timesfm-forecasting

Zero-shot time-series forecasting uses a decoder-based TimesFM model to predict univariate data like sales or sensors without training. The model outputs point forecasts and calibrated quantile intervals directly from the input history.

Yes, TimesFM forecasting supports covariates (XReg) alongside univariate time-series data. This allows you to incorporate external variables to generate point forecasts and quantile-based prediction intervals for more robust uncertainty estimation.

Running TimesFM requires passing a mandatory preflight check for RAM, VRAM, disk space, and Python compatibility. The environment must have TimesFM 2.5, PyTorch, NumPy, and Pandas installed before loading the model.

A decoder-based TimesFM model generates point forecasts alongside calibrated quantile intervals for a range of horizons. This mechanism provides robust uncertainty estimation without requiring model training on the target time series.

You can evaluate zero-shot forecasting accuracy using the provided example scripts. These scripts facilitate end-to-end workflows for loading univariate time-series data, generating forecasts with TimesFM, and measuring the output accuracy.

The preflight check fails when your system lacks sufficient RAM, VRAM, disk space, or compatible Python versions required by TimesFM 2.5. Ensuring these environment prerequisites are met allows the decoder model to load successfully.