Pith. sign in

REVIEW 6 major objections 4 minor 44 references

Assessing the Capabilities and Limitations of FinGPT Model in Financial NLP Applications

T0 review · 6 major / 4 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 FinGPT is strong at financial classification tasks but weak at reasoning and generation tasks, according to a six-task evaluation against GPT-4 and domain baselines.

desk verdict A genuinely new directional-bias analysis buried inside a benchmark whose comparative numbers don't hold up under inspection. read the letter →

arxiv 2507.08015 v1 pith:32DIRQGG submitted 2025-07-06 cs.CL cs.LG

Prudence Djagba , Chimezie A. Odinakachukwu This is my paper
classification cs.CLcs.LG
keywords FinGPTfinancialNLPsentimentanalysisquestionansweringtextsummarizationstockmovementpredictionnamedentityrecognitiondomain-specificLLMevaluation
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 paper tries to establish a precise capability profile for the financial language model FinGPT by running it on six finance-specific NLP tasks and comparing it with GPT-4, FinMA 7B, human, and baseline scores. The central claim is that FinGPT is competitive with or better than GPT-4 on structured classification tasks such as sentiment analysis and headline classification, but falls far behind on reasoning and generation tasks such as financial question answering and summarization. The authors argue that this split tracks the underlying decoder-only LLaMA architecture and the limits of its unidirectional attention, and they add a directional sensitivity analysis showing that FinGPT's stock predictions are biased toward bullish signals. A careful reader would care because the result maps which financial applications can safely rely on a low-cost open model and which still need a larger general-purpose model.

What carries the argument

The evaluation is carried through FinGPT's four-layer framework of data source, data engineering, LLM, and applications, with LLaMA2 as the base model and finance-specific LoRA adapters supplying domain adaptation. The central diagnostic contrast is between structured classification tasks and reasoning or generation tasks, and the paper identifies the decoder-only causal attention mechanism as the limiting object for summarization and complex QA because it can only attend to past tokens. For stock movement, the key machinery is a newly introduced directional sensitivity analysis that compares FinGPT's alignment with bullish versus bearish trends and simulates long-only versus short-only trading portfolios.

What would settle it

A replication that runs GPT-4, FinMA 7B, and FinGPT through identical prompts, decoding parameters, test samples, and scoring code on FLARE-FPB and ConvFinQA would settle whether FinGPT's classification parity and QA deficit are genuine: if the QA gap shrinks dramatically or the classification parity disappears, the paper's central comparative claim fails.

Watch

Extended reading notes

Core claim

FinGPT reaches F1 scores of 87.62% on FLARE-FPB sentiment analysis and 95.50% on financial headline classification, slightly above the reported GPT-4 scores of 86% on both tasks. On the financial question answering datasets ConvFinQA and FLARE-FinQA, however, its exact match scores are 28.4% and 3.8%, versus GPT-4's 76% and 69% and human scores of 89% and 91%. The paper asserts that this gap is architectural: FinGPT is a decoder-only LLaMA2 model fine-tuned with LoRA adapters, and its causal attention cannot build the full-context representation that abstractive summarization and multi-step numerical reasoning require. It also reports a consistent bullish bias in stock movement prediction and proposes a directional sensitivity analysis that measures how well the model tracks upward versus downward market phases.

Load-bearing premise

The comparison assumes that the imported GPT-4, FinMA 7B, and human scores were produced under conditions equivalent to FinGPT's runs; if prompts, samples, preprocessing, or scoring differed, the relative strengths and gaps could change or vanish.

Editorial extensions

If this is right

  • Financial sentiment analysis and headline classification pipelines can use FinGPT at accuracy levels comparable to GPT-4 while retaining the benefits of an open, lower-cost model.
  • Multi-step numerical financial question answering should not be delegated to FinGPT, since exact match scores of 3.8% to 28.4% are far below GPT-4 and human performance.
  • Text summarization of financial documents requires an encoder-decoder or retrieval-augmented architecture, because FinGPT's decoder-only attention does not capture the full document context needed for condensation.
  • Stock movement predictions from FinGPT carry a systematic bullish bias, so simulated long-only trading strategies outperform short-only strategies in the reported experiments.
  • The six-task score table provides a reference benchmark that future financial large language model work can compare against directly.

