timesfm-forecasting

Zero-shot time series forecasting uses a pre-trained foundation model to predict future values without custom training. This Skill uses Google's TimesFM to generate forecasts directly from your historical data, saving significant time and resources.

Yes, you can forecast multiple time series in batch from CSV, DataFrame, or array inputs. The Skill efficiently processes hundreds or thousands of univariate time series simultaneously to generate point forecasts.

Prediction intervals for univariate time series forecasting are generated as calibrated quantiles alongside point forecasts. This provides a measure of uncertainty for your predictions across domains like sales, sensors, energy, and weather.

Running a foundation model for forecasting requires sufficient RAM or GPU resources. This Skill includes a preflight system checker that verifies your available RAM and GPU before model inference to ensure smooth execution.

The best way to forecast sales for hundreds of product lines is batch processing with a zero-shot foundation model. You can input historical weekly sales data to generate point forecasts and 80% prediction intervals for all products simultaneously.