Author(s)

Xiawei Wang, Shengfeng Guo, Zhangting Wei

Corresponding Author

Xiawei Wang

Abstract

Gold price forecasting is a critical topic in financial quantitative research, with its volatility driven by both traditional supply-demand dynamics and market sentiment. Addressing the issue of insufficient accuracy in existing forecasting models due to the neglect of sentiment factors, this study innovatively proposes a multi-modal analysis framework integrating BERT sentiment quantification and GRU time-series modeling. The framework enhances prediction performance through the synergistic modeling of unstructured text (financial news) and structured data (gold prices). The research adopts a "Fine-tuning-Quantification-Fusion" framework. Firstly, 8115 manually labeled (Positive: 1; Neutral: 0; Negative: -1) gold-related news headlines are used as a corpus to fine-tune pre-trained large language models. Comparative analysis reveals BERT achieves optimal performance (accuracy: 86%). Secondly, the fine-tuned BERT model classifies sentiment for the remaining unlabeled headlines. The label value serves as both the classification output and sentiment score. The daily sentiment index is quantified by averaging the labels of multiple news items per day. Finally, this sentiment index is incorporated into gold price time-series forecasting. Comparing models with and without the sentiment index demonstrates that the GRU model incorporating the sentiment index delivers superior prediction performance (test set MAE=3.96, R²=0.99). Diebold-Mariano (DM) tests further validate that models with the sentiment index significantly outperform baseline models (P<0.01).

Keywords

Gold Price Forecasting, News Sentiment Quantification, BERT, GRU, Multi-Modal Fusion