Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting

2026-05-28

AI Summary

Existing time series foundation models struggle to effectively integrate multimodal, non-numerical contextual information to enhance forecasting performance. This work proposes PostTime, a post-training framework that leverages supervised fine-tuning and reinforcement learning with verifiable rewards (RLVR) to guide large language models in conditionally intervening on the raw predictions of time series foundation models—choosing to revise, retain, or discard them based on multimodal context. For the first time, the approach incorporates automatically generated reasoning traces to inform the correction process. Experimental results on the TimesX benchmark demonstrate that PostTime significantly outperforms standalone time series foundation models, pure large language models, and other multimodal forecasting methods, thereby validating its effectiveness and novelty.

Abstract

Time-Series Foundation Models (TSFMs) excel at zero-shot unimodal forecasting using numerical data, but unlike LLMs they cannot consume multimodal, non-numerical context that often shape real-world trajectories. In this work, we bridge this gap and argue for a multimodal time-series forecasting approach that post-trains LLMs to act as context-guided revisors over strong numerical TSFM priors. We introduce PostTime, a post-training recipe combining Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR), along with a methodology to generate automated reasoning traces for forecast revisions. PostTime teaches an LLM to generate context-conditioned forecast interventions -- decisions to revise, preserve, or ignore the TSFM prior based on the multimodal context. We evaluate this approach on the TimesX multimodal forecasting benchmark using a Gemma-3-4B LLM and TimesFM-2.5 TSFM, and show that it significantly outperforms standalone TSFMs, LLM-only baselines, and existing multimodal forecasting approaches.

Problem

• multimodal time-series forecasting
• time-series foundation models
• non-numerical context
• forecast revision
• context-guided prediction

Innovation

• Post-training
• Multimodal Time-Series Forecasting
• Reinforcement Learning with Verifiable Rewards
• Context-Guided Revision
• Time-Series Foundation Models