On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News Headlines
Open accessThe DOI matches its Crossref record; 89 of the 89 references it cites were found in a registry or an index; none of the 8 sources its claims cite is checked for support yet.
Abstract
The importance of sentiment analysis in the rapidly evolving financial markets is widely recognized for its ability to interpret market trends and inform investment decisions. This study delves into the target-level financial sentiment analysis (TLFSA) of news headlines related to stock. The study compares the performance in the TLFSA task of various sentiment analysis techniques, including rule-based models (VADER), fine-tuned transformer-based models (DistilFinRoBERTa and Deberta-v3-base-absa-v1.1) as well as zero-shot large language models (ChatGPT and Gemini). The dataset utilized for this analysis, a novel contribution of this research, comprises 1476 manually annotated Bloomberg headlines and is made publicly available (due to copyright restrictions, only the URLs of Bloomberg headlines with the manual annotations are provided; however, these URLs can be used with a Bloomberg terminal to reconstruct the complete dataset) to encourage future research on this subject. The results indicate that the fine-tuned Deberta-v3-base-absa-v1.1 model performs better across all evaluation metrics than other evaluated models in TLFSA. However, LLMs such as ChatGPT-4, ChatGPT-4o, and Gemini 1.5 Pro provide similar performance levels without the need for task-specific fine-tuning or additional training. The study contributes to assessing the performance of LLMs for financial sentiment analysis, providing useful insights into their possible application in the financial domain.
Overview
CoChat Summary
TL;DR
This paper evaluates target-level financial sentiment analysis (TLFSA) on 1,476 manually annotated Bloomberg stock-news headlines, comparing traditional models with zero-shot LLMs. It finds that a fine-tuned DeBERTa-v3-base-ABSA-v1.1 model performs best overall, while ChatGPT-4, ChatGPT-4o, and Gemini 1.5 Pro approach its performance without task-specific fine-tuning.
Key findings
- Fine-tuned DeBERTa-v3-base-ABSA-v1.1 achieved the best scores: accuracy 0.8679, macro precision 0.8499, macro recall 0.8414, and macro F1-score 0.8448.
- ChatGPT-4 ranked second with accuracy 0.8591 and macro F1-score 0.8337; ChatGPT-4o reached accuracy 0.8354 and macro F1 0.8187, while Gemini 1.5 Pro reached accuracy 0.8266 and macro F1 0.8123.
- Traditional document/sentence-level models performed poorly: VADER had accuracy 0.4939 and macro F1 0.4753; DistilFinRoBERTa had accuracy 0.5129 and macro F1 0.5010; pre-trained DeBERTa-v3-base-ABSA-v1.1 had accuracy 0.5169 and macro F1 0.5367.
- The dataset contains 1,476 headlines, with roughly 49% positive, 28% negative, and 23% neutral; intercoder agreement on a 150-headline subset was Krippendorff’s Alpha 0.8188.
Methods
- The authors collected Bloomberg Terminal headlines for Amazon, Netflix, Nvidia, and Alphabet from 1 September 2023 to 31 January 2024, producing 1,476 headlines annotated by three finance/economics experts using labels -1, 0, and +1.
- They compared VADER, DistilFinRoBERTa, DeBERTa-v3-base-ABSA-v1.1 (pre-trained and fine-tuned via 5-fold cross-validation), and zero-shot ChatGPT-3.5, ChatGPT-4, ChatGPT-4o, Gemini 1 Pro, Gemini 1.5 Flash, and Gemini 1.5 Pro.
- Performance was evaluated with precision, recall, F1-measure, and accuracy, reported using both macro- and weighted-averaging across the three sentiment classes.
Limitations
- The paper states that the limited range of data may affect the applicability of findings to a broader context, and recommends expanding the dataset, extending the timeframe, incorporating additional datasets, and evaluating other LLMs and few-shot performance.
- The paper states that copyright restrictions mean only headline URLs and manual annotations are released, not the actual headline text, so users need Bloomberg terminal access to reconstruct the full dataset.
