Skip to main content

​
From PyPI

Or with uv:
The core install is pure Python and depends only on narwhals, numpy, pydantic and scipy; pandas and polars are both optional. Python 3.10 or newer is required.

​
Verify the install

​
Optional extras

Everything beyond the calibration math is an extra, so the base install never pulls a forecasting framework:
Importing the package never imports a forecasting library The Nixtla adapters are re-exported lazily, so import conformalforecast does not import statsforecast, mlforecast, neuralforecast or torch. The library is imported the first time you touch the adapter that needs it, and if the extra is missing you get an ImportError naming the exact pip install command.
CallableAdapter needs no extra at all — it wraps any function and is pure numpy, which makes it the quickest way to try the package against a model you already have.

​
From a git checkout

For development, clone the repository and use the devenv target, which syncs every extra plus the dev group and installs the pre-commit hooks:
See CONTRIBUTING for the test, lint and documentation workflow.
Alpha versioning conformalforecast is in early development and the API is unstable until v0.2.0. Pin a version if you depend on exact interval values.

​
Next

  • Quick start — calibrated intervals in one screen.
  • Methods — the seven methods, and how to choose one.
  • Nixtlaverse — statsforecast, mlforecast, neuralforecast.
  • Architecture — one class per method, its two surfaces, and which one you want.