zhang, Q. (20 24). Ploutos: Towards interpretable stock movement prediction with ﬁnancial large language mod el. arXiv preprint arXiv:2403.00782

arXiv

Show all 47 references

[1]

Fatouros, G., Metaxas, K., Soldatos, J., & Kyriazis, D. (2 024). Can large language models beat wall street? Unveiling the potential of AI in stock selection. arXiv preprint arXiv:2401.03737

arXiv
[2]

Romanko, O., Narayan, A., & Kwon, R. H. (2023). Chatgpt-ba sed investment portfolio selection. In Operations Research Forum (Vol. 4, No. 4, p. 91). Cham: Springer International Publish ing

2023
[3]

Swamy, M., Shukla, A., & Purtilo, J. (2023). LLM-based sto ck market trend prediction

2023
[5]

Gupta, U. (2023). GPT-InvestAR: Enhancing stock investm ent strategies through annual report analysis with large language models. arXiv preprint arXiv:2309.03079

2023 arXiv
[6]

Han, X., Wang, N., Che, S., Yang, H., Zhang, K., & Xu, S. X. (2 024). Enhancing investment analysis: Optimizing AI-Agent collaboration in ﬁnancial research. I n Proceedings of the 5th ACM International Conference on AI in Finance (pp. 538-546)
[7]

M., & Lo, M

Liu, K. M., & Lo, M. C. (2025). LLM-Based Routing in Mixture of Experts: A Novel Framework for Trading. arXiv preprint arXiv:2501.09636

2025 arXiv
[8]

Guo, T., & Hauptmann, E. (2024). Fine-tuning large langua ge models for stock return prediction using newsﬂow. arXiv preprint arXiv:2407.18103

2024 arXiv
[9]

Y., & Wang, H

Li, L., Chang, T. Y., & Wang, H. (2023). Multimodal Gen-AI f or Fundamental Investment Research. arXiv preprint arXiv:2401.06164

2023 arXiv
[10]

Zhang, J., & Arawjo, I. (2024). ChainBuddy: An AI Agent Sy stem for Generating LLM Pipelines. arXiv preprint arXiv:2409.13588

2024 arXiv
[11]

Chu, Z., Guo, H., Zhou, X., Wang, Y., Yu, F., Chen, H., Xu, W ., Lu, X., Cui, Q., Li, L., Zhou, J., & Li, S. (2023). Data-Centric Financial Large Language Model s. arXiv preprint arXiv:2310.17784

2023 arXiv
[12]

C., & Hung, M

Chiu, I. C., & Hung, M. W. (2025). Finance-speciﬁc large l anguage models: Advancing sentiment analysis and return prediction with LLaMA 2. Paciﬁc-Basin Finance Journal , 90, 102632

2025
[13]

Yuan, H., Wang, S., & Guo, J. (2024). Alpha-GPT 2.0: Human -in-the-Loop AI for Quantitative In- vestment. arXiv preprint arXiv:2402.09746

2024 arXiv
[14]

Wang, Q., Gao, Y., Tang, Z., Luo, B., & He, B. (2024). Enhan cing LLM trading performance with fact-subjectivity aware reasoning. arXiv preprint arXiv:2410.12464

2024 arXiv
[15]

Deng, Y., He, X., Hu, J., & Yiu, S. M. (2024). Enhancing few -shot stock trend prediction with large language models. arXiv preprint arXiv:2407.09003

2024 arXiv
[16]

Bhat, R., & Jain, B. (2024). Stock price trend prediction using emotion analysis of ﬁnancial headlines with distilled LLM model. In Proceedings of the 17th International Conference on PErvas ive Technolo- gies Related to Assistive Environments (pp. 67-73)

2024
[17]

Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., & Gelly, S. (2019). Parameter-eﬃcient transfer learnin g for NLP. In International Conference on Machine Learning (pp. 2790-2799). PMLR

2019
[18]

L., & Liang, P

Li, X. L., & Liang, P. (2021). Preﬁx-tuning: Optimizing c ontinuous prompts for generation. arXiv preprint arXiv:2101.00190

2021 arXiv
[19]

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D. , Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., & Agarwal, S. (2020). Language mo dels are few-shot learners. Advances in Neural Information Processing Systems , 33, 1877-1901

2020
[20]

Kou, Z., Yu, H., Peng, J., & Chen, L. (2024). Automate stra tegy ﬁnding with LLM in quant investment. arXiv preprint arXiv:2409.06289

2024
[21]

C., Kumar, A., & Lu, C

Wang, S., Ji, T., Wang, L., Sun, Y., Liu, S. C., Kumar, A., & Lu, C. T. (2024). Stocktime: A time series specialized large language model architecture for stock price prediction. arXiv preprint arXiv:2409.08281

2024 arXiv
[22]

Zhou, L., Zhang, Y., Yu, J., Wang, G., Liu, Z., Yongchareo n, S., & Wang, N. (2025). LLM-Augmented Linear Transformer–CNN for Enhanced Stock Price Predictio n. Mathematics, 13(3), 487

2025
[24]

Zhang, C., Liu, X., Jin, M., Zhang, Z., Li, L., Wang, Z., Hu a, W., et al. (2024). When AI meets ﬁnance (StockAgent): Large language model-based stock trading in simulated real-world environments. arXiv preprint arXiv:2407.18957

2024 arXiv
[25]

Ni, H., Meng, S., Chen, X., Zhao, Z., Chen, A., Li, P., Zhan g, S., et al. (2024). Harnessing earnings reports for stock predictions: A QLoRA-enhanced LLM approa ch. In 2024 6th International Conference on Data-driven Optimization of Complex Systems (DOCS) (pp. 909-915). IEEE

