google-research/timesfm
8,602 stars·725 forks·Python·apache-2.0·19 viewsresearch.google/blog/a-decoder-only-foundation-model-for-time-series-forecasting↗
Timesfm
TimesFM is a time series foundation model designed to generalize across diverse temporal datasets for forecasting and anomaly detection. It functions as a pretrained model for predicting future values in univariate time series data, eliminating the need for manual training from scratch.
The project includes a framework for adapting pretrained weights to specific datasets using low-rank adaptation to improve accuracy. It also provides specialized capabilities for integrating time-series predictions as tools within autonomous AI agent architectures and automated workflows.
The system supports zero-shot forecasting inference, the integration of exogenous covariates and external regressors, and the identification of irregular patterns through anomaly detection. Model performance is assessed via forecast accuracy evaluation.
Features
- Time Series Forecasting - Provides a foundation model for predicting future values in univariate time series data.
- Cross-Dataset Pretraining - Learns general temporal representations by training across multiple disparate time series datasets to improve generalization.
- Time Series Anomaly Detection - Identifies irregular patterns and outliers in temporal data using a pretrained foundation model.
- Zero-Shot Forecasting - Predicts future trends on unseen datasets using a pretrained foundation model without task-specific training.
- Time Series Forecasting Models - Implements a decoder-only transformer architecture for predictive time series modeling.
- Low-Rank Adaptation - Adapts pretrained weights to specific datasets using low-rank adaptation for improved accuracy.
- Temporal Patching - Groups consecutive time steps into patches to reduce sequence length and capture local temporal patterns.
- Dynamic Covariate Integration - Incorporates external time-varying data streams and regressors into the forecasting process.
- Foundation Models - Provides a decoder-only architecture for time series forecasting.
- Machine Learning - A pretrained foundation model for time-series forecasting.
- Time Series Foundation Models - Decoder-only foundation model for time-series forecasting.
- Time Series Analysis - Pretrained foundation model for time series forecasting.
- Time-Series Forecasting - Listed in the “Time-Series Forecasting” section of the Ailia Models awesome list.
