timesfm-forecasting

You can forecast univariate time series without training by using zero-shot prediction with TimesFM. This foundation model generates point forecasts and quantile prediction intervals directly from your input data, bypassing the need to train a custom model.

Zero-shot time-series forecasting supports CSV, DataFrame, or array inputs. You can load your historical data in any of these formats to generate point forecasts and quantile prediction intervals using the ForecastConfig-driven workflow.

Yes, you need adequate RAM and GPU VRAM to run TimesFM forecasting. The Skill includes a mandatory preflight system check that verifies your available memory and Python dependencies before loading the model to ensure safe execution.

Yes, you can use external covariates to influence forecasts via the forecast_with_covariates API (XReg). This feature requires TimesFM 2.5+ to be installed with the specific xreg feature enabled.

You configure the forecast horizon and context window using a ForecastConfig-driven workflow. This setup allows you to define context, horizon, normalization, and batch sizing parameters for your time-series predictions.

To run zero-shot time-series forecasting, you need the Python dependencies torch, timesfm, numpy, and pandas. These libraries handle model execution, array operations, and data structuring for the forecasting workflow.