Draft:Financial sentiment analysis

Financial sentiment analysis combines natural language processing with concepts from financial economics to measure attitudes expressed in market-relevant documents.{{cite conference |last1=Xing |first1=Frank Z. |last2=Malandri |first2=Lorenzo |last3=Zhang |first3=Yue |last4=Cambria |first4=Erik |title=Financial Sentiment Analysis: An Investigation into Common Mistakes and Silver Bullets |conference=Proceedings of the 28th International Conference on Computational Linguistics |date=December 2020 |location=Barcelona |pages=978–987}} Typical sources include annual reports, earnings-call transcripts, newswire stories, and social-media posts. Because financial language differs from everyday usage, for example, liability is neutral in accounting contexts, FSA systems use domain-specific lexicons and models.{{cite journal |last1=Loughran |first1=Tim |last2=McDonald |first2=Bill |title=When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks |journal=The Journal of Finance |year=2011 |volume=66 |issue=1 |pages=35–65 |doi=10.1111/j.1540-6261.2010.01625.x}}

Early 2000s studies mined internet message boards for investor mood and documented links with equity returns.{{cite journal |last1=Antweiler |first1=Wolfgang |last2=Frank |first2=Murray Z. |title=Is All That Talk Just Noise? The Information Content of Internet Stock Message Boards |journal=The Journal of Finance |year=2004 |volume=59 |issue=3 |pages=1259–1294 |doi=10.1111/j.1540-6261.2004.00662.x}} The first financial lexicons—Henry’s Financial Dictionary followed by the Loughran–McDonald list—were created to capture finance-specific language. The availability of labeled datasets in the 2010s enabled supervised classifiers such as support-vector machines and random forests.{{cite journal |last1=Sohangir |first1=Sahar |last2=Wang |first2=Dingding |last3=Pomeranets |first3=Anna |last4=Khoshgoftaar |first4=Taghi M. |title=Big Data: Deep Learning for Financial Sentiment Analysis |journal=Journal of Big Data |year=2018 |volume=5 |issue=3 |pages=1–25 |doi=10.1186/s40537-017-0111-6|doi-access=free }} Advances in deep learning and transformer models, including BERT and its domain-adapted variant FinBERT, now set state-of-the-art benchmarks.{{cite arXiv | eprint=1908.10063 | last1=Araci | first1=Dogu | title=FinBERT: Financial Sentiment Analysis with Pre-trained Language Models | date=2019 | class=cs.CL }}

• Lexicon-based methods  – Score text using finance-specific word lists such as Loughran–McDonald.
• Traditional supervised learning  – Employ algorithms including naïve Bayes, random forests, and support-vector machines trained on labeled financial corpora.
• Deep learning  – Use convolutional neural networks, recurrent neural networks, or transformers to model contextual cues.
• Hybrid and ensemble models  – Combine lexicon features with machine-learning or neural outputs for improved robustness.

• Market prediction – Sentiment metrics derived from news and social media help forecast short-term price movements and volatility.{{cite journal |last=Chen |first=Tianyu |title=EFSA: Towards Event-Level Financial Sentiment Analysis |journal=Proceedings of ACL 2024 |year=2024 |pages=7455–7467}}
• Algorithmic trading – Funds incorporate real-time sentiment signals into automated trading strategies.{{cite conference |last1=Iacovides |first1=Giorgos |last2=Konstantinidis |first2=Thanos |last3=Xu |first3=Mingxue |last4=Mandic |first4=Danilo |title=FinLlama: LLM-Based Financial Sentiment Analysis for Algorithmic Trading |conference=5th ACM International Conference on AI in Finance |year=2024 |pages=134–142 |doi=10.1145/3677052.3698696}}
• Risk assessment and portfolio management – Analysts gauge managerial tone in filings and calls to inform credit and valuation models.

Large language models—including FinGPT{{cite arXiv | eprint=2306.05429 | last1=Li | first1=Yan | last2=Sun | first2=Jingbo | last3=Wen | first3=Yongzheng | last4=Xiong | first4=Xiaoyu | last5=Zhou | first5=Ji | title=Spin photonics on chip based on a twinning crystal metamaterial | date=2023 | class=physics.optics }}, BloombergGPT{{cite arXiv | eprint=2303.17564 | last1=Wu | first1=Shijie | last2=Irsoy | first2=Ozan | last3=Lu | first3=Steven | last4=Dabravolski | first4=Vadim | last5=Dredze | first5=Mark | last6=Gehrmann | first6=Sebastian | last7=Kambadur | first7=Prabhanjan | last8=Rosenberg | first8=David | last9=Mann | first9=Gideon | title=BloombergGPT: A Large Language Model for Finance | date=2023 | class=cs.LG }}, and FinLlama{{cite conference |last1=Iacovides |first1=Giorgos |last2=Konstantinidis |first2=Thanos |last3=Xu |first3=Mingxue |last4=Mandic |first4=Danilo |title=FinLlama: LLM-Based Financial Sentiment Analysis for Algorithmic Trading |conference=5th ACM International Conference on AI in Finance |year=2024 |pages=134–142 |doi=10.1145/3677052.3698696}}—are being explored for few-shot or zero-shot financial sentiment tasks. Multimodal approaches combine textual sentiment with market microstructure data, and explainable-AI techniques aim to increase interpretability in regulated environments.

{{Reflist}}

• [https://www.aaii.com/sentimentsurvey AAII Investor Sentiment Survey]
• [https://www.sentix.de Sentix Market Sentiment]
• [https://huggingface.co/bloomberg/bloomberggpt BloombergGPT on Hugging Face]
• [https://www.nasdaq.com/market-activity Nasdaq Market Sentiment Data]