Reading between the lines

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

  • Because the GPT-4, FinMA 7B, and human numbers are imported from prior work rather than produced under the paper's exact prompts and decoding conditions, the comparative strengths and gaps should be treated as approximate until a same-condition rerun is done.
  • The reported reasoning gap suggests a testable extension the paper does not try: adding chain-of-thought prompting or an external calculator to FinGPT may recover part of the ConvFinQA deficit without changing the architecture.
  • The bullish bias in stock movement could come from the training distribution or label balance rather than from the model itself, so rebalancing fine-tuning data and checking calibration on bearish periods would separate those causes.
  • A practical consequence the paper leaves implicit is model routing: financial systems could use FinGPT for high-volume classification tasks and reserve a stronger general model for reasoning-heavy queries.
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.

Referee Report

6 major / 4 minor

Summary. This paper evaluates FinGPT across six financial NLP tasks (sentiment analysis, text classification, named entity recognition, financial question answering, stock movement prediction, and summarization) using public finance datasets. The authors report that FinGPT achieves strong classification performance, sometimes comparable to GPT-4, but lags on reasoning-heavy QA and summarization. The evaluation uses FinGPT LoRA adapters with LLaMA-2 base models, compares against numbers imported from prior work for GPT-4, FinMA 7B, and human performance, and draws conclusions about architectural limitations of decoder-only models.

Significance. If the reported findings were methodologically sound, the paper would provide a useful multi-task benchmark for a widely used open-source financial LLM. The authors make a reasonable attempt to cover diverse tasks and real-world datasets, and they identify a plausible directional limitation of FinGPT in numerical reasoning and abstractive summarization. However, the current evidence is compromised by evaluation-protocol problems, unverified baseline comparability, and possible data circularity, so the paper does not yet establish its central comparative claims. The effort to include a directional sensitivity analysis for stock prediction is a positive feature, though it is not enough to offset the methodological issues.

major comments (6)
  1. [Table 2 and Section 6] The same FinGPT score on FLARE-FPB appears as 87.62 (F1) in Table 2 and 87.62% (Acc) in Table 3, while the GPT-4 comparison number is an imported F1 from prior work. Accuracy and F1 are not interchangeable, especially on class-imbalanced sentiment data, and no same-condition, same-metric runs of GPT-4 or FinMA 7B are reported. The abstract's claim that FinGPT 'often achieves results comparable to GPT-4' is therefore not supported by the evidence as presented.
  2. [Section 4.4.4] The NER max_new_tokens parameter was reduced from 64 to 34 because this improved macro F1 from 38% to approximately 69%, but no validation split is described. This constitutes tuning on the test set and inflates the NER result reported in Tables 2 and 5; the authors must either use a validation split or report the test-score selection bias explicitly.
  3. [Sections 4.3.2, 4.3.5, and 4.5.2] The evaluation excludes ambiguous outputs ('maybe'), 'unknown' outputs, and any example where either the prediction or the ground truth is not a valid number. The number and proportion of excluded samples are not reported. These post-hoc exclusions inflate the sentiment and QA scores and prevent unbiased interpretation of the F1 and EM results.
  4. [Sections 4.3.1, 4.4.1, and 4.5.1] The FinGPT Headline, FinGPT-NER, and ConvFinQA evaluation datasets are hosted by the FinGPT project, and the adapters used are from the same FinGPT project. The paper does not disclose whether the fine-tuning data of these adapters includes the evaluation splits. Without a demonstration of non-overlap, the strong classification and NER scores could reflect memorization rather than generalization; this is a load-bearing concern for the paper's conclusions.
  5. [Section 4.7 and Table 2] The summarization evaluation is entirely qualitative: Table 2 lists a dash for FinGPT, and Section 4.7 reports only that FinGPT 'failed to generate coherent or informative summaries.' No quantitative metric (e.g., ROUGE or BERTScore) or error-analysis statistics are provided, so the paper's claim that FinGPT is significantly limited in summarization is not supported by the quantitative evidence.
  6. [Section 4.5.1 and Table 6] The QA evaluation uses only the first 200 ConvFinQA and first 50 FLARE-FinQA test examples, with no confidence intervals or statistical comparison. A 3.8% EM on 50 examples corresponds to roughly two correct answers, making the reported gap to GPT-4's 69% EM fragile and not robust to sampling choice.
minor comments (4)
  1. [Tables 2 and 7] The FinGPT scores for the same stock movement datasets differ between Table 2 (e.g., CIKM18 45.00 F1) and Table 7 (47.03 Acc) without explanation; the relationship between these metrics should be clarified.
  2. [References] Reference [7] is labeled 'Cikm18 stock dataset' but points to the flare-sm-acl URL, which is described elsewhere as StockNet; the dataset citations need to be reconciled.
  3. [Section 2.1] The statement that 'accuracy and F1 often fall short in assessing financial task performance' is presented without support or follow-up; either cite relevant work or remove the claim.
  4. [Throughout] Some table and figure captions are minimal (e.g., Figure 8 and Figure 9 lack descriptive detail), and the paper would benefit from reporting the exact number of excluded samples in each filtering step.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity: NER macro F1 and FLARE-FinQA EM are partly constructed by test-set hyperparameter tuning and ground-truth-based answer selection; the headline 'comparable to GPT-4' claim is otherwise a comparability/evidence issue, not a circular one.

