Elamur Research⌕
☰
home / papers / 2026 / Enhancing Brazilian Inflation Forecasts through Sentiment Analysis Using Large Language Models

Enhancing Brazilian Inflation Forecasts through Sentiment Analysis Using Large Language Models

Paper · 2026 · Lucas Miranda Mendonça Rezende, Cézio Luiz Ferreira Junior, Mateus Tarcinalli Machado et al.

Summary

The paper asks whether the tone of central bank communication can sharpen short-term inflation forecasts for Brazil. The authors scraped 251 COPOM minutes (1996–2025) and split them into sentences. They kept only phrases semantically close to "inflation" and had nine LLMs label each phrase as optimistic or pessimistic using a Portuguese prompt that defined both stances. Because each model judges tone differently, they measured every model's bias against human-annotated datasets and derived corrected sentiment scores. These scores were added as exogenous inputs to ARIMA and LSTM models predicting monthly IPCA one step ahead. Accuracy was evaluated with rolling-origin RMSE, t-tests and Diebold-Mariano tests. Adding sentiment gave a modest but statistically significant improvement over sentiment-free baselines. ARIMA benefited strongly and consistently, while LSTM gains appeared only in aggregate. LLMs that agreed more closely with human judgments, such as grok-4-fast and llama-4-maverick, produced better forecasts. The work offers a reusable way to profile an LLM's sentiment bias and turn policy texts into forecasting features.

Key points

  • Corpus: 251 COPOM minutes (January 1996 – July 2025) were parsed from HTML and PDF, using SpaCyLayout with pt_core_news_lg for the PDFs. Qwen3-Embedding-0.6B similarity filtering (cosine > 0.6 to "inflation") produced 9,378 phrases, about 37.4 per meeting.
  • Labelling: nine LLMs from six companies were used, including gpt-5, gemini-2.5-pro, claude-sonnet-4, grok-4-fast, llama-4-maverick and phi-4. Each phrase was sent in a separate request without context, and neutral labels were discarded. Bias was calibrated against three human evaluation datasets.
  • Forecasting models: an ARIMA(1,1,1) model, with its order chosen via PACF, and a multivariate LSTM. Both were evaluated on a 70/30 split with a rolling-origin scheme.
  • Results: t-tests showed significant RMSE gains (p < 0.001) in all four configurations (ARIMA and LSTM, each with and without bias correction). In Diebold-Mariano tests, ARIMA-Uncorrected beat the baseline in 99.9% of runs (mean DM 0.7582), versus 50.9% for LSTM-Uncorrected (mean DM 0.1571). Bias correction weakened ARIMA's gains but strengthened LSTM's.
  • Limitations: a cubic spline fitted over all meeting dates and bias parameters estimated on the full period introduce indirect look-ahead. LLM outputs also vary with reruns, choice of model and prompt wording.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works

🧭 Topics 1Research subjects and languages the work is about.Sentiment & emotion · 1,035 works

Cite

Cite as Lucas Miranda Mendonça Rezende, Cézio Luiz Ferreira Junior, Mateus Tarcinalli Machado and Evandro Eduardo Seron Ruiz. 2026. Enhancing Brazilian Inflation Forecasts through Sentiment Analysis Using Large Language Models. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{rezende2026enhancing,
  title = {{Enhancing Brazilian Inflation Forecasts through Sentiment Analysis Using Large Language Models}},
  author = {Lucas Miranda Mendonça Rezende, Cézio Luiz Ferreira Junior, Mateus Tarcinalli Machado and Evandro Eduardo Seron Ruiz},
  booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1},
  year = {2026},
  url = {https://aclanthology.org/2026.propor-1.17/}
}

