Research map / Time series / Foundation models for time series
CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning
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 CALF, a cross-modal LLM fine-tuning framework for multivariate time series forecasting that aligns a temporal branch with a textual GPT2-based branch via a cross-modal match module and feature/output losses.
Problem
LLM-based multivariate time series forecasting suffers from a distribution discrepancy between temporal tokens and textual tokens, which limits accuracy and few/zero-shot generalization.
Approach
Two parallel GPT2 branches (a textual source branch and a temporal target branch) are aligned at input via PCA-reduced word embeddings with cross-attention, at every intermediate layer via a feature regularization loss, and at the output via an output consistency loss; only the temporal branch is used at inference.
Key results
- Reduces MSE/MAE by 7.05%/6.53% versus PatchTST and 5.98%/5.34% versus TimeLLM on long-term forecasting averaged across horizons.
- On few-shot ETT (10% training data), CALF averages about 8% lower error than TimeLLM and 9% lower than PatchTST.
- Efficiency table reports CALF training time of 135 s on ETTm1 versus 1476 s for Time-LLM and 626 s for GPT4TS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CALF instead of PatchTST | Long-term forecasting · ETTh2 | MSE: 0.366 -> 0.349 | Helped |
| Use CALF instead of MICN | Long-term forecasting · Weather | MSE: 0.242 -> 0.250 | Hurt |
| Use CALF instead of iTransformer | Long-term forecasting · Traffic | MSE: 0.428 -> 0.439 | Hurt |
| Add feature regularization loss (ablation) | CALF with supervised loss on ETTh1 · ETTh1 | MSE: 0.446 -> 0.434 | Helped |
| Add output consistency loss (ablation) | CALF with supervised loss on ETTh1 · ETTh1 | MSE: 0.446 -> 0.438 | Helped |
| Add feature regularization loss (ablation) | CALF with supervised loss on Weather · Weather | MSE: 0.263 -> 0.254 | Helped |
| Add output consistency loss (ablation) | CALF with supervised loss on Weather · Weather | MSE: 0.263 -> 0.258 | Helped |
| Add both feature regularization and output consistency losses (ablation) | CALF with supervised loss on ETTh1 · ETTh1 | MSE: 0.446 -> 0.432 | Helped |
| Add both feature regularization and output consistency losses (ablation) | CALF with supervised loss on Weather · Weather | MSE: 0.263 -> 0.250 | Helped |
| Vary the number of principal components (ablation) | CALF cross-modal match module · ETTh1 and ETTh2 | forecasting performance | Mixed |
Surprising findings
- Ablation on the PCA dimension d shows performance is not very sensitive to the number of principal components.
Limitations
- Relies on PCA over word embeddings, which loses information (explained variance ratio of 88% at d=500)
- Alignment is evaluated mostly via visualizations and ablations; the paper does not provide a theoretical justification
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather ECL Traffic M4
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather ECL Traffic M4
Metrics: MSE MAE SMAPE MASE OWA
Baselines: TimeLLM GPT4TS PatchTST iTransformer Crossformer FEDformer TimesNet DLinear TiDE N-BEATS
Builds on: TimeLLM GPT4TS LoRA PatchTST
Closest papers by meaning
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.92)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.90)
- LangTime: A Language-Guided Unified Model for Time Series Forecasting with Proximal Policy Optimization (ICML 2025 · similarity 0.88)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.88)
- $S^2$IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting (ICML 2024 · similarity 0.88)
- TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · 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.