  1. fitted input called prediction [Section 4.4.4, Inference Procedure (NER)]
    "Max New Tokens: Reduced from 64 to 34, which improved macro F1 from 38% to approximately 69%, minimizing hallucination."

    The reported NER result (Table 5: 69.76 Entity F1) is obtained after choosing max_new_tokens by observing its effect on macro F1 on the same test set (98 examples used as-is per Section 4.4.1). The generation parameter is fitted to the test metric, so the final score is the maximum of the tuned values rather than an independent evaluation. Because the same test examples are used both to select the hyperparameter and to report the score, the improvement from 38% to 69% is by construction the result of test-set optimization, not a prediction of generalization.

  2. fitted input called prediction [Section 4.5.4, Inference Strategy (Financial QA)]
    "For FLARE-FinQA, if multiple numbers were present, the one with the smallest absolute difference from the true value was used as the final prediction."

    The reported EM score for FLARE-FinQA (3.8%) is computed after choosing, among several candidate numbers in the model output, the candidate closest to the ground-truth value. This selection uses the true answer to define the prediction, so the resulting exact-match score is partially constructed from the target it is meant to predict. Combined with Section 4.5.2's retention of only examples where both prediction and ground truth are valid numbers, the QA evaluation reduces to a best-case selection rather than a fixed decoding rule.

full rationale

The reported NER score is not an independent prediction: max_new_tokens was selected by watching macro F1 on the 98 test examples, so the jump from 38% to 69% is a test-set optimization artifact. The FLARE-FinQA EM is likewise partially built from ground truth, since the model's multiple numeric candidates are resolved by choosing the value closest to the true answer and then only valid-number pairs are scored. These are concrete reductions of reported results to their own targets, meeting the bar for partial circularity. The classification-strength claim (sentiment, headline) is not itself circular, since those numbers come from the paper's own runs and are not derived from GPT-4, but the comparison with GPT-4 and FinMA is unsupported because Table 2 and Table 3 mix Accuracy with F1 and import baseline scores from unknown conditions; that is a correctness and comparability problem rather than a definitional circle. No self-citation chain or dataset-overlap evidence is established, so the score is 6 rather than higher.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central evaluation rests on the authenticity of the adapter checkpoints, the comparability of imported baselines, and the acceptability of post-hoc filtering. None of these are independently verified in the paper.

