Research map / Time series / Foundation models for time series
Overcoming the Modality Gap in Context-Aided Forecasting
Vincent Zhihao Zheng, Étienne Marcotte, Arjun Ashok, Andrew Robert Williams, Lijun Sun, Alexandre Drouin, Valentina Zantedeschi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train DoubleCast on CAF-7M and give it the correct context | DoubleCast · CAF-7M test (ALL) | CRPS: 0.278 -> 0.231 (vs Chronos) | Improved |
| remove context at test time (ablation) | DoubleCast · CAF-7M test (ALL) | CRPS: +18.3% | Hurt |
| swap context from another window (ablation) | DoubleCast · CAF-7M test (ALL) | CRPS: +26.5% | Hurt |
| train TimeLLM on CAF-7M and remove or swap context (ablation) | TimeLLM (CAF) · CAF-7M test (ALL) | CRPS: -8.0% (no context) | No clear effect |
| keep only contexts accepted by the GPT-5.2 verifier (ablation) | Direct Prompt forecasters · CAF-7M test (HARD) | CRPS | Improved |
| provide swapped context from another instance (ablation) | Direct Prompt · CAF-7M test (HARD, accepted set) | CRPS | No clear effect |
| use larger LLM backbones for Direct Prompt (ablation) | Direct Prompt · CAF-7M test (HARD) | CRPS | Improved |
| pre-train on CAF-7M and fine-tune on CGTSF | DoubleCast · CGTSF (PTF, LEU, MSPG) | CRPS: MSPG hist 480: 0.4191 -> 0.3891 (vs Chronos) | Improved |
| use zero-shot DoubleCast trained only on CAF-7M | DoubleCast · CGTSF PTF and LEU vs MSPG | CRPS: PTF 0.1598 -> 0.1534; MSPG 0.5568 -> 0.7683 | Mixed |
| fine-tune DoubleCast on CGTSF instead of zero-shot | DoubleCast · CGTSF MSPG (hist 192, pred 96) | CRPS: 0.7683 -> 0.6173 | Improved |
| add context-aided training on top of Chronos | Chronos · GIFT-Eval (no context) | normalized MASE: 0.870 -> 0.879 | No clear effect |
| filter training windows with the DP verifier (ablation) | DoubleCast training data · in-domain validation | validation cross-entropy: ~2.95 -> ~2.89 (50k budget) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ChatTime — ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data (AAAI 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Context is Key: A Benchmark for Forecasting with Essential Textual Information (ICML 2025 · similarity 0.92)
- Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting (NeurIPS 2026 · similarity 0.92)
- TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.89)
- Time-VLM: Exploring Multimodal Vision-Language Models for Augmented Time Series Forecasting (ICML 2025 · similarity 0.89)
- TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · similarity 0.88)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