2024
[26]

Li, Z., Tam, W. L. V., & Yeung, L. K. (2025). A Multimodal an d Sentiment-Based Trading System for Financial Portfolio Optimisation. In 43rd IEEE International Conference on Consumer Electronic s

2025
[28]

Koa, K. J. L., Ma, Y., Ng, R., & Chua, T. S. (2024). Learning to generate explainable stock predictions using self-reﬂective large language models. In Proceedings of the ACM Web Conference 2024 (pp. 4304- 4315)

2024
[29]

Yu, Y., Yao, Z., Li, H., Deng, Z., Jiang, Y., Cao, Y., Chen, Z., et al. (2025). FinCon: A synthesized LLM multi-agent system with conceptual verbal reinforceme nt for enhanced ﬁnancial decision making. Advances in Neural Information Processing Systems , 37, 137010-137045

2025
[30]

Xiao, Y., Sun, E., Luo, D., & Wang, W. (2024). TradingAgen ts: Multi-Agents LLM Financial Trading Framework. arXiv preprint arXiv:2412.20138

2024 arXiv
[32]

Joshi, S. (2025). Leveraging prompt engineering to enha nce ﬁnancial market integrity and risk man- agement

2025
[34]

Xu, C., Liu, Z., & Li, Z. (2025). FinArena: A Human-Agent C ollaboration Framework for Financial Market Analysis and Forecasting. arXiv preprint arXiv:2503.02692

2025 arXiv
[35]

Iacovides, G., Konstantinidis, T., Xu, M., & Mandic, D. ( 2024). FinLlama: LLM-Based Financial Sentiment Analysis for Algorithmic Trading. In Proceedings of the 5th ACM International Conference on AI in Finance (pp. 134-141)

2024
[36]

Elahi, A., & Taghvaei, F. (2024). Combining ﬁnancial dat a and news articles for stock price movement prediction using large language models. In 2024 IEEE International Conference on Big Data (BigData) (pp. 4875-4883). IEEE

2024
[37]

Kirtac, K., & Germano, G. (2025). Leveraging LLM-based s entiment analysis for portfolio allocation with proximal policy optimization. In ICLR 2025 Workshop on Machine Learning Multiscale Processe s

2025
[38]

Yang, Y., Zhang, Y., Wu, M., Zhang, K., Zhang, Y., Yu, H., H u, Y., & Wang, B. (2025). TwinMarket: A Scalable Behavioral and Social Simulation for Financial M arkets. arXiv preprint arXiv:2502.01506

2025
[42]

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin , V., Goyal, N., K¨ uttler, H., et al. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474

2020
[43]

V., & Zhou, D

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi , E., Le, Q. V., & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large lan guage models. Advances in Neural Information Processing Systems, 35, 24824-24837

2022
[44]

J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wan g, S., Wang, L., & Chen, W

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wan g, S., Wang, L., & Chen, W. (2022). LoRA: Low-rank adaptation of large language models. ICLR, 1(2), 3

2022
[45]

Papasotiriou, K., Sood, S., Reynolds, S., & Balch, T. (20 24). AI in Investment Analysis: LLMs for Equity Stock Ratings. In Proceedings of the 5th ACM International Conference on AI in Finance (pp. 419-427)
[46]

M., Spoon, R., Han, J., & Camerer, C

Henning, T., Ojha, S. M., Spoon, R., Han, J., & Camerer, C. F. (2025). LLM Trading: Analysis of LLM Agent Behavior in Experimental Asset Markets. arXiv preprint arXiv:2502.15800

2025
[47]

Ma, T., Du, J., Huang, W., Wang, W., Xie, L., Zhong, X., & Zh ou, J. T. (2025). LLM Knows Geom- etry Better than Algebra: Numerical Understanding of LLM-B ased Agents in A Trading Arena. arXiv preprint arXiv:2502.17967

2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [39]

W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., et al

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., et al. (2021). Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (pp. 8748-8763). PMLR

work page 2021

[2] [33]

A., Jordan, M

Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation , 3(1), 79-87

work page 1991

[3] [48]

figure1_llm_models.png

Lee, H., Choi, Y., & Kwon, Y. (2024). Quantifying Qualita tive Insights: Leveraging LLMs to Market Predict. arXiv preprint arXiv:2411.08404 . This figure "figure1_llm_models.png" is available in "png" format from: http://arxiv.org/ps/2507.01990v1

arXiv 2024

[4] [40]

F., Leike, J., Brown, T., Martic, M., Legg , S., & Amodei, D

Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg , S., & Amodei, D. (2017). Deep reinforcement learning from human preferences. Advances in Neural Information Processing Systems

work page 2017

[5] [41]

Liu, F. (2020). Learning to summarize from human feedbac k. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 583-592)

work page 2020

[6] [27]

Li, X., Li, Z., Shi, C., Xu, Y., Du, Q., Tan, M., Huang, J., & Lin, W. (2024). Alphaﬁn: Benchmarking ﬁnancial analysis with retrieval-augmented stock-chain f ramework. arXiv preprint arXiv:2403.12582

arXiv 2024

[7] [23]

Fatouros, G., Metaxas, K., Soldatos, J., & Karathanassi s, M. (2025). MarketSenseAI 2.0: Enhancing Stock Analysis through LLM Agents. arXiv preprint arXiv:2502.00415

arXiv 2025

[8] [4]

Tong, H., Li, J., Wu, N., Gong, M., Zhang, D., & Zhang, Q. (20 24). Ploutos: Towards interpretable stock movement prediction with ﬁnancial large language mod el. arXiv preprint arXiv:2403.00782

arXiv