Time Series Foundation Model
A massive, pre-trained neural network capable of performing zero-shot forecasting or anomaly detection on novel sequential data without task-specific tuning.
Think of It Like This
Like an experienced meteorologist who can predict weather patterns in a brand-new city just by looking at a few days of their thermometer data.
Models like TimeGPT or Chronos are trained on billions of varied time series data points (finance, weather, web traffic). They represent a paradigm shift from traditional statistical models (like ARIMA) which require fitting parameters to every individual dataset. They bring the immense generalization power of LLMs to temporal forecasting.