free parameters (4)
  • max_new_tokens (NER) = 34
    Reduced from 64 to 34 after observing test-set macro F1 improve from 38% to approximately 69%. This is hand-tuning on the evaluation data and directly affects the reported 69.76% entity F1.
  • ConvFinQA sample size = first 200 test examples
    The first 200 test examples were selected without justification, which can bias the QA score if the subset is not representative.
  • FLARE-FinQA sample size = first 50 test examples
    The first 50 test examples were selected, an ad hoc subsample that may not reflect the full benchmark difficulty.
  • QA numeric filtering rule = retain only numeric pairs
    Only examples where both prediction and ground truth are valid numbers were retained for EM scoring. This changes the test set composition and can inflate exact match scores.
assumptions (4)
  • domain assumption The numbers reported for GPT-4, FinMA 7B, and human performance are directly comparable to the FinGPT runs in this paper.
    The paper does not describe running these models; it appears to import scores from prior work, which may use different prompts, subsets, and preprocessing. Without same-condition runs, the comparison is unwarranted.
  • domain assumption The LoRA adapters used (e.g., fingpt-mt_llama2-7b_lora) were not fine-tuned on the test datasets used in this evaluation.
    No disclosure of adapter training data. If the adapters were trained on FinGPT-headline, fingpt-ner, or fingpt-convfinqa, the strong classification and NER scores are in-distribution measurements, not evidence of generalization.
  • domain assumption Excluding 'unknown' and 'maybe' outputs does not bias the classification results.
    The paper removes ambiguous outputs 'to maintain reliability', but this eliminates exactly the cases where the model is uncertain, inflating F1 scores.
  • domain assumption The selected datasets are representative of the six financial NLP tasks.
    Datasets are a convenience sample from HuggingFace, with no justification that they cover the breadth of financial NLP or that small subsamples (first 200, first 50) are representative.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Assessing the Capabilities and Limitations of FinGPT Model in Financial NLP Applications." pith.science (2026). https://pith.science/paper/32DIRQGG

@misc{pith2026250708015,
  author       = {Pith},
  title        = {Pith review of: Assessing the Capabilities and Limitations of FinGPT Model in Financial NLP Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/32DIRQGG}},
  note         = {Machine review of arXiv:2507.08015}
}
read the original abstract

This work evaluates FinGPT, a financial domain-specific language model, across six key natural language processing (NLP) tasks: Sentiment Analysis, Text Classification, Named Entity Recognition, Financial Question Answering, Text Summarization, and Stock Movement Prediction. The evaluation uses finance-specific datasets to assess FinGPT's capabilities and limitations in real-world financial applications. The results show that FinGPT performs strongly in classification tasks such as sentiment analysis and headline categorization, often achieving results comparable to GPT-4. However, its performance is significantly lower in tasks that involve reasoning and generation, such as financial question answering and summarization. Comparisons with GPT-4 and human benchmarks highlight notable performance gaps, particularly in numerical accuracy and complex reasoning. Overall, the findings indicate that while FinGPT is effective for certain structured financial tasks, it is not yet a comprehensive solution. This research provides a useful benchmark for future research and underscores the need for architectural improvements and domain-specific optimization in financial language models.

Figures

Figures reproduced from arXiv: 2507.08015 by the authors.

Figure 1
Figure 1. Timeline showing the evolution of selected PLM/LLM releases from the general domain to the financial [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Summary of FinPLMs and FinLLMs. Abbreviations include Paras. = Parameter Size, PT = Pretraining, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FinGPT system architecture highlighting the end-to-end data and modeling pipeline [41]. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (14 more)
Figure 6
Figure 6. Figure 6: Performance Comparison on Financial Sentiment Datasets [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Performance on Headline-Based Text Classification [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Comparison for Named Entity Recognition (NER) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Comparison of Model Performance on Financial QA Datasets [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Stock Movement Prediction Accuracy Across Datasets [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Rolling Mean of FinGPT vs Ground Truth on CIKM18 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Portfolio Value (Bearish Trades) — CIKM18 [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Portfolio Value (Bullish Trades) — CIKM18 [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: FinGPT vs Ground Truth (BigData22) [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Portfolio Value from Bearish Trades (BigData22) [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Portfolio Value from Bullish Trades (BigData22) [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: Smoothed Stock Movement on StockNet Dataset [PITH_FULL_IMAGE:figures/full_fig_p017_17.png]
Figure 18
Figure 18. Figure 18: Short-Only Strategy Portfolio — StockNet [PITH_FULL_IMAGE:figures/full_fig_p018_18.png]
Figure 19
Figure 19. Figure 19: Long-Only Strategy Portfolio — StockNet Insights: FinGPT’s predictive alignment with bullish markets leads to positive trading returns in upward trends, but its underperformance in bearish contexts highlights a key limitation. This directional imbalance is important f…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 26 canonical work pages

