Autoformer is a Transformer architecture enhanced with auto-correlation mechanisms and deep decomposition architecture, enabling series-wise connection for time series forecasting. The model progressively decomposes trend and seasonal components during prediction, achieving state-of-the-art results across six benchmarks covering energy, traffic, economics, weather, and disease domains. Originally published at NeurIPS 2021, it has since been integrated into Hugging Face and the broader Time-Series-Library.