Decision support systems use LLM embeddings to convert market data into actionable trading insights. However, current financial prediction models often overlook valuable information within short-term intervals (e.g., 4 hours) within longer ones (e.g., a day). The dissemination of news within these shorter periods significantly impacts market movements even for multiple days. This study aims to determine the effectiveness of incorporating fine-grained information into market prediction models that traditionally rely on coarse-grained data. We design a neural network to simultaneously attend to important news and influential indicators present in short-term time slices as well as benefit from market data available in long-term timeframes. With the advancement of contrastive learning-based NLP, we utilize the Angle-optimized Embedding (AoE) sentence transformer for news representation, which generates discriminative embeddings leveraging angle-optimized loss. Besides, to tackle the problem of non-stationary series regression, we employed reversible instance normalization. Comparative results with baseline articles within Forex and cryptocurrencies demonstrate the superiority of the proposed method. Our ablation studies demonstrate that the simultaneous use of financial market data in both fine-grained hourly time slices and coarse-grained daily time slices improves prediction accuracy by up to 60%. Furthermore, utilizing the AoE method to generate informative vector representations for news documents outperformed other embeddings by up to 9.5%.
Farimani et al. (Mon,) studied this question.