- Evident from the text: the study covers only four stock companies and Bloomberg headlines over a five-month period, so generalizability to other firms, sources, languages, or time periods is not established; the paper also notes LLM black-box interpretability and external model updates as practical challenges.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 1 Not valid · 5 Not checked · 39 Own results and interpretations
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, LLMs such as ChatGPT-4, ChatGPT-4o, and Gemini 1.5 Pro provide similar performance levels without the need for task-specific fine-tuning or additional training.”
§ On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News Headlines > Abstract
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The dataset consists of 1476 distinct news headlines, which are each linked to a particular stock company (referred to as the”
§ On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News Headlines > 3.1. Dataset Creation and Annotation
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Indeed, the fine-tuned version of the DeBERTa-v3-base-ABSA-v1.1 model exhibits exceptional performance, attaining the top scores in all evaluation metrics.”
§ On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News Headlines > 5. Conclusions
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The fine-tuned DeBERTa-v3-base-ABSA-v1.1 model stands out by achieving the highest recorded accuracy of 0.8679 in this investigation.”
§ On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News Headlines > 4. Results and Discussion
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Pre-trained models such as DistilFinRoBERTa and DeBERTa-v3-base-ABSA-v1.1 exhibit some improvement compared to VADER.”
§ On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News Headlines > 4. Results and Discussion
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
12 more claims
Create a free account to see all 45 claims and every source verdict.
- It may state two separate claims as one. 22
- Probably not a substantial claim. 13
- The quote only partly states this claim. 1
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Content Analysis: An Introduction to Its Methodology
Krippendorff, K. (2019) · doi:10.4135/9781071878781 · Cited by 15,448
Opinion Mining and Sentiment Analysis
Pang, B., & Lee, L. (2008) · now publishers, Inc. eBooks · doi:10.1561/9781601981516 · Cited by 6,720
Opinion mining and sentiment analysis
Pang, B., & Lee, L. (2008) · Foundations and Trends® in Information Retrieval · doi:10.1561/1500000011 · Cited by 6,372
Chicco, D., & Jurman, G. (2020) · BMC Genomics · doi:10.1186/s12864-019-6413-7 · Cited by 6,143
VADER: a parsimonious rule-based model for sentiment analysis of social media text
Hutto, C., & Gilbert, E. (2014) · Proceedings of the International AAAI Conference on Web and Social Media · doi:10.1609/icwsm.v8i1.14550 · Cited by 6,087
When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10‐Ks
LOUGHRAN, T., & MCDONALD, B. (2011) · The Journal of Finance · doi:10.1111/j.1540-6261.2010.01625.x · Cited by 6,015
The general theory of employment
Keynes, J. M. (1937) · The Quarterly Journal of Economics · doi:10.2307/1882087 · Cited by 5,173
Giving content to investor sentiment: The role of media in the stock market
TETLOCK, P. C. (2007) · The Journal of Finance · doi:10.1111/j.1540-6261.2007.01232.x · Cited by 4,827
Research works printed but not linked: 2
- Rizinski, M., Mishev, K., Chitkushev, L.T., Vodenska, I., and Trajanov, D. (2023, January 12–15). Using NLP transformer models to evaluate the relationship between ethical principles in finance and machine learning. Proceedings of the 13th International Conference on Information Society and Technology (ICIST) Conference, Kopaonik, Serbia.
- Yang, H., and Li, K. (2021). Improving implicit sentiment learning via local sentiment aggregation. arXiv.
Cited by
Sentiment Analysis Algorithm Based on Deep Transfer Learning for Multi-Source Data Fusion
Qin, L. (2025) · IEEE Access · doi:10.1109/access.2025.3593022 · Cited by 4
“The experimental results showed that Deberta-v3 performed the best after fine-tuning, while large language models such as ChatGPT performed well without fine-tuning, providing new insights for financial SA [11].”
