Chronos

Chronos, introduced by Amazon Science researchers in March 2024, reframes time-series forecasting as a language modeling task. Continuous values are scaled and quantized into discrete bins, letting off-the-shelf decoder-only transformer architectures (based on the T5 family, 20M to 710M parameters) train via cross-entropy loss on a large corpus of public time-series data plus synthetic series from Gaussian processes.

The approach produces probabilistic forecasts by sampling multiple future trajectories, with no task-specific fine-tuning required. Across 42 benchmark datasets, Chronos outperformed classical and deep-learning baselines on in-corpus data and matched or beat dataset-specific models zero-shot. Checkpoints and code are public under Apache 2.0, and a faster patch-based variant, Chronos-Bolt, followed later.