timesfm-forecasting

Zero-shot time series forecasting uses Google's TimesFM foundation model to predict future values directly from univariate data without custom training. You can forecast sales or sensor data by providing historical inputs via CSV, DataFrame, or arrays.

Yes, time series forecasting with TimesFM generates both point forecasts and calibrated prediction intervals using quantiles. This provides a single best estimate alongside a range of likely probabilistic outcomes for your univariate input data.

Loading the TimesFM model requires sufficient RAM and GPU resources. A mandatory preflight system checker script verifies your system resources before model loading to ensure the forecasting environment can support the foundation model.

You can forecast future sales from a CSV file by passing the univariate historical sales data to the TimesFM forecasting skill. The foundation model processes the input and generates future point estimates and prediction intervals automatically.

Yes, TimesFM forecasting accepts pandas DataFrames and numpy arrays as valid input formats for univariate time series data. It also supports direct CSV file inputs to generate point forecasts and probabilistic prediction intervals.

Zero-shot forecasting with TimesFM is limited to univariate data and requires sufficient RAM and GPU verified by a preflight checker. It focuses on generating point forecasts and intervals without incorporating external covariates or custom model training.