References

29 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Fahri Alfiansyah, Mahendra Dwifebri Purbolaksono, and Alfian Akbar Gozali. 2025. Llm-based sentiment analysis in private social media: Case study of my tel-u app. In 2025 International Conference on Data Science and Its Applications (ICoDSA), pages 1100– 1105. IEEE. find
  2. Andres Algaba, David Ardia, Keven Bluteau, Samuel Borms, and Kris Boudt. 2020. Econometrics meets sentiment: An overview of methodology and applications. Journal of Economic Surveys, 34(3):512–547. find
  3. Banco Central do Brasil. 2025a. Atas do Comitê de Política Monetária – Copom. https://www.bcb.gov.br/publicacoes/atascopom/cronologicos. ↗
  4. Banco Central do Brasil. 2025b. IPCA Monthly Series 433 – API. https://api.bcb.gov.br/dados/ serie/bcdata.sgs.433/dados?formato=json. ↗
  5. Gregory P. Casey and Ann L. Owen. 2013. Good news, bad news, and consumer confidence. Social Science Quarterly, 94(1):292–315. find
  6. Michael P. Clements and J. James Reade. 2020. Forecasting and forecast narratives: The Bank of England inflation reports. International Journal of Forecasting, 36(4):1488–1500. find
  7. Francis X. Diebold and Roberto S. Mariano. 1995. Comparing Predictive Accuracy. Journal of Business & Economic Statistics, 13(3):253–263. find
  8. Timothy Dozat. 2016. Incorporating Nesterov Momentum into Adam. find
  9. Technical Report, Stanford University. ICLR 2016 Workshop track. find
  10. Patrick Eugster and Matthias W. Uhl. 2024. Forecasting inflation using sentiment. Economics Letters, 236:111575. find
  11. Explosion AI. 2024. spaCy-layout: Process PDFs, Word documents and other files with spaCy. GitHub repository. Accessed: 2026-03-10. find
  12. Sam Farisa Chaerul Haviana, Sri Mulyono, and Badie’Ah. 2023. The effects of stopwords, stemming, and lemmatization on pre-trained language models for text classification: A technical study. In 2023 10th International Conference on Electrical Engineering, Computer Science and Informatics (EECSI), pag… find
  13. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation, 9(8):1735–1780. find
  14. Klemens Katterbauer and Philippe Moschetta. 2022. An innovative Artificial Intelligence and Natural Language Processing framework for asset price forecasting based on islamic finance: A case study of the Saudi stock market. Econometric Research in Finance, 6(2):183–196. find
  15. Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980. Published at ICLR 2015.
  16. Roman Kräussl and Elizaveta Mirgorodskaya. 2017. Media, sentiment and market performance in the long run. The European Journal of Finance, 23(11):1059– 1082. find
  17. Yang Liu, Xichou Zhu, Zhou Shen, Yi Liu, Min Li, Yujun Chen, Benzi John, Zhenzhen Ma, Tao Hu, Zhi Li, Zhiyang Xu, Wei Luo, and Junhui Wang. 2025. Do large language models possess sensitivity to sentiment? Preprint, arXiv:2409.02370. — [arXiv:2409.02370](https://arxiv.org/abs/2409.02370) find
  18. Greta M Ljung and George EP Box. 1978. On a measure of lack of fit in time series models. Biometrika, 65(2):297–303. find
  19. Sydney C. Ludvigson. 2004. Consumer confidence and consumer spending. Journal of Economic Perspectives, 18(2):29–50. find
  20. Sidra Mehtab and Jaydip Sen. 2019. A robust predictive model for stock price prediction using deep learning and natural language processing. arXiv preprint arXiv:1912.07700. — [arXiv:1912.07700](https://arxiv.org/abs/1912.07700) find
  21. Zahra Moslemi, Logan Clark, Sarah Kernal, Samantha Rehome, Scott Sprengel, Ahoora Tamizifar, Shawna Tuli, Vish Chokshi, Mo Nomeli, Ella Liang, Moury Bidgoli, Jeff Lu, Manish Dasaur, and Marty Hod-gett. 2024. Comprehensive Forecasting of California’s Energy Consumption: A Multi-Source and Sectoral An… — [arXiv:2402.04432](https://arxiv.org/abs/2402.04432) find
  22. Rylan Schaeffer, Mikail Khona, Zachary Robertson, Akhilan Boopathy, Kateryna Pistunova, Jason W. Rocks, Ila Rani Fiete, and Oluwasanmi Koyejo. 2023. Double Descent Demystified: Identifying, Interpreting & Ablating the Sources of a Deep Learning Puzzle. arXiv preprint arXiv:2303.14151. — [arXiv:2303.14151](https://arxiv.org/abs/2303.14151) find
  23. Paul C. Tetlock. 2007. Giving content to investor sentiment: The role of media in the stock market. The Journal of Finance, 62(3):1139–1168. find
  24. Klodiana Istrefi and Michael McMahon. 2025. Communication and inflation expectations. In Research Handbook on Inflation, pages 358–376. Edward Elgar Publishing. find
  25. Volkan Ülke, Afsin Sahin, and Abdulhamit Subasi. 2018. A comparison of time series and machine learning models for inflation forecasting: empirical evidence from the USA. Neural Computing and Applications, 30:1519–1527. find
  26. Carlos M Jarque and Anil K Bera. 1980. Efficient tests for normality, homoscedasticity and serial independence of regression residuals. Economics letters, 6(3):255–259. find
  27. Tong Zhan, Chenxi Shi, Yadong Shi, Huixiang Li, and Yiyu Lin. 2024. Optimization techniques for sentiment analysis based on LLM (GPT-3). Applied and Computational Engineering, 77:251–257. find
  28. Kai Zhang and Yupeng Han. 2025. Harnessing commonsense: Llm-driven knowledge integration for fine-grained sentiment analysis. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 4106–4116. find
  29. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv preprint arXiv:2506.05176.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rezende-2026-enhancing-brazilian-inflation-forecasts-through/