ViTime: A Vision Intelligence-Powered Foundation Model for Time Series Forecasting

Paper | Code

This is the official repository for the paper: "ViTime: Foundation Model for Time Series Forecasting Powered by Vision Intelligence", published in Transactions on Machine Learning Research (TMLR), 10/2025.

ViTime is a pioneering foundation model for both point and probabilistic Time Series Forecasting (TSF) that fundamentally shifts the paradigm from traditional numerical fitting to a novel approach powered by vision intelligence. Instead of operating directly on numerical sequences, ViTime transforms time series into binary images, leveraging a formally defined metric space to perform forecasting tasks. This vision-based paradigm enhances robustness, generalizability, and performance, especially in zero-shot and few-shot scenarios.

🌟 Key Contributions

• Novel Theoretical Framework: We introduce ViTime, a TSF foundation model grounded in a novel theoretical framework that shifts from conventional numerical fitting to operations within a formally defined binary image-based time series metric space.

• Advanced Data Synthesis (RealTS): To train a truly generalizable model, we developed RealTS, a sophisticated data generation algorithm that synthesizes diverse and realistic time series by modeling fundamental "trend" and "periodicity" components. This mitigates the risk of test-set leakage inherent in models trained on large real-world corpora.

• State-of-the-Art Performance: ViTime sets a new SOTA in zero-shot, few-shot, and robust forecasting. It significantly outperforms existing foundation models like TimesFM and supervised benchmarks, particularly under data perturbations and missing value scenarios.

• Unified Point and Probabilistic Forecasting: ViTime's framework naturally supports both deterministic (point) and probabilistic forecasting. By generating a full probability distribution over the value-axis for each time step, it provides not only precise predictions but also robust uncertainty quantification.

🏛️ Architecture Overview

The ViTime framework comprises four key modules:

1. RealTS Synthesis: Generates diverse synthetic training data to ensure the model learns universal time series patterns.
2. Mapping Function: Converts numerical time series into high-resolution binary images, transforming temporal correlations into spatial patterns.
3. ViTime Model: A Vision Transformer (ViT) based network that learns patterns from the binary image representation. It includes a Visual Time Tokenizer, a Decoder, and a Refining Module to ensure prediction continuity.
4. Inverse Mapping: Converts the predicted binary image back into a numerical time series forecast.

The complete architecture is illustrated below:

Figure 1: ViTime architecture overview. (a) Pipeline comparison. (b) ViTime network modules. (c) Complete architecture.

📊 Zero-Shot Performance

Point Forecasting

ViTime demonstrates exceptional zero-shot generalization. To ensure a fair and rigorous comparison, we introduced scale-invariant metrics (ReMAE and ReMSE), which rescale the test set to mitigate potential test set leakage from training on large, comprehensive real-world datasets.

The table below summarizes the zero-shot point forecasting performance against other leading foundation models. ViTime achieves state-of-the-art results on 11 out of 14 evaluation settings, highlighting its robust generalization.

Model	ETTh1	ETTh2	ETTm1	ETTm2	Electricity	Traffic	Weather
	ReMSE / ReMAE	ReMSE / ReMAE	ReMSE / ReMAE	ReMSE / ReMAE	ReMSE / ReMAE	ReMSE / ReMAE	ReMSE / ReMAE
Numerical Models
Moirai	1.144 / 0.722	0.754 / 0.467	1.448 / 0.849	0.455 / 0.397	0.859 / 0.676	1.416 / 0.894	0.706 / 0.414
Moment	1.089 / 1.240	0.498 / 0.321	0.894 / 0.618	0.542 / 0.582	0.907 / 0.743	1.138 / 0.690	0.545 / 0.349
VisionTS	0.988 / 1.016	0.524 / 0.350	0.873 / 0.559	0.773 / 0.516	0.851 / 0.669	1.173 / 0.669	0.519 / 0.327
TimesFM	0.490 / 0.467	0.374 / 0.396	0.671 / 0.503	0.355 / 0.359	0.367 / 0.404	0.744 / 0.519	0.284 / 0.306
PatchTST-ZS	1.477 / 0.903	1.097 / 0.775	1.295 / 0.798	0.805 / 0.613	1.414 / 0.921	2.054 / 1.002	0.911 / 0.584
Vision-Assisted Models
ViTime-TFM	0.481 / 0.451	0.314 / 0.354	0.519 / 0.455	0.276 / 0.325	0.301 / 0.350	0.718 / 0.460	0.237 / 0.261
ViTime (Ours)	0.457 / 0.431	0.290 / 0.346	0.473 / 0.420	0.237 / 0.301	0.225 / 0.308	0.730 / 0.400	0.203 / 0.228

Probabilistic Forecasting

ViTime also excels at zero-shot probabilistic forecasting, providing well-calibrated uncertainty estimates. It achieves state-of-the-art performance on both standard (CRPS) and scale-invariant (ReCRPS) metrics.

