新聞特徵融合於多模態深度強化學習之投資組合最佳化：以台灣股市為例

Abstract

隨著大型語言模型的崛起，新聞等非結構化資訊成為量化投資中可提取alpha的訊號來源。然而，將新聞情緒整合至多股票深度強化學習（Deep ReinforcementLearning，DRL）投資組合最佳化的研究仍屬稀缺。 本研究以VisionTrader 為基礎架構，於台股元大台灣50 成分股上探索新聞情緒融入多股票DRL 投資組合框架的可行性，圍繞三個遞進目的展開：（一）檢視MSU（Market Score Unit）的貢獻以確立基準架構；（二）比較不同情緒整合配置在不同覆蓋率股票池的效能，並比較四種融合機制以找出最佳整合方式；（三）跨八個時間區段驗證穩健性。 研究結果表明，MSU 在月頻設定下貢獻有限，可安全移除。融合效益高度依賴股票池新聞覆蓋率，高覆蓋率子集可獲顯著提升，混合覆蓋率池則反而衰退；四種融合機制中Gate 表現最佳。跨期分析顯示，純價量模型在八個區段中有六個優於0050；新聞融合相對於純價量的優勢具市場環境依賴性——多頭市場帶來顯著增益，但平盤高雜訊市場此優勢消失。 本研究揭示新聞覆蓋率為融合效益的關鍵調節因素，填補新聞情緒與多股票DRL 投資組合整合的文獻缺口，並為從業者提供實務指引。

With the rise of large language models, unstructured information such as financial news has emerged as a valuable source of alpha signals in quantitative investing. However, research integrating news sentiment into multi-stock deep reinforcement learning (DRL) portfolio optimization remains scarce. This study takes VisionTrader as the backbone architecture and explores the feasibility of incorporating news sentiment into a multi-stock DRL portfolio framework on Taiwan's Yuanta Taiwan 50 ETF (0050) constituents, organized around three progressive objectives: (1) examining the contribution of the Market Score Unit (MSU) to establish a baseline architecture; (2) comparing different sentiment integration configurations across stock pools with varying news coverage rates, and comparing four fusion mechanisms to identify the optimal integration method; (3) verifying cross-period robustness across eight rolling time windows. Results show that MSU contributes marginally under monthly settings and can be safely removed. Fusion effectiveness is highly dependent on news coverage rate: high-coverage subsets achieve significant improvements, while mixed-coverage pools suffer performance degradation; among the four fusion mechanisms, Gate performs best. Cross-period analysis shows that the pure price-volume model outperforms 0050 in six out of eight segments; the advantage of news fusion over the pure price-volume model is market-environment-dependent—yielding significant gains in bull markets but diminishing in sideways, high-noise markets. This study reveals news coverage rate as a critical moderating factor for fusion effectiveness, fills the literature gap at the intersection of news sentiment and multi-stock DRL portfolio optimization, and provides practical guidelines for practitioners.