alterlab-timesfm

Zero-shot time-series forecasting uses Google's TimesFM to predict univariate data like sales or sensor readings without custom training. This Skill loads the pre-trained model to deliver point forecasts and quantile intervals directly from your input arrays or DataFrames.

Yes, time-series forecasting with this Skill accepts CSV, pandas DataFrame, or array inputs. It processes your univariate series to generate point forecasts and calibrated quantile intervals using the TimesFM model.

A GPU is not strictly required but is checked during the preflight system check. The Skill verifies available RAM, GPU, and disk space before loading TimesFM to ensure your environment can handle the forecasting workload.

The best way to get prediction intervals is using a zero-shot model like TimesFM. It provides calibrated quantile intervals alongside point forecasts, allowing you to quantify uncertainty without training a custom model.

Yes, TimesFM supports batch forecasting for multiple univariate series. The Skill compiles batch forecasting capabilities using ForecastConfig to process several time series like sensor readings or sales data simultaneously.

TimesFM might fail to load if preflight system checks detect insufficient RAM, GPU unavailability, or low disk space. The Skill runs these checks before model loading to prevent crashes during the forecasting process.