  1. [1]

    Sentiment classification: Review of text vectorization methods: Bag of words, tf-idf, word2vec and doc2vec

    Haisal Dauda Abubakar, Mahmood Umar, and Muhammad Abdullahi Bakale. Sentiment classification: Review of text vectorization methods: Bag of words, tf-idf, word2vec and doc2vec. SLU Journal of Science and Technology, 4(1):27–33, 2022

    work page 2022

  2. [2]

    Ai in financial services: Fraud detection, algorithmic trading, and risk assessment

    Moses Alabi. Ai in financial services: Fraud detection, algorithmic trading, and risk assessment. 2022

    work page 2022

  3. [3]

    Ai competencies for internal auditors in the public sector.Edpacs, 69(1):3–21, 2024

    Ceray Aldemİr and Tuğba Uçma Uysal. Ai competencies for internal auditors in the public sector.Edpacs, 69(1):3–21, 2024

    work page 2024

  4. [4]

    Finbert: Financial sentiment analysis with pre-trained language models

    Dogu Araci. Finbert: Financial sentiment analysis with pre-trained language models. arxiv 2019.arXiv preprint arXiv:1908.10063, 2019

    arXiv 2019

  5. [5]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  6. [6]

    Flare-fiqa-sa: Financial sentiment dataset.https://huggingface.co/datasets/ChanceFocus/ flare-fiqasa, 2023

    ChanceFocus. Flare-fiqa-sa: Financial sentiment dataset.https://huggingface.co/datasets/ChanceFocus/ flare-fiqasa, 2023. Accessed: 2025-06-10

    work page 2023

  7. [7]

    Cikm18 stock dataset

    ChanceFocus. Cikm18 stock dataset. https://huggingface.co/datasets/ChanceFocus/flare-sm-acl,

    work page

  8. [8]

    Flare-ectsum: Financial summarization dataset

    ChanceFocus. Flare-ectsum: Financial summarization dataset. https://huggingface.co/datasets/ ChanceFocus/flare-ectsum, 2024. Accessed: 2025-06-10

    work page 2024

Show all 44 references
  1. [9]

    Flare-finqa: Financial qa dataset

    ChanceFocus. Flare-finqa: Financial qa dataset. https://huggingface.co/datasets/ChanceFocus/ flare-finqa, 2024. Accessed: 2025-06-10

    2024

  2. [10]

    Stocknet dataset

    ChanceFocus. Stocknet dataset. https://huggingface.co/datasets/ChanceFocus/flare-sm-stocknet,

  3. [11]

    How artificial intelligence is reshaping the financial services industry, 2024

    Dr Kostis Chlouverakis. How artificial intelligence is reshaping the financial services industry, 2024

    2024

  4. [12]

    Accessed: 2025-06-10

    2025

  5. [13]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langua...

