Text-Attention Guided In-Context Forecasting for Multimodal Time Series Prediction

Real-world time series forecasting often contains both numerical histories and textual context, such as news, reports, or event descriptions. Existing multimodal forecasting methods typically incorporate text through LLM-based prompting or representation-level fusion, which presents a modelling challenge when textual signals are heterogeneous, noisy and unevenly informative. We propose TAG-ICF, a text-attention guided in-context learning (ICL) forecasting framework that uses a frozen, pretrained ICL backbone for numerical prediction while using text to modulate attention over historical examples. Textual inputs are analyzed using LLMs and embedded as lagged features to compute query-context relevance scores, serving as an attention gate that prioritizes contextually relevant historical patterns. TAG-ICF consistently outperforms all multimodal baselines across nine datasets, yielding superior average ranks of 1.50 and 1.22 in MSE and MAE, respectively. Our framework surpasses numerical-only ICL on nearly all datasets and delivers up to 23.94% MAE improvement relative to the strongest competing methods in diverse settings.