Research map / Time series / Foundation models for time series
VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters
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 VisionTS, which reformulates time series forecasting as patch-level image reconstruction using a pretrained ImageNet masked autoencoder, achieving strong zero-shot forecasting without time-series training.
Problem
Building time series forecasting (TSF) foundation models is hampered by the cross-domain gap of text LLMs and the heterogeneity of time series datasets; the paper asks whether an image-pretrained visual model can serve as a zero-shot forecaster.
Approach
Segment a univariate series by period into a 2D matrix, render it as a grayscale image, and align the forecast window with masked patches so an ImageNet-pretrained MAE reconstructs the future without time-series training; optional one-epoch fine-tuning of LayerNorm layers adds full-shot gains.
Key results
- Zero-shot VisionTS averages MSE 0.309 on six long-term datasets, best in 7 of 14 settings and about 6% lower MSE than Moirai Small and Large.
- Full-shot VisionTS with one epoch of LayerNorm fine-tuning wins 46 of 80 long-term results.
- VisionTS ranks first in normalized MASE on the GIFT-Eval leaderboard at the time of release, and surpasses individually trained models on the Monash benchmark in zero-shot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VisionTS instead of Moirai Small for zero-shot forecasting | Long-term TSF · Long-term TSF average | MSE: 0.327 -> 0.309 | Helped |
| Use VisionTS instead of TimeLLM for zero-shot forecasting | Long-term TSF · Long-term TSF average | MSE: 0.336 -> 0.309 | Helped |
| Use MAE Large instead of MAE Base (ablation) | VisionTS · Long-term TSF average | MSE: 0.309 -> 0.311 | Hurt |
| Use MAE Huge instead of MAE Base (ablation) | VisionTS · Long-term TSF average | MSE: 0.309 -> 0.315 | Hurt |
| Set periodicity P=1 instead of frequency-selected periodicity (ablation) | VisionTS · Long-term TSF average | MSE: 0.344 -> 0.559 | Hurt |
| Set periodicity P=1 instead of frequency-selected periodicity (ablation) | VisionTS · Long-term TSF average | MAE: 0.370 -> 0.492 | Hurt |
| Replace bilinear interpolation with another tested interpolation strategy (ablation) | VisionTS alignment · Long-term TSF benchmark | forecasting performance | Hurt |
| Apply horizontal or vertical image flips (ablation) | VisionTS alignment · Long-term TSF benchmark | forecasting performance | No clear effect |
| Add a learned convolutional color transformation to the grayscale input (ablation) | VisionTS rendering · Long-term TSF benchmark | forecasting performance | No clear effect |
| Increase context length (ablation) | VisionTS · Weather and ETTm1/ETTm2 | MSE | Helped |
| Fine-tune LN layers instead of freezing all parameters (ablation) | VisionTS · Full-shot long-term TSF | forecasting performance | Helped |
| Fine-tune MLP or attention layers (ablation) | VisionTS · Full-shot long-term TSF | forecasting performance | Hurt |
Surprising findings
- A vanilla ImageNet MAE with no time-series training outperforms several pure time-series foundation models in zero-shot forecasting.
- Swapping the MAE for the LaMa inpainting model gives similar zero-shot performance to Moirai, suggesting the gains come from the image-time series similarity rather than the specific MAE.
Limitations
- Uses only MAE and LaMa; more advanced visual models such as diffusion models are unexplored.
- Cannot capture multivariate inter-variable interactions (channel independence) or produce distributional forecasts.
Experimental setup
Datasets: ImageNet ETTh1 ETTh2 ETTm1 ETTm2 Electricity Weather Traffic Illness Monash GIFT-Eval
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Weather Traffic Illness Monash GIFT-Eval
Metrics: MSE MAE MASE MAPE
Baselines: Moirai TimesFM LLMTime Time-LLM GPT4TS DLinear PatchTST TimesNet ETS ARIMA
Builds on: masked autoencoder Moirai TimesNet LaMa
Closest papers by meaning
- Time-VLM: Exploring Multimodal Vision-Language Models for Augmented Time Series Forecasting (ICML 2025 · similarity 0.88)
- IMTS is Worth Time $\times$ Channel Patches: Visual Masked Autoencoders for Irregular Multivariate Time Series Prediction (ICML 2025 · similarity 0.88)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.88)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.87)
- TARFVAE: Efficient One-Step Generative Time Series Forecasting via TARFLOW based VAE (NeurIPS 2025 · similarity 0.86)
- Multi-Modal View Enhanced Large Vision Models for Long-Term Time Series Forecasting (NeurIPS 2025 · similarity 0.86)
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.