time-series-foundation-models/lag-llama
Lag Llama
Lag-llama is a probabilistic machine learning foundation model designed for time series forecasting. It generates predictive distributions and uncertainty bounds for sequential data across arbitrary frequencies by leveraging pre-trained foundational weights.
The system supports zero-shot transfer inference, allowing it to predict future values on entirely new and unseen datasets without requiring prior retraining. It achieves this by combining generalized representations from foundational training with adjustable context lengths, where historical context lengths and lagged feature values feed directly into the model inputs during iterative autoregressive predictions.
Additionally, the framework provides capabilities for foundation model fine-tuning. Users can adapt pre-trained weights and adjust forecasting models on specialized datasets using targeted hyperparameters and validation splits to optimize domain-specific predictive accuracy.
Features
- Zero-Shot Forecasting - Predicts future values on new datasets without prior training by leveraging pre-trained foundational knowledge and adjustable context lengths.
- Autoregressive Lag Tuning - Predicts future values iteratively by feeding historical context lengths and lagged feature values directly into model inputs.
- Probabilistic Forecasting - Generates predictive distributions and uncertainty bounds for sequential data across arbitrary frequencies using probabilistic deep learning.
- Time Series Foundation Models - Serves as a probabilistic machine learning model pretrained on sequential data to forecast future values and quantify uncertainty across arbitrary frequencies.
- Finetuning Workflows - Adjusts foundational weights on custom datasets using dedicated hyperparameters and validation splits to maximize prediction accuracy.
- Zero-Shot Inference - Predicts future values on unseen target datasets without prior retraining by leveraging generalized representations from foundational training.
- Probabilistic Hybrid Models - Generates predictive distributions and uncertainty bounds for sequential data across arbitrary frequencies using probabilistic modeling.
- Foundational Model Adaptation - Adapts pre-trained neural network weights to specialized downstream domains using targeted hyperparameter tuning and validation splits.
- Time Series Forecasting Fine-Tuners - Tunes pre-trained time series models on specialized datasets to improve predictive accuracy for domain-specific workloads.
- Probability Distributions - Produces probability distributions for future timesteps to quantify uncertainty across any prediction length and data frequency.
