簡易檢索 / 詳目顯示

回結果列表
研究生： 顧以恩
Ku, Yi-En
論文名稱： 結合大型語言模型情緒標註與VMD分解之歐盟碳排放權期貨價格預測
EUA Carbon Futures Price Forecasting via LLM Sentiment Labeling and VMD Decomposition
指導教授： 鄭順林
Jeng, Shuen-Lin
學位類別： 碩士
Master
系所名稱： 管理學院 - 數據科學研究所
Institute of Data Science
論文出版年： 2026
畢業學年度： 114
語文別： 英文
論文頁數： 152
中文關鍵詞： EUA 碳期貨價格預測 、大型語言模型 、新聞情緒 、情緒分析 、變分模態分解 、機器學習
外文關鍵詞： EUA Carbon Futures Price Prediction, Large Language Model, News Sentiment, Sentiment Analysis, Variational Mode Decomposition, Machine Learning
相關次數： 點閱：103  下載：1 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 本研究旨在探討大型語言模型（Large Language Models, LLM）所產生之新聞情緒標註，是否能提升歐盟碳排放權期貨（European Union Allowance, EUA）價格預測之表現。EUA 價格同時受到能源價格、歐盟減碳政策、總體經濟與地緣政治事件影響，呈現高度非線性、非穩態與多尺度波動特徵；因此，單純依賴歷史價格或傳統技術指標，可能難以及時捕捉新聞事件對市場預期的影響。
    本研究建立一套端對端預測流程，整合市場資料、OilPrice.com 能源新聞、FinBERT情緒分數、Gemini LLM 結構化標註，以及變分模態分解（Variational Mode Decomposition,VMD）。在特徵設計上，本研究將資料分為四組：僅含市場與技術指標的G0、加入FinBERT 情緒特徵的G1、加入Gemini 產生之EUA 相關性與方向信心等結構化標籤但不含FinBERT 的G2，以及同時納入FinBERT 與Gemini LLM 的G3。預測模型則包含ElasticNet、XGBoost、LSTM 與MARS 等方法，並以方向準確率（Directional Accuracy, DA）、RMSE 與MAPE 作為評估指標。
    實驗結果顯示，LLM 情緒特徵並非在所有模型與預測期程下皆能穩定改善預測表現；其效果具有模型依賴性與時間尺度差異。短期預測中，市場技術特徵仍具主導性；在部分中長期預測與高波動體制下，新聞情緒特徵較可能提供額外訊號。此結果意味著，LLM 標註的價值不僅取決於模型能力，也取決於新聞與交易日對齊方式、EUA 相關性篩選、VMD 分頻段建模設計，以及是否嚴格避免未來資訊洩漏。
    本研究之主要貢獻在於：第一，提出結合LLM 情緒標註與VMD 分解之EUA 碳期貨預測架構；第二，以G0/G1/G2/G3 消融實驗檢驗FinBERT 與LLM 新聞特徵的個別與互補增量預測力；第三，將資訊時間對齊與滾動式分解納入實驗設計，以降低未來資訊洩漏風險；第四，從波動體制與IMF 分頻段觀點討論情緒訊號在不同市場狀態下的適用性。

    This study investigates whether news-sentiment annotations produced by Large Language Models (LLMs) can improve the prediction of European Union Allowance (EUA) futures prices. EUA prices are jointly driven by energy prices, EU decarbonization policy, macroeconomic conditions, and geopolitical events, exhibiting highly nonlinear, non-stationary, and multi-scale volatility. Relying solely on historical prices or conventional technical indicators may therefore fail to timely capture the impact of news events on market expectations.
    This study builds an end-to-end forecasting pipeline that integrates market data, OilPrice.com energy news, FinBERT sentiment scores, Gemini LLM structured annotations, and Variational Mode Decomposition (VMD). For feature design, the data are organized into four groups: G0, containing only market and technical indicators; G1, which adds FinBERT sentiment features; G2, which adds Gemini-generated structured labels such as EUA relevance and directional confidence without FinBERT; and G3, which incorporates both FinBERT and Gemini LLM features. The forecasting models include ElasticNet, XGBoost, LSTM, and MARS, evaluated using Directional Accuracy (DA), RMSE, and MAPE.
    Empirical results show that LLM sentiment features do not consistently improve predictive performance across all models and forecast horizons; their effectiveness is model-dependent and varies across time scales. In short-horizon forecasts, market technical features remain dominant, whereas news sentiment features are more likely to provide additional signal in certain medium-to-long-horizon forecasts and under high-volatility regimes. This result implies that the value of LLM annotations depends not only on model capability but also on the news-to-trading-day alignment scheme, EUA relevance filtering, the VMD frequency-band modeling design, and whether look-ahead information leakage is strictly avoided.
