ChronoScribe: Semantic-Numerical Specialization for Instruction-Driven Time Series Forecasting

Key points

• Large language models (LLMs) can interpret unstructured requests but may be inaccurate when asked to produce forecast values directly, whereas time series foundation models (TSFMs) provide numerical forecasts but typically require structured inputs specifying the target series, context window, and horizon.1
• We present ChronoScribe, which applies semantic-numerical specialization to this workflow.1
• A LoRA-adapted LLM proposes an execution path and context length; deterministic rules resolve the executable configuration.1
• We evaluate a Qwen3.5-9B--Chronos-2 implementation across 23 evaluation dimensions and 30 public datasets against eleven proprietary direct LLMs and three open TSFMs. ChronoScribe has the lowest error on sMAPE, MASE, WAPE, RMSE, and MAE among all evaluated systems, including fixed-context controls.1

Sources (1)

• [1]ChronoScribe: Semantic-Numerical Specialization for Instruction-Driven Time Series Forecasting
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 08:03 AM

  Recent advances in instruction-driven time series forecasting highlight a fundamental tension: systems must simultaneously master natural-language instruction following and precise numerical prediction.

  Large language models (LLMs) can interpret unstructured requests but may be inaccurate when asked to produce forecast values directly, whereas time series foundation models (TSFMs) provide numerical forecasts but typically require structured inputs specifying the target series, context window, and horizon.

Extractive summary: sentences quoted from the sources.