HIC Robustness to Extreme Text Noise

Background: While current methods utilize LLMs for feature extraction or reasoning in time series forecasting, challenges remain in fully unlocking the textual context available for prediction.

Question / Future Work: Investigating the scalability and robustness of the Historical In-Context Learning (HIC) mechanism in scenarios with extremely sparse or highly noisy exogenous text data remains an open question, as current tests only cover moderate degradation (10-20% masking/noise). Further work should explore the limits where performance decline becomes significant and develop mitigation strategies beyond simple quality flagging.