Research map / Time series / Foundation models for time series

Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting

NeurIPS 2026 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PostTime, a post-training recipe (CoT-SFT with forecast-time traces and fallback, then GRPO with an improvement-ratio reward) that trains Gemma-3-4B to revise TimesFM-2.5 forecasts using textual context on the TimesX benchmark.

Problem

Multimodal time-series forecasting needs to combine strong numerical TSFM priors with non-numerical textual context, but LLMs underperform as standalone numerical forecasters and existing post-training recipes borrowed from math/code are poorly suited.

Approach

Treat the LLM as a context-guided reviser of a TSFM base forecast rather than a direct forecaster; post-train with SFT on multiple forecast-time reasoning traces (top-3 effective plus fallback for non-revisable cases) and RL using an improvement-ratio reward relative to the base forecast instead of absolute accuracy.</dd>

Key results

• PostTime SFT+RL (Gemma-3-4B + TimesFM-2.5) reaches ID nMAE 0.738 / nMSE 0.638 vs TimesFM-2.5 0.788 / 0.729, i.e. 6.38% and 12.43% improvements.
• On OOD, PostTime SFT+RL reaches nMAE 0.746 / nMSE 0.597 vs TimesFM-2.5 0.777 / 0.673 (3.93% and 11.24% gains).
• ImpRatio reward raises mean within-group reward STD to 0.0719 from 0.0171 for ExpMAE and cuts near-zero-variance groups from 27.97% to 4.66%.

Surprising findings

• Fitting ground-truth observed horizons as SFT targets performs worse than the TimesFM-2.5 prior, so ground truth is better used as a verifier than an imitation target.
• Keeping hard and non-revisable (0/5) cases with fallback supervision improves tail nMSE (p99) over filtering to easy samples, which hurt the tail.

Limitations

• Evaluated only on univariate forecasting with text context on one benchmark (TimesX); no image or other modalities.
• Revision restricted to textual forecast edits; no tool use, code execution, or web retrieval explored.
• Relies on a frontier LLM (Gemini) for trace generation, and the trace generator choice affects results (G3.1 stronger on ID, G3 on OOD).

Experimental setup

Datasets: TimesX (refreshed)

Benchmarks: TimesX (refreshed) ID split (88 variables) TimesX OOD split (11 variables)

Metrics: nMAE nMSE average rank valid-window rate fallback rate

Baselines: TimesFM-2.5 Moirai-2.0 Sundial TTS TimeOmni-1-7B Aurora ChatTS-8B GPT-5 Gemini-3.1-Pro Qwen-3-VL-32B GPT-OSS-120B

Builds on: TimesFM-2.5 GRPO TimesX Time-R1 Time-Omni

Closest papers by meaning

• Context is Key: A Benchmark for Forecasting with Essential Textual Information (ICML 2025 · similarity 0.89)
• Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.89)
• LangTime: A Language-Guided Unified Model for Time Series Forecasting with Proximal Policy Optimization (ICML 2025 · similarity 0.89)
• TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · similarity 0.89)
• Time-VLM: Exploring Multimodal Vision-Language Models for Augmented Time Series Forecasting (ICML 2025 · similarity 0.89)
• Empowering Time Series Analysis with Large-Scale Multimodal Pretraining (NeurIPS 2026 · similarity 0.89)