English

ReasonCast: Towards Explainable Time Series Forecasting with Reasoning

Artificial Intelligence 2026-08-03 v1 Machine Learning

Abstract

Most time series (TS) models are specialized for a single task, either understanding (i.e., returning text answers about a TS) or generation (i.e., returning a numeric forecast). Only recently have unified models begun to handle the two within a single architecture. Even these models, however, produce the two outputs as task-separated paths and cannot predict a series and explain why that prediction arises within a single coherent response. In this paper, we argue for a task-fused model that jointly produces 1) prediction (generation) and 2) selfexplanation (understanding), thereby integrating 1) numerical TS forecasting and 2) interpretable text reasoning within a single response. To enable the systematic study of this capability, we present both a benchmark and a recipe that jointly address the two tasks. The benchmark, ReasonTS-Bench, identifies five fundamental patterns underlying TS and enables the joint evaluation of both tasks. ReasonCast, our recipe for finetuning any LLM to perform both tasks jointly, yields a model that generates a reasoning chain and a forecast together in a single autoregressive pass. Extensive experiments show that ReasonCast outperforms both LLMs and TS models on prediction accuracy while producing verifiable, causal reasoning. Code is available at: https://github.com/seunghan96/reasoncast.

Cite

@article{arxiv.2608.01875,
  title  = {ReasonCast: Towards Explainable Time Series Forecasting with Reasoning},
  author = {Seunghan Lee and Jun Seo and Jaehoon Lee and Junhyeok Kang and Sangjun Han and Sungdong Yoo and Minjae Kim and Tae Yoon Lim and Dongwan Kang and Hwanil Choi and Soonyoung Lee and Wonbin Ahn},
  journal= {arXiv preprint arXiv:2608.01875},
  year   = {2026}
}