TimesFM 1.0 200M
| Property | Value |
|---|---|
| License | Apache 2.0 |
| Paper | Link to Paper |
| Architecture | Decoder-only foundation model |
| Developer | Google Research |
What is timesfm-1.0-200m?
TimesFM is a groundbreaking decoder-only foundation model developed by Google Research specifically for time-series forecasting. This first open checkpoint (1.0-200m) represents a significant advancement in the field of time series analysis, offering robust forecasting capabilities with remarkable flexibility in terms of input context and horizon lengths.
Implementation Details
The model architecture features 20 layers with 1280 model dimensions, utilizing input patch lengths of 32 and output patch lengths of 128. It's designed to handle context lengths up to 512 time points and can work with variable horizon lengths, making it highly versatile for different forecasting scenarios.
- Supports univariate time series forecasting
- Handles contiguous context lengths up to 512 time points
- Flexible frequency indicators (high, medium, and low frequency data)
- Point forecast focus with experimental quantile forecasting
Core Capabilities
- Flexible frequency handling (daily, weekly, monthly, etc.)
- Automatic padding/truncation for different input lengths
- Efficient batch processing with parallel computation support
Frequently Asked Questions
Q: What makes this model unique?
TimesFM stands out for its decoder-only architecture specifically designed for time-series forecasting, offering exceptional flexibility in handling various time frequencies and forecast horizons while maintaining high accuracy across different types of time series data.
Q: What are the recommended use cases?
The model is ideal for univariate time series forecasting across different frequencies: high frequency (up to daily granularity), medium frequency (weekly and monthly data), and low frequency (quarterly or yearly data). It's particularly useful for businesses and researchers requiring reliable point forecasts for various time horizons.
