Autoformer

Why we included this project

Long-term forecasting is where a forecasting model actually proves itself, and this repository holds the official PyTorch implementation of Autoformer, a Transformer-style architecture built for that job. Instead of reading a series as one flat sequence, Autoformer progressively separates it into trend and seasonal components, then swaps classic self-attention for an auto-correlation mechanism that picks out period-based dependencies while keeping complexity log-linear as the forecast horizon grows. The code ships with scripts covering six public benchmarks in energy, traffic, economics, weather, and disease, so reproducing results or borrowing the decomposition idea for your own pipeline is straightforward. The same approach was even used for venue-level weather prediction at the 2022 Winter Olympics, which says a lot about how it holds up outside the lab.