TimeBraid: Unifying Time Series and Language for Understanding and Forecasting

TimeBraid. Unifying Time Series and Language for Understanding and Forecasting. By Xinyue Wang and nine others, from the University of California San Diego, the University of Southern California, and one other institution. Published on the twenty-fourth of September. Its code repository is worked out: about fifty source files, with a readme and a licence.

Two kinds of pretrained model each hold half of what time-series work needs. Language models carry knowledge, instruction following and reasoning. Time-series foundation models carry perception of continuous signals and zero-shot forecasting. Getting one system that both understands and generates in both modalities means joining them, and the joint raises hard questions: where to align the two representation spaces, how to ground language in temporal structure, how to balance understanding against generation, and how to keep joint optimisation stable.

TimeBraid is a series of unified models that align a pretrained language model with a pretrained time-series foundation model through interleaved global residual attention layers. Each model inherits its strengths from both sides and fuses them in a shared representation space, where both modalities are understood and generated.

The paper studies those design questions directly, and the answers become a recipe with three parts. A unified prompting scheme covers diverse time-series and text tasks. Joint training is stabilised. And supervision comes from two point two million curated pairs of series and text, plus four point nine million instruction-tuning samples.

Evaluation spans time-series perception, understanding and reasoning, as well as forecasting, both context-aided and unimodal. Across those benchmarks TimeBraid remains competitive with far larger general-purpose models and with task-specific counterparts. That is two comparisons at once: against models much bigger than it, and against models built for a single job. Holding its own on both fronts is the case the paper makes for unifying the two modalities rather than keeping them apart.

The braid in the name is the point. Rather than building one model to do everything from nothing, TimeBraid weaves two pretrained models together through interleaved layers, so each keeps what it was already good at while sharing one space with the other. Language brings the reasoning and the instructions. The time-series side brings the feel for continuous signals. The recipe is what makes the weave hold together in training.