    The main contributions of this study are as follows: first, proposing an EUA carbon futures forecasting framework that combines LLM sentiment annotation with VMD decomposition; second, using G0/G1/G2/G3 ablation experiments to examine the individual and complementary incremental predictive power of FinBERT and LLM news features; third, incorporating information-time alignment and rolling decomposition into the experimental design to reduce the risk of look-ahead information leakage; and fourth, discussing the applicability of sentiment signals under different market states from the perspectives of volatility regimes and IMF frequency bands.

    摘要 i Abstract ii 誌謝 iii Table of Contents iv List of Tables vii List of Figures ix Notation and Abbreviations xi Chapter 1 Introduction 1 1.1 Background and Motivation 1 1.1.1 The EU ETS and the Importance of EUA Prices 1 1.1.2 The Predictive Value of News Information and Market Sentiment 2 1.1.3 The Need for Time-Series Decomposition and Multi-Frequency Modeling 2 1.2 Research Questions and Objectives 3 1.2.1 A Note on Method Selection 4 1.3 Research Contributions 5 1.4 Research Workflow and Thesis Structure 5 Chapter 2 Literature Review 8 2.1 Financial Time-Series Forecasting Methods 8 2.2 Carbon-Market Background and EUA Futures Price Forecasting 9 2.2.1 The EU ETS and EUA Futures Price Formation 9 2.2.2 The Methodological Evolution of Carbon-Price Forecasting 10 2.3 Financial News and Text-Feature Construction: From Dictionaries and BERT to LLMs 10 2.3.1 Dictionary Methods and Traditional Text-Sentiment Features 10 2.3.2 BERT, FinBERT, and Financial Language Representation 11 2.3.3 LLM Structured Annotation and Financial Forecasting 12 2.4 LLM Label Reliability and Consistency Checks 12 2.5 Time-Series Decomposition and Multi-Scale Modeling 13 2.6 Summary and Research Gaps 14 Chapter 3 Methodology 16 3.1 Data Collection 16 3.1.1 Market Data 16 3.1.2 News Data 18 3.2 Data Pre-processing and Feature Construction 19 3.2.1 Market Data Pre-processing 19 3.2.2 Market Technical Feature Construction 21 3.2.3 News Data Pre-processing, Annotation Models, and Article-Level Fields 23 3.2.4 News Daily Aggregation 32 3.2.5 Feature Alignment and Ablation Feature Groups 35 3.3 Dependent-Variable Time-Series Processing and Training Design 36 3.3.1 Prediction Target and Time Index 36 3.3.2 VMD Decomposition and Hyperparameter Optimization 37 3.3.3 Rolling VMD Look-Ahead-Free Design 43 3.3.4 IMF Routing Strategy (Route A / B / C) 44 3.3.5 Window-Length Settings 47 3.3.6 Training, Validation, and Test Set Splitting 49 3.4 Prediction Models and Hyperparameter Selection 49 3.4.1 Rationale for Model Selection and Overview of Parameter Tuning 49 3.4.2 ElasticNet 50 3.4.3 XGBoost 52 3.4.4 MARS 52 3.4.5 LSTM 53 3.5 Evaluation Metrics 53 3.6 Testing Workflow 55 3.7 Implementation Environment and Model Versions 59 Chapter 4 Empirical Results and Analysis 61 4.1 Exploration and Pre-processing of Experimental Data 61 4.1.1 Market Data and Pre-processing 61 4.1.2 News Data and Coverage 64 4.2 LLM Annotation Results and Feature Analysis 69 4.2.1 Annotation Setup and Annotation Results 69 4.2.2 Reliability and Consistency Check of the Large Language Model Annotations 71 4.2.3 Three Examples of LLM Annotations 74 4.2.4 Distribution, Anomalies, and Missingness Analysis of the LLM Annotations 78 4.2.5 Construction of LLM Feature Groups 80 4.2.6 Feature Analysis 81 4.3 Time-Series Decomposition and Prediction Routes 83 4.3.1 Rolling VMD Parameter Selection and Leakage-Prevention Mechanism 84 4.3.2 Comparison of Prediction Routes (Route A / B / C) 85 4.4 Model Training and Feature Ablation Analysis 88 4.4.1 Evaluation Basis and Best Model Configurations 88 4.4.2 Ablation Analysis of NLP Features (G0–G3) 94 4.4.3 Volatility-Regime Robustness Analysis (Research Question 2) 97 4.4.4 Statistical Significance Tests 105 4.5 Chapter Summary 109 Chapter 5 Conclusion and Future Work 111 5.1 Research Conclusions 111 5.1.1 Answers to the Research Questions 111 5.1.2 Main Findings 113 5.2 Research Contributions 114 5.2.1 Academic Contributions 114 5.2.2 Practical Contributions 114 5.3 Limitations 115 5.3.1 Data Limitations 115 5.3.2 Methodological Limitations 115 5.3.3 Evaluation and Interpretation Limitations 117 5.3.4 Model Limitations 117 5.4 Future Research Directions 118 References 120 Appendix A Full Distribution of LLM Annotation Fields 125 Appendix B Robustness Results under Tercile Volatility Regimes 127 Appendix C Grouped Summary of News-Derived (FinBERT and Gemini LLM) Feature Columns 129

    [1] Piia Aatola, Markku Ollikainen, and Anne Toppinen. Price determination in the EU ETS market: Theory and econometric analysis with market fundamentals. Energy Economics, 36:380–395, 2013.
