Text-Conditioned Low-Rank Adaptation of Time-Series Foundation Models for Multimodal Forecasting

Time-series foundation models (TSFMs) provide transferable numerical forecasting priors. However, forecast-relevant information may be conveyed in text without being reflected in the observed series. Many multimodal forecasting methods integrate such context through model inputs or intermediate representations. We investigate a complementary mechanism: text-conditioned parameter adaptation. We propose Context-to-LoRA for Time Series (C2LoRA-TS), which encodes text and generates temporary Low-Rank Adaptation (LoRA) updates for selected weight matrices of a frozen TSFM. Both the numerical backbone and text encoder remain frozen, and the method exactly recovers the frozen TSFM output when context is absent. Evaluated on Time-MMD across nine domains and three TSFM architectures, C2LoRA-TS achieves lower aggregate forecasting error than the frozen-TSFM and Static LoRA baselines for every architecture and seed. Replacing the instance-specific text with a single globally fixed context degrades performance for every architecture. These results support text-conditioned parameter generation as a viable strategy for multimodal forecasting with frozen TSFMs.