Method	ETTh1	ETTh2	ETTm1	ETTm2	Electricity	Traffic	Weather
CRPS
Moirai	0.506	0.274	0.538	0.309	0.522	0.626	0.506
Lag-Llama	0.441	0.401	0.435	0.432	0.470	0.548	0.394
ViTime	0.356	0.319	0.344	0.286	0.267	0.327	0.244
ReCRPS
Moirai	0.537	0.382	0.631	0.329	0.609	0.684	0.620
Lag-Llama	0.478	0.442	0.463	0.420	0.514	0.629	0.377
ViTime	0.358	0.318	0.346	0.283	0.266	0.324	0.241

🚀 Getting Started

1. Installation

First, clone the repository and navigate to the project directory:

git clone https://github.com/IkeYang/ViTime.git
cd ViTime

Next, install the required dependencies. We recommend using a virtual environment.

pip install -r requirements.txt

2. Download Pre-trained Model

Download our pre-trained model weights from the link below and place them in the appropriate directory.

• Model: ViTime_Model.pth
• Download Link: Google Drive

After downloading, please ensure the path to the model is correctly specified in your configuration file.

💻 Usage / Inference

We provide a simple and efficient ViTimePrediction class for running inference for both point and probabilistic forecasts.

Deterministic (Point) Forecasting

This is the default mode. The model will output a single, deterministic forecast.

import numpy as np
import matplotlib.pyplot as plt
from main import ViTimePrediction

# 1. Prepare your historical data
xData = np.sin(np.arange(512) / 10) + np.sin(np.arange(512) / 5 + 50) + np.cos(np.arange(512) + 50)

# 2. Define the forecast horizon
prediction_length = 720

# 3. Initialize the predictor for deterministic forecasting
vitime = ViTimePrediction(device='cuda:0', model_name='MAE', lookbackRatio=None)

# 4. Generate the forecast
prediction = vitime.prediction(xData, prediction_length)

# 5. Visualize the results
plt.figure(figsize=(12, 6))
plt.plot(np.concatenate([xData, prediction.flatten()], axis=0), label='Prediction')
plt.plot(xData, label='Input Sequence')
plt.title("ViTime Deterministic Forecast")
plt.xlabel("Time")
plt.ylabel("Value")
plt.legend()
plt.show()

Figure: Example of a deterministic forecast generated by ViTime.

Probabilistic Forecasting

To generate probabilistic forecasts, simply specify the number of samples (sampleNumber) you want to generate and adjust the tempature to control the sharpness of the predictive distribution.

import numpy as np
import matplotlib.pyplot as plt
from main import ViTimePrediction

# 1. Prepare your historical data
xData = np.sin(np.arange(512) / 10) + np.sin(np.arange(512) / 5 + 50) + np.cos(np.arange(512) + 50)
prediction_length = 720

# 2. Initialize the predictor for probabilistic forecasting
# A higher tempature (e.g., 8) creates a softer distribution for sampling.
# Set sampleNumber to the desired number of forecast paths.
vitime_prob = ViTimePrediction(device='cuda:0', model_name='MAE', lookbackRatio=None, tempature=8)

# 3. Generate prediction samples
prediction_samples = vitime_prob.prediction(xData, prediction_length, sampleNumber=100)

# 4. Calculate statistics for visualization (e.g., median, prediction intervals)
median_prediction = np.median(prediction_samples, axis=1)
p95 = np.percentile(prediction_samples, 95, axis=1)
p5 = np.percentile(prediction_samples, 5, axis=1)
p75 = np.percentile(prediction_samples, 75, axis=1)
p25 = np.percentile(prediction_samples, 25, axis=1)

# 5. Visualize the probabilistic forecast
plt.style.use('seaborn-v0_8-whitegrid')
fig, ax = plt.subplots(figsize=(12, 6))
x_axis_input = np.arange(len(xData))
x_axis_pred = np.arange(len(xData), len(xData) + prediction_length)

ax.fill_between(x_axis_pred, p5, p95, color='#A8DADC', alpha=0.9, label='90% Prediction Interval')
ax.fill_between(x_axis_pred, p25, p75, color='#457B9D', alpha=0.8, label='50% Prediction Interval')
ax.plot(x_axis_input, xData, color='black', linewidth=1.5, label='Input Sequence')
ax.plot(x_axis_pred, median_prediction, color='#E63946', linestyle='--', linewidth=2.0, label='Median Prediction')

ax.set_title('ViTime Probabilistic Forecast', fontsize=16)
ax.set_xlabel('Time Step', fontsize=12)
ax.set_ylabel('Value', fontsize=12)
ax.legend(loc='upper left')
plt.tight_layout()
plt.show()

> Figure: Example of a probabilistic forecast with 50% and 90% prediction intervals.

📜 Citation

If you find our work useful in your research, please consider citing our paper:

@article{yang2025vitime,
  title={{ViTime}: Foundation Model for Time Series Forecasting Powered by Vision Intelligence},
  author={Yang, Luoxiao and Wang, Yun and Fan, Xinqi and Cohen, Israel and Chen, Jingdong and Zhang, Zijun},
  journal={Transactions on Machine Learning Research},
  year={2025},
  url={https://openreview.net/forum?id=XInsJDBIkp},
  note={Published in Transactions on Machine Learning Research (10/2025)}
}