Large Language Models for Smarter Market Sentiment and Trend Prediction

Abstract

The fast development of financial sectors has provided an enormous amount of unstructured data which ranges from the organization files, reports collected by the analyst for the social media discussion and the real-time news. Thus, collecting the useful information is very complex task for traders, financial organizations and the investors. The current advancements in the Large Language Models (LLMs) provides a promising way forward. LLM understands the tone, context and other hidden signals inside the text. It can also identify the market sentiment and enhances the trend prediction accuracy. In the traditional models, the dictionaries which are predefined will be considered or it uses the shallow statistical methods, but in LLM it offers a richer insight by identifying the subtle changes in the investor sentiment and uncovers the potential risks or opportunities. This research examines how the LLMs will be used for the financial sentiment analysis and the trend forecasting by reviewing the core architectures, benchmark datasets and various evaluation strategies. It also points out the specific challenges such as biased data, privacy concerns and regulatory compliance. once these challenges are addressed, LLM-powered systems have the potential to provide smarter, more adaptive, and human-like financial insights, enabling faster and more confident decision-making in dynamic market environments. To demonstrate the proposed approach, the Financial Phrase Bank dataset will be considered for implementation, with the aim of evaluating LLM-based models against traditional approaches and showcasing measurable improvements in sentiment detection and trend prediction accuracy.

Introduction

• Financial markets generate massive volumes of structured (e.g., stock prices) and unstructured (e.g., news, reports, social media) data daily.

• Traditional sentiment analysis methods (lexicon-based, shallow models) struggle to capture context and subtle financial language.

• There’s growing demand for context-aware models that support reliable decision-making in fast-moving markets.

II. Rise of Large Language Models (LLMs)

• Advanced models like GPT, FinBERT, and BloombergGPT excel at:

  • Understanding financial context

  • Recognizing sentiment (optimism, fear, caution)

  • Generating insights from unstructured financial text

• Unlike fixed-dictionary models, LLMs dynamically interpret nuanced language, enabling real-time market sentiment detection.

III. Application: Sentiment & Trend Prediction

• Investor sentiment, influenced by news and social media, significantly impacts market trends.

• LLMs can:

  • Process real-time textual data

  • Assess sentiment in news, reports, or earnings announcements

  • Predict short- and long-term market reactions

IV. Dataset: Financial PhraseBank

• Contains 4,840 sentences from financial news, each labeled as positive, neutral, or negative.

• Used to fine-tune and test LLMs (like FinBERT).

• Offers a benchmark for comparing LLMs against traditional models (SVM, Logistic Regression).

V. Implementation

A. Data Preparation

• Cleaned and standardized text: lowercasing, punctuation removal, domain-specific tokenization.

B. Model Selection

• FinBERT chosen for its financial-specific training.

• Baselines: SVM and Logistic Regression with TF-IDF features.

C. Fine-Tuning & Training

• FinBERT fine-tuned using cross-entropy loss and AdamW optimizer.

• Hyperparameters: learning rate = 2e-5, batch size = 16, epochs = 3.

D. Evaluation

• Metrics: Accuracy, Precision, Recall, F1-score.

• FinBERT significantly outperformed traditional models.

E. Sentiment to Trend Forecasting

• Aggregated sentiment scores can feed into ARIMA or LSTM models for market trend prediction.

VI. Results Summary

• F1-Score:

  • FinBERT: 0.915

  • SVM: ~0.77

  • Logistic Regression: ~0.80

VII. Discussion

• FinBERT provides:

  • High accuracy and stability

  • Improved sentiment detection in complex financial texts

  • Greater consistency and lower error margins

• Traditional models fail to capture nuanced expressions in financial documents.

• Visual analysis confirms FinBERT's superior performance across all metrics.

Conclusion

This study shows that Large Language Models, such as FinBERT, can effectively capture complex financial sentiment from the Financial PhraseBank dataset. In comparison to conventional machine learning models like Support Vector Machines and Logistic Regression, FinBERT exhibits substantially higher accuracy, precision, recall, and F1-score, highlighting the benefits of domain-adapted transformer models in understanding intricate financial terminology. Deploying Large Language Models in actual financial settings poses challenges that necessitate careful consideration, comprising guaranteeing model robustness, managing resource demands, and preserving reliability across a wide range of data scenarios. This research offers fundamental knowledge on integrating sophisticated natural language processing techniques into financial analysis systems and highlights the promising capabilities of large language models to facilitate more informed investment choices and accurate risk evaluation.Future research can focus on expanding Large Language Model applications to more diverse and multilingual financial datasets to improve generalizability. Enhancing model interpretability and efficiency for real-time deployment will be vital for practical adoption. Additionally, integrating multimodal data and addressing ethical concerns like bias and privacy will advance responsible and accurate financial forecasting.

References

[1] D. Araci, “FinBERT: Financial Sentiment Analysis with Pre-trained Language Models,” arXiv preprint arXiv:1908.10063, 2019. [2] B. Wu et al., “BloombergGPT: A Large Language Model for Finance,” arXiv preprint arXiv:2303.17564, 2023. [3] Y. Ma and Y. Zhang, “Applications of Large Language Models in Financial Decision-Making,” Journal of Financial Data Science, vol. 4, no. 2, pp. 45– 59, 2022. [4] FiQA Challenge, “Financial Opinion Mining and Question Answering,” 2018. [5] P. Malo, A. Sinha, and P. M. G. Apweiler, \"Good debt or bad debt: Detecting semantic orientations in economic texts,\" Journal of the Association for Information Science and Technology, vol. 65, no. 4, pp. 782–796, 2014. [Online]. Available: https://doi.org/10.1002/asi.23063 [6] Hugging Face, \"Financial PhraseBank,\" 2025. [Online]. Available: https://huggingface.co/datasets/takala/financial_phrasebank [7] Sun, Y., Yuan, H., & Xu, F. (2025). Financial sentiment analysis for pre-trained language models incorporating dictionary knowledge and neutral features. Natural Language Processing Journal, 11, 100148. https://doi.org/10.1016/j.nlp.2025.100148 [8] Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv preprint arXiv:1908.10063. [9] Research on financial sentiment analysis using FinBERT and transformer models: ProsusAI/finBERT GitHub repository. Large language models in finance: what is financial sentiment? (2025). SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5166656 [10] Survey on natural language processing in finance:Natural Language Processing in Finance: A Survey, ScienceDirect, 2024. https://www.sciencedirect.com/science/article/abs/pii/S1566253524005335 [11] Financial sentiment analysis for pre-trained language models incorporating dictionary knowledge and neutral features, Natural Language Processing Journal, Volume 11, June 2025, Elsevier.https://doi.org/10.1016/j.nlp.2025.100148 [12] Kirtac, K., & Germano, G. (2025). Large language models in finance: what is financial sentiment? arXiv preprint. DOI: 10.48550/arXiv.2503.03612 [13] Shen, Y., & Zhang, P. K. (2024). Financial Sentiment Analysis on News and Reports Using Large Language Models and FinBERT. arXiv preprint. DOI: 10.48550/arXiv.2410.01987 [14] Kirtac, K., & Germano, G. (2024). Enhanced Financial Sentiment Analysis and Trading Strategy Development Using Large Language Models. Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis.DOI: 10.18653/v1/2024.wassa-1.1 [15] Zafar, M. B. (2025). FinAI-BERT: A Transformer-Based Model for Sentence-Level Detection of AI Disclosures in Financial Reports. arXiv preprint.DOI: 10.48550/arXiv.2507.01991

Copyright

Copyright © 2025 Dr. C. Bhuvaneshwari, Dr. R. Gayathri. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.