    [2] Gerald Appel. Technical Analysis: Power Tools for Active Investors. Financial Times Prentice Hall, 2005.
    [3] Dogu Araci. FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063, 2019.
    [4] Eva Benz and Stefan Trück. Modeling the price dynamics of CO2 emission allowances. Energy Economics, 31(1):4–15, 2009.
    [5] Tim Bollerslev. Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31(3):307–327, 1986.
    [6] George E. P. Box and Gwilym M. Jenkins. Time Series Analysis: Forecasting and Control. Holden-Day, San Francisco, 1976.
    [7] Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001.
    [8] Yassine Charouif, Antony Sommerfeld, and Yves Plancherel. AI-driven meta-learning for long-term commodity price forecasting: A case study of european carbon markets (EU ETS). SSRN Working Paper No. 5292236, 2025.
    [9] Q. Chen and Hiroyuki Kawashima. Stock price prediction using LLM-based sentiment analysis. In 2024 IEEE International Conference on Big Data (BigData), pages 4846–4853, 2024.
    [10] Rui Chen, Haiqi Jiang, Tingyu Guo, and Chenyou Fan. Can large language models forecast carbon price movements? evidence from chinese carbon markets. Research in International Business and Finance, 77:102951, 2025.
    [11] Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, 2016.
    [12] Robert B. Cleveland, William S. Cleveland, Jean E. McRae, and Irma J. Terpenning. STL: A seasonal-trend decomposition procedure based on Loess. Journal of Official Statistics, 6(1):3–73, 1990.
    [13] Rama Cont. Empirical properties of asset returns: Stylized facts and statistical issues. Quantitative Finance, 1(2):223–236, 2001.
    [14] Anna Creti, Pierre-André Jouvet, and Valérie Mignon. Carbon price drivers: Phase i versus phase ii equilibrium? Energy Economics, 34(1):327–334, 2012.
    [15] George Daskalakis, Dimitris Psychoyios, and Raphael N. Markellos. Modeling CO2 emission allowance prices and derivatives: Evidence from the european trading scheme. Journal of Banking & Finance, 33(7):1230–1241, 2009.
    [16] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. pages 4171–4186, 2019.
    [17] Francis X. Diebold and Roberto S. Mariano. Comparing predictive accuracy. Journal of Business & Economic Statistics, 13(3):253–263, 1995.
    [18] Konstantin Dragomiretskiy and Dominique Zosso. Variational mode decomposition. IEEE Transactions on Signal Processing, 62(3):531–544, 2014.
    [19] Robert F. Engle. Autoregressive conditional heteroscedasticity with estimates of the variance of united kingdom inflation. Econometrica, 50(4):987–1007, 1982.
    [20] Robert F. Engle and C. W. J. Granger. Co-integration and error correction: Representation, estimation, and testing. Econometrica, 55(2):251–276, 1987.
    [21] European Energy Exchange. EU ETS Spot, Futures and Options. https://www.eex.com/en/markets/environmental-markets/eu-ets-spot-futures-options, 2026. Accessed 2026-05-27.
    [22] Christian Oliver Ewald and Yaoyu Li. A novel text-based framework for forecasting carbon price. 2025. Working paper, version April 2025.
    [23] Benoit Frenay and Michel Verleysen. Classification in the presence of label noise: A survey. IEEE Transactions on Neural Networks and Learning Systems, 25(5):845–869, 2014.
    [24] Jerome H. Friedman. Multivariate adaptive regression splines. The Annals of Statistics, 19(1):1–67, 1991.
    [25] Xuefeng Gao, Wenbin Wu, and Yun Ye. ChatGPT and commodity return. Journal of Futures Markets, 2025. Early online publication.
    [26] Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. ChatGPT outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30):e2305016120, 2023.
    [27] Joseph E. Granville. Granville's New Key to Stock Market Profits. Prentice-Hall, 1963.
    [28] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997.
    [29] Yumeng Huang, Xingyu Dai, Qunwei Wang, and Dequn Zhou. A hybrid model for carbon price forecasting using GARCH and long short-term memory network. Applied Energy, 285:116485, 2021.
    [30] Jack K. Hutson. Good TRIX. Technical Analysis of Stocks & Commodities, 1(5):105–108, 1983.