    2019

  6. [14]

    An investigation of artificial intelligence application in auditing

    Tamanna Abdul Rahman Dalwai, Araby Madbouly, and Syeeda Shafiya Mohammadi. An investigation of artificial intelligence application in auditing. InArtificial intelligence and COVID effect on accounting, pages 101–114. Springer, 2022

    2022

  7. [15]

    Convfinqa dataset

    FinGPT. Convfinqa dataset. https://huggingface.co/datasets/FinGPT/fingpt-convfinqa, 2024. Ac- cessed: 2025-06-10

    2024

  8. [16]

    Enhancing financial risk management with federated ai

    Vineet Dhanawat, Varun Shinde, Vishal Karande, and Kartik Singhal. Enhancing financial risk management with federated ai. In2024 8th SLAAI International Conference on Artificial Intelligence (SLAAI-ICAI), pages 1–6. IEEE, 2024

    2024

  9. [17]

    Fingpt-ner dataset

    FinGPT. Fingpt-ner dataset. https://huggingface.co/datasets/FinGPT/fingpt-ner, 2024. Accessed: 2025-06-10

    2024

  10. [18]

    Fingpt-headline dataset

    FinGPT. Fingpt-headline dataset. https://huggingface.co/datasets/FinGPT/fingpt-headline, 2024. Accessed: 2025-06-10

    2024

  11. [19]

    Is chatgpt a financial expert? evaluating language models on financial natural language processing

    Yue Guo, Zian Xu, and Yi Yang. Is chatgpt a financial expert? evaluating language models on financial natural language processing. arXiv preprint arXiv:2310.12664, 2023

    2023 arXiv

  12. [20]

    MIT press Cambridge, 2016

    Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio.Deep learning, volume 1. MIT press Cambridge, 2016

    2016

  13. [21]

    Analyzing the role of artificial intelligence in it audit: current practices and future prospects

    U Iwuanyanwu, AJ Apeh, OR Adaramodu, EC Okeleke, and OG Fakeyede. Analyzing the role of artificial intelligence in it audit: current practices and future prospects. Computer Science & IT Research Journal, 4(2):54–68, 2023

    2023

  14. [22]

    Finbert: A large language model for extracting information from financial text

    Allen H Huang, Hui Wang, and Yi Yang. Finbert: A large language model for extracting information from financial text. Contemporary Accounting Research, 40(2):806–841, 2023. 20

    2023

  15. [23]

    A survey of large language models in finance (finllms)

    Jean Lee, Nicholas Stevens, Soyeon Caren Han, and Minseok Song. A survey of large language models in finance (finllms). arXiv preprint arXiv:2402.02315, 2024

    2024 arXiv

  16. [24]

    A comprehensive review of generative ai in finance

    David Kuo Chuen Lee, Chong Guan, Yinghui Yu, and Qinxu Ding. A comprehensive review of generative ai in finance. FinTech, 3(3):460–478, 2024

    2024

  17. [25]

    Are chatgpt and gpt-4 general-purpose solvers for financial text analytics? a study on several typical tasks.arXiv preprint arXiv:2305.05862, 2023

    Xianzhi Li, Samuel Chan, Xiaodan Zhu, Yulong Pei, Zhiqiang Ma, Xiaomo Liu, and Sameena Shah. Are chatgpt and gpt-4 general-purpose solvers for financial text analytics? a study on several typical tasks.arXiv preprint arXiv:2305.05862, 2023

    2023 arXiv

  18. [26]

    Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

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

    2020

  19. [27]

    A proposal for the dartmouth summer research project on artificial intelligence.WIRED, 2012

    John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. A proposal for the dartmouth summer research project on artificial intelligence.WIRED, 2012. Accessed: 2025-04-16

    2012

  20. [28]

    Fingpt: Democratizing internet-scale data for financial large language models.arXiv preprint arXiv:2307.10485, 2023

    Xiao-Yang Liu, Guoxuan Wang, Hongyang Yang, and Daochen Zha. Fingpt: Democratizing internet-scale data for financial large language models.arXiv preprint arXiv:2307.10485, 2023

    2023 arXiv

  21. [29]

    Fino1: On the transferability of reasoning enhanced llms to finance.arXiv preprint arXiv:2502.08127, 2025

    Lingfei Qian, Weipeng Zhou, Yan Wang, Xueqing Peng, Han Yi, Jimin Huang, Qianqian Xie, and Jianyun Nie. Fino1: On the transferability of reasoning enhanced llms to finance.arXiv preprint arXiv:2502.08127, 2025

