Model Overview
MOIRAI is a foundation model for time series forecasting that leverages transformer architecture. It can handle arbitrary number of covariates via its unique any-variate encoding and any-variate attention mechanism. MOIRAI is trained on probabilistic forecasting objective, with the target distribution being a mixture of Gaussian, T-student, log-normal distributions.
Key Features
- Any-variate Encoding: Concatenate all covariates into a single univariate time series
- Multi-variate Foundation Model: Pre-trained on large-scale time series data with various number of covariates
- Zero-Shot Performance: Outperforms domain-specific baselines even in zero-shot settings