    [31] Clayton J. Hutto and Eric Gilbert. Vader: A parsimonious rule-based model for sentiment analysis of social media text. In Proceedings of the International AAAI Conference on Web and Social Media, volume 8, pages 216–225, 2014.
    [32] Intercontinental Exchange. EUA Futures contract specifications. https://www.ice.com/products/197/EUA-Futures, 2026. Accessed 2026-05-27.
    [33] Kexin Jiang, Chuhan Wu, and Yaoran Chen. Revisiting the efficacy of signal decomposition in AI-based time series prediction. arXiv preprint arXiv:2405.06986, 2024.
    [34] Nicolas Koch, Sabine Fuss, Godefroy Grosjean, and Ottmar Edenhofer. Causes of the EU ETS price drop: Recession, CDM, renewable policies or a bit of everything? new evidence. Energy Policy, 73:676–685, 2014.
    [35] Nicolas Koch, Godefroy Grosjean, Sabine Fuss, and Ottmar Edenhofer. Politics matters: Regulatory events as catalysts for price formation under cap-and-trade. Journal of Environmental Economics and Management, 78:121–139, 2016.
    [36] Bryan Lim, Sercan O. Arik, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4):1748–1764, 2021.
    [37] Yongliang Liu, Chunling Tang, Aiying Zhou, Kai Yang, and Huaiyu Yuan. A sentiment-driven three-stage approach for multi-scale carbon price prediction. Discover Sustainability, 6:553, 2025.
    [38] Tim Loughran and Bill McDonald. When is a liability not a liability? textual analysis, dictionaries, and 10-Ks. The Journal of Finance, 66(1):35–65, 2011.
    [39] Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala. Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology, 65(4):782–796, 2014.
    [40] Maria Mansanet-Bataller, Angel Pardo, and Enric Valor. CO2 prices, energy and weather. The Energy Journal, 28(3):73–92, 2007.
    [41] Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2):153–157, 1947.
    [42] A. Menéndez Medina and J. A. Heredia Álvaro. Using generative pre-trained transformers (GPT) for electricity price trend forecasting in the Spanish market. Energies, 17(10):2338, 2024.
    [43] Nagarajan Natarajan, Inderjit S. Dhillon, Pradeep K. Ravikumar, and Ambuj Tewari. Learning with noisy labels. In Advances in Neural Information Processing Systems, 2013.
    [44] Zuokun Ouyang, Philippe Ravier, and Meryem Jabloun. Stl decomposition of time series can benefit forecasting done by statistical methods but not by machine learning ones. In Engineering Proceedings, volume 5, page 42, 2021.
    [45] Marc S. Paolella and Luca Taschini. An econometric analysis of emission allowances prices. Journal of Banking & Finance, 32(10):2022–2032, 2008.
    [46] Alvaro Paredes Amorin and Andre Python. Not all news is equal: Topic- and event-conditional sentiment from finetuned LLMs for aluminum price forecasting. arXiv preprint arXiv:2603.09085, 2026.
    [47] Joshua C. Peterson, Ruairidh M. Battleday, Thomas L. Griffiths, and Olga Russakovsky. Human uncertainty makes classification more robust. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019.
    [48] David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit. Deep learning is robust to massive label noise. arXiv preprint arXiv:1705.10694, 2017.
    [49] Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr. Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. In International Conference on Learning Representations (ICLR), 2024.
    [50] Victor S. Sheng, Foster Provost, and Panagiotis G. Ipeirotis. Get another label? improving data quality and data mining using multiple, noisy labelers. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 614–622, 2008.
    [51] Christopher A. Sims. Macroeconomics and reality. Econometrica, 48(1):1–48, 1980.
    [52] Rion Snow, Brendan O'Connor, Daniel Jurafsky, and Andrew Y. Ng. Cheap and fast – but is it good? evaluating non-expert annotations for natural language tasks. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 254–263, 2008.
    [53] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996.
    [54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017.
    [55] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations (ICLR), 2023.
    [56] J. Welles Wilder. New Concepts in Technical Trading Systems. Trend Research, 1978.
    [57] Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In Advances in Neural Information Processing Systems, 2021.
    [58] Xinyi Yang, Jingyi Li, and Xuchu Jiang. Research on information leakage in time series prediction based on empirical mode decomposition. Scientific Reports, 14:28362, 2024.
    [59] Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11106–11115, 2021.
    [60] Mingyu Zhou and Pei Du. Multivariate events enhanced pre-trained large language model for carbon price forecasting. Energy, 336:138377, 2025.
    [61] Baolong Zhu and Yiming Wei. Carbon price forecasting with a novel hybrid ARIMA and least squares support vector machines methodology. Omega, 41(3):517–524, 2013.
    [62] Hui Zou and Trevor Hastie. Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2):301–320, 2005.

    下載圖示
    校外：立即公開
    QR CODE