timesfm-forecasting

Zero-shot time-series forecasting is possible using TimesFM to generate predictions without custom training. You supply univariate series via CSV, DataFrame, or arrays, and the model outputs point forecasts alongside calibrated quantile prediction intervals.

Covariates are supported to improve time-series forecasting accuracy. You can pass dynamic numerical or categorical covariates alongside static covariates through CSV, DataFrame, or array inputs to provide contextual signals to the TimesFM model.

Prediction intervals for univariate forecasts are returned as calibrated quantiles. By passing your time-series data into the forecasting workflow, the TimesFM model outputs both point forecasts and quantile intervals to quantify forecast uncertainty.

Hardware checks required before running TimesFM include a mandatory preflight verification of RAM, GPU, disk space, and Python version. This precheck ensures your machine is ready to load the model and execute the forecasting workflow without memory failures.

Zero-shot forecasting works with CSV, DataFrame, and array inputs. You can supply your univariate time series and optional dynamic or static covariates directly in these formats to receive configurable context and horizon forecasts.

Limitations of zero-shot time-series forecasting include strict machine readiness requirements enforced by preflight checks. Additionally, the model focuses on univariate series, requiring sufficient RAM, GPU, and disk capacity to load TimesFM and process configurable context windows.