    2025 arXiv

  22. [30]

    Deep learning for financial applica- tions: A survey.Applied soft computing, 93:106384, 2020

    Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, and Omer Berat Sezer. Deep learning for financial applica- tions: A survey.Applied soft computing, 93:106384, 2020

    2020

  23. [31]

    Generative adversarial networks applied to synthetic financial scenarios generation.Physica A: Statistical Mechanics and its Applications, 623:128899, 2023

    Matteo Rizzato, Julien Wallart, Christophe Geissler, Nicolas Morizet, and Noureddine Boumlaik. Generative adversarial networks applied to synthetic financial scenarios generation.Physica A: Statistical Mechanics and its Applications, 623:128899, 2023

    2023

  24. [32]

    Artificial intelligence and the evolution of finance: opportunities, challenges and ethical considerations.EdTech Journal, 3(1):20–23, 2023

    Marko Ranković, Elena Gurgu, Oliva Martins, and Milan Vukasović. Artificial intelligence and the evolution of finance: opportunities, challenges and ethical considerations.EdTech Journal, 3(1):20–23, 2023

    2023

  25. [33]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  26. [34]

    International Monetary Fund, 2023

    Mr Ghiath Shabsigh and El Bachir Boukherouaa.Generative artificial intelligence in finance: Risk considera- tions. International Monetary Fund, 2023

    2023

  27. [35]

    Flare-fpb: Financial phrase bank dataset

    TheFinAI. Flare-fpb: Financial phrase bank dataset. https://huggingface.co/datasets/TheFinAI/ flare-fpb, 2023. Accessed: 2025-06-10

    2023

  28. [36]

    Bigdata22 dataset.https://huggingface.co/datasets/TheFinAI/flare-sm-bigdata, 2023

    TheFinAI. Bigdata22 dataset.https://huggingface.co/datasets/TheFinAI/flare-sm-bigdata, 2023. Ac- cessed: 2025-06-10

    2023

  29. [37]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  30. [38]

    Flare-fpb: Financial language analysis for real-world events - financial phrasebank

    TheFinAI. Flare-fpb: Financial language analysis for real-world events - financial phrasebank. https:// huggingface.co/datasets/TheFinAI/flare-fpb, 2024. Accessed: 2025-05-18

    2024

  31. [39]

    Bloomberggpt: A large language model for finance.arXiv preprint arXiv:2303.17564, 2023

    Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kam- badur, David Rosenberg, and Gideon Mann. Bloomberggpt: A large language model for finance.arXiv preprint arXiv:2303.17564, 2023

    2023 arXiv

  32. [40]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  33. [41]

    Fingpt: Open-source financial large language models

    Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. Fingpt: Open-source financial large language models. arXiv preprint arXiv:2306.06031, 2023

    2023

  34. [42]

    Pixiu: A large language model, instruction data and evaluation benchmark for finance

    Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. Pixiu: A large language model, instruction data and evaluation benchmark for finance. arXiv preprint arXiv:2306.05443, 2023. 21

    2023 arXiv

  35. [43]

    Gpt (generative pre-trained transformer)–a comprehensive review on enabling technologies, potential applications, emerging challenges, and future directions.IEEE Access, 2024

    Gokul Yenduri, M Ramalingam, G Chemmalar Selvi, Y Supriya, Gautam Srivastava, Praveen Kumar Reddy Maddikunta, G Deepti Raj, Rutvij H Jhaveri, B Prabadevi, Weizheng Wang, et al. Gpt (generative pre-trained transformer)–a comprehensive review on enabling technologies, potential ...

    2024

  36. [44]

    Investlm: A large language model for investment using financial domain instruction tuning.arXiv preprint arXiv:2309.13064, 2023

    Yi Yang, Yixuan Tang, and Kar Yan Tam. Investlm: A large language model for investment using financial domain instruction tuning.arXiv preprint arXiv:2309.13064, 2023

    2023 arXiv

Pith tools

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