// the find
winedarksea/AutoTS
Automated Time Series Forecasting
AutoTS is a Python AutoML library for time series forecasting that runs genetic algorithm search over a large space of models (naive through deep learning) and preprocessing transforms. It targets data scientists who need production-ready forecasts across many series without manually tuning each one. Won the M6 forecasting competition in 2023, which is a real signal that it works.
The genetic algorithm over both model selection AND preprocessing transforms is genuinely useful — most AutoML tools ignore that the transform pipeline matters as much as the model. The horizontal/mosaic ensembles are smart: different series can get different best models, which matters when your dataset has heterogeneous series. Template import/export means you can run search once, save the winners, and deploy fast without re-running the full genetic search every time. The MCP server is a nice practical addition — if you already have an LLM-based workflow, you can wire forecasting in without writing glue code.
The dependency situation is a real problem — 'additional packages required for some models' undersells it; some of the best models (Prophet, NeuralForecast, GluonTS) each pull in heavy optional installs that often conflict in practice. Training time on large datasets is brutal; the tips section is essentially an apology for this, and the workarounds (subset sampling, upsampling, limiting generations) all trade accuracy for speed. Documentation is scattered across README, extended_tutorial.md, and Sphinx docs with overlapping content and no clear learning path for someone starting from scratch. The Rust PWA frontend is a curious architectural choice that adds Cargo + Trunk as build dependencies for what appears to be a demo app wrapper around Pyodide — it works but it's a significant maintenance surface for a side feature.