Malaquias, R. F., & Hwang, Y. (2025) · Journal of Management History · doi:10.1108/jmh-01-2025-0008 · Cited by 3
García‐Magariño, I., Bravo, J., & Lacuesta, R. (2025) · AI · doi:10.3390/ai6050095 · Cited by 3
“Muhammad et al. [13] evaluated the effectiveness of traditional sentiment analysis techniques compared to ChatGPT and Gemini in the financial domain.”
Theodorakopoulos, L., & Theodoropoulou, A. (2026) · Electronics · doi:10.3390/electronics15071405 · Cited by 1
Bittla, S. R., Mhsnhasan, M., Ayshwarya, B., et al. (2025) · doi:10.1109/icdsis65355.2025.11070322
“One of the primary challenges is the complexity of financial text data which often includes domain-specific terminology, jargon, and nuanced language [4].”
Positive sentiments in early academic literature on DeepSeek: a cross-disciplinary mini review
He, Y., Giangan, A., Vu, N., & Watters, C. (2026) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2025.1725853 · Cited by 2
Gupta, S. (2026) · International Journal of Advanced Research in Science Communication and Technology · doi:10.48175/ijarsct-31649
Patil, R., & Sekar, A. M. (2026) · Biomedical Signal Processing and Control · doi:10.1016/j.bspc.2026.110126
FinLumen: A Game-Theoretic Multi-Agent Framework for Rational Portfolio Management
Zhang, Z., Sen, P., Gu, F., Wang, H., Jiang, Z., & Su, J. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11462217
Suwanwijit, N., Viriyapant, K., & Bhumpenpein, N. (2025) · Artificial Intelligence and Cloud Computing Conference (AICCC) · doi:10.1145/3789982.3790000
Sriveni, B., Pramod Kumar, P., & Muppidi, S. (2026) · IEEE International Conference on Emerging Computing and Intelligent Technologies (ICoECIT) · doi:10.1109/icoecit68303.2026.11497436
Bonifacio, K. L. A., & Regla, A. I. (2026) · International Conference on Information and Education Technology (ICIET) · doi:10.1109/iciet69664.2026.11561568
“Comments that use other Philippine dialects were also removed. mBERT and DistilBERT utilized WordPiece, while DeBERTa and RoBERTa employed Byte-Pair Encoding (BPE) to manage regional sub-words [9], [12].”
InfluentialBenchmarking large language models for target-based financial sentiment and stock return
Muhammad, I., Rospocher, M., Knez, T., & Žitnik, S. (2026) · International Journal of Data Science and Analytics · doi:10.1007/s41060-026-01201-x · Cited by 1
“VADER, 6 developed by Hutto and Gilbert [15], a rule-based framework, incorporates lexical, grammatical, and syntactic heuristics—validated against LIWC and ANEW—and has also been extensively employed in financial contexts [21, 43–46].”
Self-citationSetiawan, A. M., & Wibowo, A. (2026) · International Conference on Intelligent Engineering and Next-Gen Healthcare Systems (IEHNS) · doi:10.1109/iehns68708.2026.11606723
“This comes through the advantage from large scale pretraining and the capacity to change through finetuning for targeted applications [13].”
Metadata
- Authors
- Muhammad, I., & Rospocher, M.
- Affiliations
- University of Verona (IT)
- Corresponding author
- Rospocher, Marco
- Venue
- Algorithms
- Publisher
- MDPI AG
- Published
- 2025-01-13
- Volume
- 18
- Issue
- 1
- Pages
- 46
- ISSN
- 1999-4893 (ISSN-L)
- OpenAlex
- W4406308807
- Cited by
- 14
- References
- 70
- Citations by year
- 2026: 9 · 2025: 5
- FWCI
- 5.66
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Algorithms (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
- Social Sciences › Decision Sciences › Management Science and Operations Research › Stock Market Forecasting Methods
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- financial sentiment analysis, financial news headlines, stock market sentiment, large language models, DeBERTa, VADER, ChatGPT, Gemini, annotated dataset
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.