融合新聞特徵與總體經濟指標於多模態原油價格預測

Abstract

原油市場是全球最具波動性且最具影響力的市場之一，其變動受到市場供需動態、宏觀經濟變化及地緣政治事件等複雜因素的交互影響。本研究提出一個多模態深度學習框架，透過整合結構化的數值指標與非結構化的文本特徵，預測原油價格。以西德州中級原油（WTI）即期價格為預測目標，模型結合了9,000多項與商品和宏觀經濟相關的變數，並輔以來自金融新聞的增強文本特徵。為克服新聞標題篇幅短且語意模糊的限制，本研究採用基於GPT的提示工程技術生成具情境感知摘要，並以FinBERT嵌入向量進行編碼。該框架將WTI價格序列分解為趨勢、季節性與殘差三個成分，並分別使用長短期記憶網路（LSTM）進行建模。研究進一步實作並比較三種模態融合策略—早期、中期與後期融合，以找出整合不同資料類型的最佳時機。自訂的混合損失函數則同時考慮預測誤差與價格走勢方向的準確性，以提升對趨勢變化的靈敏度。模型能預測未來最多八週的油價變化，為能源採購與金融風險管理提供決策支援。透過同時運用量化訊號與事件驅動的情緒資訊，本研究提供一個穩健、可解釋且自適應的解決方案，以因應原油市場的高度複雜性。

Crude oil markets are among the most volatile and globally influential, shaped by complex interactions between supply-demand dynamics, macroeconomic shifts, and geopolitical events. This study develops a multimodal deep-learning framework to forecast crude oil prices by integrating structured numerical indicators with unstructured textual features. Using the West Texas Intermediate (WTI) spot price as the target, the model combines over 9,000 commodity-related and macroeconomic variables with enriched textual features derived from financial news. To address the limitations of short and ambiguous headlines, the study employs GPT-based prompt engineering to generate context-aware summaries and encodes them using FinBERT embeddings. The framework decomposes the WTI price series into trend, seasonal, and residual components, modeling each separately with Long Short-Term Memory (LSTM) networks. It implements and compares three fusion strategies (early, intermediate, and late) to determine the optimal stage for modality integration. A hybrid loss function combines prediction error with directional accuracy to enhance trend sensitivity. The model forecasts oil prices up to eight weeks in advance, supporting informed decisions in energy procurement and financial risk management. By jointly leveraging quantitative signals and event-driven sentiment, this study presents a robust, interpretable, and adaptive solution for navigating the complexities of the oil market.