amazon-science/chronos-forecasting

Features

• LLM-Based Forecasting - Adapts a pretrained transformer architecture to treat time series values as discrete tokens for forecasting.
• Probabilistic Forecasting - Produces value ranges and uncertainty intervals through quantile forecasts rather than simple point estimates.
• Zero-Shot Forecasting - Predicts future values for diverse time series using pretrained models without requiring task-specific training.
• Probabilistic Generative Sampling - Generates a distribution of future outcomes by repeatedly sampling from the model's learned probability distribution.
• Supervised Model Weight Optimization - Updates model parameters using specific datasets and configurations to improve prediction accuracy.
• Dynamic Covariate Integration - Incorporates time-varying external features and known future events to improve the precision of time series forecasts.
• Time Series Model Tuning - Allows adjusting model weights and parameters on specific datasets to increase accuracy for unique data patterns.
• Uncertainty Estimation - Quantifies prediction error margins by producing multiple future trajectories through model sampling.
• Time Series Foundation Models - Learns the language of time series for forecasting.

Star history