Oct 2026· Journal of Computational Social Science· Vol 9· 0 citations· 49 references
TL;DR
Experimental results show that the incorporation of aligned textual information improves the accuracy of forecasting, underscoring the importance of robust integration of news data in financial prediction tasks.
Abstract
Accurate financial time series forecasting hinges on effectively integrating both numerical price data and external textual signals, such as news events. While Time-LLM provides a cross-modal time-series prediction mechanism for price sequences, our main contribution focuses on a novel method to fuse these time-series embeddings with learned representations of news text within a large language model (LLM) framework. In contrast to traditional approaches, often relying on text-only training or simple numeric similarity for retrieval, our method leverages the LLM’s capacity to capture intricate correlations between language-based context and price fluctuations. To facilitate this, we introduce aligned datasets that pair daily (or periodic) financial news with the corresponding market data, ensuring temporal consistency. Our model performs token-level concatenation of textual embeddings extracted from a domain-specific financial language model and patched numerical embeddings, followed by frozen LLaMA self-attention in a shared embedding space. By jointly encoding textual and numerical features, the model captures interactions between historical market patterns and temporally aligned textual context. We evaluate the approach on three datasets spanning gold prices, foreign exchange rates, and Bitcoin, which include structured macroeconomic-calendar descriptions and news-derived textual summaries. Experimental results show that the incorporation of aligned textual information improves the accuracy of forecasting, underscoring the importance of robust integration of news data in financial prediction tasks.
Financial time-series forecasting must capture price dynamics across heterogeneous assets while incorporating news available at prediction time. We introduce DualCast, a dual-path framework that extends a frozen language model with a discrete financial vocabulary. Each log-return patch is represented by a learned summa...
Wen-Tao Zhao, Hong-Qiang Wu, Shang-Hang Liu et al.· arXiv.org· 0 citations
Time-series language models typically align numerical signals with language using labels, captions, or question-answer pairs, without explicitly requiring their representations to preserve how the underlying signals evolve. We introduce forecast grounding, a pretraining principle that requires language-aligned time-ser...
Van Dai Do, H. Nguyen, M. Nguyen et al.· 0 citations
Adapting large language models (LLMs) for time series forecasting introduces unique challenges due to the mismatch between continuous time series tokens and the discrete word tokens LLMs are trained on. While LLMs have shown success on natural language tasks, aligning time series tokens with the semantic space of LLMs...
Results indicate that freezing the pretrained backbone and training lightweight projection and adapter modules provides a favourable accuracy--efficiency trade-off with stable behaviour across varying context lengths.
Tamanna S. Kumavat, Georg Brunner, Kyriakos Flouris· arXiv.org· 0 citations
TAC-Time is proposed, a unified framework that transforms textual information into additional temporal channels that preserves temporal continuity and periodic structures while remaining efficient and scalable and identifies predictive textual signals whose correlation-aware alignment yields partial forecasting improve...
Financial forecasting from earnings conference calls requires models to reason over complex corporate disclosures, market expectations, and subtle communication signals. However, existing financial benchmarks are often limited to unimodal inputs or single-task settings, making it difficult to evaluate whether multimoda...