timesfm-forecasting

Zero-shot time-series forecasting on a CSV file is done by passing your data directly to the pre-trained TimesFM model. The Skill accepts CSV, DataFrame, or 1-D numpy arrays as input and generates point forecasts with quantile prediction intervals without requiring any model training.

The preflight check verifies available RAM, GPU/VRAM, and disk space before loading the TimesFM model. This mandatory resource verification step prevents out-of-memory crashes and runtime failures during the univariate forecasting process.

Batch forecasting for univariate time-series using numpy arrays is supported in TimesFM 2.5 and above. You provide inputs as a list of 1-D numpy arrays, enabling scalable batch processing for richer forecasts with covariates support.

Zero-shot forecasting generates predictions directly from pre-trained models without requiring historical data training. Traditional machine-learning time-series models require explicit training on your specific dataset, whereas TimesFM uses pre-trained weights to deliver immediate forecasts with quantile prediction intervals.

Checking GPU and VRAM before execution is required because the TimesFM model needs significant hardware resources to run the pre-trained weights and context-length handling. The preflight check prevents crashes by ensuring sufficient RAM, GPU memory, and disk space are available.