timesfm-forecasting

Zero-shot forecasting with TimesFM removes the need to train new models. It accepts CSV, DataFrame, or NumPy array inputs and directly returns point forecasts and calibrated prediction intervals for your univariate time series.

Zero-shot forecasting generates predictions without requiring model training on your specific dataset. You need it when you want to quickly forecast univariate time series across different domains without training new models from scratch.

A GPU is not strictly required, but the Skill enforces a mandatory preflight system check to evaluate your available RAM and GPU. This safety check prevents runtime failures before you execute the from_pretrained, compile, and forecast steps.

Yes, you can get prediction intervals. The Skill returns calibrated prediction intervals alongside point forecasts for your univariate time series data without requiring any training data.

You must prepare a Python environment with TimesFM, torch, numpy, and pandas installed. After setup, you run the preflight system checker and use the from_pretrained, compile, and forecast functions as described to get your results.

TimesFM focuses on univariate time-series data and does not support covariates. It generates point forecasts and calibrated prediction intervals directly from your single series inputs without needing external covariate data.