TimeGPT-1

Zero-shot generative pre-trained model for time series forecasting and anomaly detection, trained on over 100 billion data points.

TimeGPT-1 is the first foundation model for time series forecasting and anomaly detection, developed by Nixtla. Unlike large language models, TimeGPT is independently trained on over 100 billion rows of time series data spanning financial, weather, energy, and web domains, using an encoder-decoder architecture built on self-attention with residual connections and layer normalization. It produces accurate zero-shot forecasts without any training on your specific data, making it immediately usable across retail, electricity, finance, and IoT applications.

TimeGPT-1 supports forecasting at daily, weekly, monthly, and hourly frequencies, anomaly detection (both historical batch and real-time online detection), probabilistic forecasting with configurable prediction intervals, exogenous variable integration, and multi-series forecasting at scale. Performance can be further improved through fine-tuning on domain-specific datasets using custom loss functions and adjustable fine-tuning depth. Nixtla also offers a dedicated long-horizon variant (timegpt-1-long-horizon) for forecasts exceeding one seasonal period.

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