| Property | Value |
|---|---|
| Parameter Count | 91 Million |
| Model Type | Time Series Forecasting Transformer |
| Author | Salesforce |
| Paper | arXiv:2402.02592 |
What is moirai-1.0-R-base?
Moirai-1.0-R-base is a sophisticated time series forecasting model that employs a Masked Encoder-based Universal Time Series Forecasting Transformer architecture. It represents the medium-sized variant in the Moirai family, specifically designed to handle multi-variate time series data with both target variables and dynamic covariates.
Implementation Details
The model utilizes a patch-based approach for processing time series data, where each variate is divided into tokens based on a configurable patch size (8 to 128). It incorporates sequence and variate IDs into the embedding layer and employs a Transformer architecture to generate mixture distribution parameters for forecasting.
- Flexible patch size configuration (8, 16, 32, 64, 128, or auto)
- Support for multiple target variables and dynamic covariates
- Customizable prediction and context lengths
- Batch processing capabilities for efficient inference
Core Capabilities
- Multi-variate time series forecasting
- Dynamic covariate handling
- Probabilistic forecasting with mixture distributions
- Rolling window evaluation support
- Integration with GluonTS dataset format
Frequently Asked Questions
Q: What makes this model unique?
Moirai's unique architecture combines patch-based processing with a universal forecasting approach, allowing it to handle various time series tasks while maintaining computational efficiency through its patch-based tokenization strategy.
Q: What are the recommended use cases?
The model is particularly well-suited for complex time series forecasting tasks involving multiple variables and dynamic covariates. It's designed for research purposes and should be carefully evaluated before deployment in production environments, especially for high-risk scenarios.