TimeBraid Unifies Language and Time-Series Models via Residual Attention

TimeBraid aligns pretrained language models with time-series foundation models using interleaved global residual attention layers. The architecture uses 2.2M curated series-text pairs and 4.9M instruction-tuning samples to enable unified prompting for both zero-shot forecasting and text-based reasoning.