TimesFM 2.0 500M
| Property | Value |
|---|---|
| Author | Google Research |
| Model Size | 500M parameters |
| Paper | A decoder-only foundation model for time-series forecasting, ICML 2024 |
| Architecture | Decoder-only, 50 layers, 1280 model dimensions |
What is timesfm-2.0-500m-pytorch?
TimesFM 2.0 is Google's advanced foundation model specifically designed for time-series forecasting. It represents a significant evolution in time-series analysis, capable of handling sequences up to 2048 time points with flexible horizon lengths. The model specializes in univariate time series forecasting and includes experimental quantile forecasting capabilities through 10 dedicated heads.
Implementation Details
The model is implemented as a decoder-only architecture with 50 layers and 1280 model dimensions. It uses a patch-based approach with input_patch_len=32 and output_patch_len=128, operating without positional embeddings. The model handles missing values through linear interpolation and supports three frequency categories for different time granularities.
- Supports context lengths up to 2048 time points
- Flexible horizon length configuration
- Handles three frequency categories (high, medium, low)
- Built-in missing value handling through linear interpolation
- PyTorch implementation with efficient batch processing
Core Capabilities
- Univariate time series forecasting
- Point forecasts with experimental quantile predictions
- Supports multiple time frequencies (from minutes to yearly)
- Handles contiguous time series data
- Efficient batch processing with configurable batch sizes
Frequently Asked Questions
Q: What makes this model unique?
TimesFM 2.0 stands out for its versatility in handling different time frequencies and its ability to process long sequences up to 2048 points. It's pretrained on a comprehensive dataset including various time-series sources from the LOTSA pretraining data, making it robust for different forecasting scenarios.
Q: What are the recommended use cases?
The model is ideal for various time-series forecasting applications, from high-frequency data (up to daily granularity) to low-frequency cases (quarterly or yearly). It's particularly suitable for scenarios requiring flexible horizon lengths and dealing with missing values, such as financial forecasting, demand prediction, and resource utilization planning.
