timesfm-forecasting

Zero-shot time series forecasting uses Google's TimesFM foundation model to generate predictions directly. This approach skips custom model training and hyperparameter tuning, providing immediate forecasts for univariate temporal data like sales or sensor readings.

Probabilistic forecasting is supported natively alongside point forecasts. TimesFM generates calibrated quantile prediction intervals, allowing you to quantify uncertainty ranges for diverse temporal data such as sales, energy metrics, or weather readings.

You validate local hardware compatibility by running the built-in system preflight scripts. These preflight checks ensure your local environment meets resource requirements, preventing crashes before you execute zero-shot univariate forecasting operations.

The best way to forecast thousands of products simultaneously is using TimesFM's foundation model. It performs zero-shot univariate forecasting across large datasets, generating 24-step or 12-month predictions without manual hyperparameter tuning.

Yes, the forecasting skill works with CSV files and processes data using Pandas. It ingests your univariate time series data from standard formats, leveraging Pandas and NumPy arrays to structure inputs for the TimesFM model.

Forecasting fails locally because TimesFM requires specific hardware resources to run Torch operations. Executing the preflight scripts first validates your system environment, ensuring local hardware compatibility and preventing runtime crashes during prediction.