TimeGPT-1-long-horizon

Zero-shot generative pre-trained model for time series forecasting optimised for long forecast horizons exceeding one seasonal period.

TimeGPT-1-long-horizon is a specialised variant of TimeGPT-1 developed by Nixtla, optimised for forecasts that exceed one seasonal period. It shares the same foundation model architecture as TimeGPT-1 — an encoder-decoder built on self-attention with residual connections and layer normalisation, trained on over 100 billion rows of time series data — and is purpose-built for use cases where the standard TimeGPT-1 variant may lose accuracy at extended horizons.

TimeGPT-1-long-horizon supports forecasting at daily, weekly, monthly, and hourly frequencies, probabilistic forecasting with configurable prediction intervals, exogenous variable integration, and multi-series forecasting at scale. It is recommended when forecast horizons span multiple seasonal cycles, such as annual demand planning, long-range energy load forecasting, or multi-month financial projections.

Note: Legacy model. Intended for testing or benchmarking purposes.