Discover the open-source massive time mannequin Time-MoE and apply it in a small experiment utilizing Python
12 min learn
10 hours in the past
Historically, the sphere of time sequence forecasting relied on data-specific fashions, the place a mannequin was skilled on a selected dataset and process. If the information or the forecast horizon modified, the mannequin additionally needed to be modified.
Since October 2023, researchers have been actively creating basis forecasting fashions. With these massive time fashions, a single mannequin can now deal with totally different forecasting duties from totally different domains, at totally different frequencies, and with just about any forecast horizon.
Such massive time fashions embrace:
- TimeGPT, which is accessed by way of API making it simple to carry out forecasting, fine-tuning with out utilizing native sources
- Lag-Llama, an open-source mannequin for probabilistic forecasting that constructs options from lagged values
- Chronos, a mannequin primarily based on T5 that translated the unbounded time sequence area to the bounded language area by means of tokenization and quantization
- Moirai, a mannequin that helps exogenous options and the primary to publicly share their dataset LOTSA containing greater than 27B knowledge factors.