創新文本情感分析方法在報酬率預測中的應用

Abstract

本論文旨在探討「創新文本情感分析方法應用於報酬率預測」之可行性與成效，並著重於先進的自然語言處理技術與具關聯權重的事件指標在財務領域中的應用。全論文包含兩篇主要研究。第一篇研究利用LLaMA-2訓練財務特化之 大型語言模型，結合AI驅動文本摘要策略，針對 10-K 財務報表中的管理披露進行情感分析。實證結果顯示，經過AI摘要的文本能保留重要且具價格影響力的資訊，並在經財務特化訓練後之大語言模型預測後，可大幅提升買進持有報酬(BHR)與累積異常報酬(CAR)的預測準確度。在第二篇文章中，研究聚焦於RavenPack 提供的公司新聞文本情感資料，並創建了多項結合關聯度指標的情感度量，進一步檢驗其與股票報酬率之間的關係。結果顯示，關聯度較高的負面新聞往往帶來更強的異常報酬，說明投資人對悲觀情緒的過度反應在後期可能出現修正。上述兩篇研究綜合展現了大型語言模型、財務文本摘要技術與新聞關聯權重結合的價值，不僅深化了我們對市場動態與情感交互影響的理解，也為投資策略提供了操作上的參考方向。未來可將此框架應用於更多元的金融與經濟場域，進一步發揮 AI 駕馭大規模文本資料的潛力。

This dissertation investigates innovative textual sentiment analysis techniques for return prediction, highlighting the role of advanced natural language processing and relevance-weighted event metrics in financial contexts. The dissertation is composed of two distinct papers, each addressing a crucial aspect of sentiment-based analysis. In Paper 1, a fine-tuned LlaMA-2-based finance-specific large language model (LLM) model is coupled with an AI-driven summarization strategy, enabling the extraction of accurate sentiment signals from lengthy MD&A management disclosures of 10-K filings. Empirical results demonstrate that summarizing documents preserves critical price information and enables the sentiment predictability of finance-specific LLM, significantly improving buy-and-hold return (BHR) and cumulative abnormal return (CAR) predictions. In Paper 2, firm-specific news sentiment is examined using RavenPack data, with newly constructed sentiment indicators integrating relevance metrics. The analysis reveals that negative sentiment linked to higher event relevance often yields stronger abnormal returns, underscoring an investor overreaction mechanism. Together, these papers highlight the value of combining large language models, domain-specific summarization, and relevance weighting to capture subtle investor sentiment. The findings not only enhance our understanding of the interplay between textual data and market dynamics but also offer practical insights for portfolio managers seeking to leverage sentiment signals in trading strategies. Moreover, the frameworks proposed here provide a foundation for future research on AI-driven textual analysis in diverse financial and economic settings.