NX-AI/tirex

What it solves

TiRex is designed to solve the challenge of time series forecasting across both long and short horizons without requiring the user to train the model on their specific dataset. It provides a zero-shot approach to forecasting, meaning it can predict future values based on a provided context of past data without further training.

How it works

TiRex is a 35M parameter pre-trained model based on the xLSTM architecture. It uses enhanced in-context learning to perform forecasting, classification, and regression tasks on time series data. The model can provide both point estimates and quantile estimates for predictions, allowing for uncertainty quantification.

Who it’s for

This project is aimed at data scientists and machine learning engineers who need high-performance time series forecasting, classification, or regression without the overhead of training a model from scratch. It is particularly useful for those working with diverse time series datasets where zero-shot capabilities are necessary.

Highlights

• Zero-Shot Forecasting: Predicts future values without needing training on user data.
• Quantile Predictions: Offers both point and quantile estimates for more robust forecasting.
• xLSTM-based Architecture: Built on the 35M parameter xLSTM model for efficient processing of time series.
• High Performance: Achieves state-of-the-art performance on benchmarks like GiftEval and ChronosZS.
• Multimodal Task Support: Supports forecasting, classification, and regression tasks.
• Optimized Inference: Available as an ONNX model for optimized performance across different hardware platforms.