Reproduce “Text Reinforcement for Multimodal Time Series Forecasting” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chen Su, Yuanhe Tian, Yan Song, Yongdong Zhang

TeR-TSF uses a language model to generate descriptions of time-series patterns from numeric histories and available text, then feeds them to a multimodal forecaster. It refines the generator with preference pairs scored by forecast error and keyword overlap. On Time-MMD’s eight-domain test evaluation, average MSE/MAE fell from 0.953/0.647 without reinforcement to 0.631/0.478 with both rewards. This measures forecasting, not description factuality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.