timesfm-forecasting

Zero-shot time series forecasting is performed using Google's TimesFM foundation model, which requires no custom training. By loading the pre-trained model from Hugging Face, you can generate point forecasts with calibrated prediction intervals directly from your historical data.

You can forecast sales data directly from a CSV file using the TimesFM model without custom machine learning code. The skill accepts CSV, DataFrame, and array inputs, allowing you to generate predictions for future horizons like the next 30 days immediately.

Yes, you need PyTorch installed because the skill lists both the TimesFM library and Torch as dependencies. A preflight system checker runs before loading the model to ensure your environment has sufficient resources to execute the forecasting operations.

You adjust the forecast horizon and context window by configuring the ForecastConfig parameters within the skill. This allows you to customize how much historical data the model considers and how far into the future it generates point forecasts and prediction intervals.

A limitation of using this foundation model is that it is designed specifically for univariate data, meaning it analyzes a single variable over time. It is applicable to various domains like sales, sensors, energy, and weather, but requires downloading the model from Hugging Face.